Actually I made this because I much prefer the API sandboxes/playgrounds of the AI providers, versus their consumer apps. You can edit context, delete bad context, rewrite AI messages. So you can do all that in Cumbersome, and a bunch of other things too.
I also threw in some features like Face/Off Mode where you can get 3 responses at once, and the AI picks the best. (Cursor recently added an ensemble feature like this.) This can be really spendy, but it's worth it for high-value AI sessions. (And it's probably never something that the all-you-can-eat consumer plans will offer as they are incentivized to cheap out.)"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["z",".","ai"],"value":"Show HN: Cumbersome \u2013 iOS/macOS API client for OpenAI, Anthropic, Z.ai"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://apps.apple.com/us/app/cumbersome-ai-api-client/id6753016821"}},"_tags":["story","author_bluepeter","story_46425990","show_hn"],"author":"bluepeter","created_at":"2025-12-29T21:33:30Z","created_at_i":1767044010,"num_comments":0,"objectID":"46425990","points":2,"story_id":46425990,"story_text":"It's called Cumbersome because you have to wrangle API keys (versus just using ChatGPT directly). So, I think of it like driving with a manual transmission, and all the benefits you get there, versus everything-on-auto. Not always easy, but often better (well, for some). This means you can bypass the monthly subscriptions for AI providers and just pay directly. This can work out to be a lot cheaper (or not!) depending on your use.
Actually I made this because I much prefer the API sandboxes/playgrounds of the AI providers, versus their consumer apps. You can edit context, delete bad context, rewrite AI messages. So you can do all that in Cumbersome, and a bunch of other things too.
I also threw in some features like Face/Off Mode where you can get 3 responses at once, and the AI picks the best. (Cursor recently added an ensemble feature like this.) This can be really spendy, but it's worth it for high-value AI sessions. (And it's probably never something that the all-you-can-eat consumer plans will offer as they are incentivized to cheap out.)","title":"Show HN: Cumbersome \u2013 iOS/macOS API client for OpenAI, Anthropic, Z.ai","updated_at":"2026-03-05T23:15:29Z","url":"https://apps.apple.com/us/app/cumbersome-ai-api-client/id6753016821"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"reddec"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["z",".","ai"],"value":"Z.ai phasing out original subscription plans"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["z",".","ai"],"value":"https://docs.z.ai/devpack/transition"}},"_tags":["story","author_reddec","story_47871339"],"author":"reddec","created_at":"2026-04-23T01:25:15Z","created_at_i":1776907515,"num_comments":0,"objectID":"47871339","points":1,"story_id":47871339,"title":"Z.ai phasing out original subscription plans","updated_at":"2026-04-23T01:30:08Z","url":"https://docs.z.ai/devpack/transition"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"indigodaddy"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["z",".","ai"],"value":"Claude Code with Z.ai Vision MCP"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://jpcaparas.medium.com/claude-code-with-z-ai-vision-mcp-master-the-full-toolbelt-4447c2f953a0"}},"_tags":["story","author_indigodaddy","story_46899452"],"author":"indigodaddy","created_at":"2026-02-05T13:32:07Z","created_at_i":1770298327,"num_comments":0,"objectID":"46899452","points":1,"story_id":46899452,"title":"Claude Code with Z.ai Vision MCP","updated_at":"2026-03-05T23:32:19Z","url":"https://jpcaparas.medium.com/claude-code-with-z-ai-vision-mcp-master-the-full-toolbelt-4447c2f953a0"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"jeudesprits"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["z",".","ai"],"value":"https://twitter.com/Zai_org/status/2093354097122455713
https://z.ai/blog/glm-5.3"},"title":{"matchLevel":"none","matchedWords":[],"value":"GLM-5.3 is now open-weight"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://huggingface.co/zai-org/GLM-5.3"}},"_tags":["story","author_jeudesprits","story_49479878"],"author":"jeudesprits","children":[49480188,49480282,49480347,49480393,49480440,49480591,49480594,49480611,49480727,49480810,49480839,49480864,49481270,49481363,49481697,49481955,49482353,49482578,49483041,49483051,49483167,49483481,49483807,49483853,49485626,49486004,49486055,49486158,49486240,49487294,49487800,49488352,49488640,49488976,49495741,49497740,49502961],"created_at":"2026-08-28T15:20:13Z","created_at_i":1787930413,"num_comments":280,"objectID":"49479878","points":800,"story_id":49479878,"story_text":"https://twitter.com/Zai_org/status/2093354097122455713
https://z.ai/blog/glm-5.3","title":"GLM-5.3 is now open-weight","updated_at":"2026-08-31T08:24:37Z","url":"https://huggingface.co/zai-org/GLM-5.3"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"merge-conflict"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["z",".","ai"],"value":"GAC is a tool I built to help users spend less time summing up what was done and more time building. It uses LLMs to generate contextual git commit messages from your code changes. And it can be a drop-in replacement for `git commit -m "..."`.
Example:
feat(auth): add OAuth2 integration with GitHub and Google\n\n - Implement OAuth2 authentication flow\n\n - Add provider configuration for GitHub and Google\n\n - Create callback handler for token exchange\n\n - Update login UI with social auth buttons\nYou can try it out with uvx (no install):
uvx gac init # config wizard\n\n uvx gac\nTech details:
14 providers - Supports local (Ollama & LM Studio) and cloud (OpenAI, Anthropic, Gemini, OpenRouter, Groq, Cerebras, Chutes, Fireworks, StreamLake, Synthetic, Together AI, & Z.ai (including their extremely cheap coding plans!)).
Three verbosity modes - Standard with bullets (default), one-liners (`-o`), or verbose (`-v`) with detailed Motivation/Architecture/Impact sections.
Secret detection - Scans for API keys, tokens, and credentials before committing. Has caught my API keys on a new project when I hadn't yet gitignored .env.
Flags - Automate common workflows:
`gac -h "bug fix"` - pass hints to guide intent\n\n `gac -yo` - auto-accept the commit message in one-liner mode\n\n `gac -ayp` - stage all files, auto-accept the commit message, and push (yolo mode)\nGitHub: https://github.com/cellwebb/gac"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Git Auto Commit (GAC) \u2013 LLM-powered Git commit command line tool"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/cellwebb/gac"}},"_tags":["story","author_merge-conflict","story_45723533","show_hn"],"author":"merge-conflict","children":[45723793,45723986,45724220,45724309,45724426,45724576,45724644,45724655,45724850,45724911,45724938,45725025,45725162,45725430,45725991,45726244,45726331,45726417,45733397,45743483,45758713],"created_at":"2025-10-27T17:07:05Z","created_at_i":1761584825,"num_comments":36,"objectID":"45723533","points":56,"story_id":45723533,"story_text":"GAC is a tool I built to help users spend less time summing up what was done and more time building. It uses LLMs to generate contextual git commit messages from your code changes. And it can be a drop-in replacement for `git commit -m "..."`.
Example:
feat(auth): add OAuth2 integration with GitHub and Google\n\n - Implement OAuth2 authentication flow\n\n - Add provider configuration for GitHub and Google\n\n - Create callback handler for token exchange\n\n - Update login UI with social auth buttons\nYou can try it out with uvx (no install):
uvx gac init # config wizard\n\n uvx gac\nTech details:
14 providers - Supports local (Ollama & LM Studio) and cloud (OpenAI, Anthropic, Gemini, OpenRouter, Groq, Cerebras, Chutes, Fireworks, StreamLake, Synthetic, Together AI, & Z.ai (including their extremely cheap coding plans!)).
Three verbosity modes - Standard with bullets (default), one-liners (`-o`), or verbose (`-v`) with detailed Motivation/Architecture/Impact sections.
Secret detection - Scans for API keys, tokens, and credentials before committing. Has caught my API keys on a new project when I hadn't yet gitignored .env.
Flags - Automate common workflows:
`gac -h "bug fix"` - pass hints to guide intent\n\n `gac -yo` - auto-accept the commit message in one-liner mode\n\n `gac -ayp` - stage all files, auto-accept the commit message, and push (yolo mode)\nGitHub: https://github.com/cellwebb/gac","title":"Show HN: Git Auto Commit (GAC) \u2013 LLM-powered Git commit command line tool","updated_at":"2026-03-05T22:56:15Z","url":"https://github.com/cellwebb/gac"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"aaronSong"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["z",".","ai"],"value":"Clink lets you use the coding agents you already pay for (Claude Code, Codex CLI, Gemini CLI, Z.ai GLM) to build \u2192 live-preview \u2192 ship apps in an isolated container.
No token purchases, no extra cost for coding. Just link your existing Claude/OpenAI/Gemini account and start building and deploying instantly.
Why we built this:
Claude Code is our go-to for coding, but it lacked preview + deploy capabilities. We didn't want to pay Lovable again just for that.
Different agents excel at different tasks - Claude Code for versatility, Codex for complex work, GLM for speed. We needed one platform to leverage them all.
CLI agents offer more freedom than traditional web builders. We wanted to unlock their full potential with proper dev tooling.
What it does:
- Prompt \u2192 Build \u2192 Live \u2192 Deploy - The fastest path from idea to live website. Deploy for free.
- BYO Subscription - Use your existing plans efficiently (Claude Code $20 = 10x Lovable $25 usage, GLM $3 = 3x Claude Code $20)
DEV Mode (Beta):
- Multi-stack support - Build with Node, Python, Go, Rust and deploy containers to public URLs instantly
- Repo imports - Upgrade and deploy your existing projects across any stack
Links:
- Clink: https://clink.new
- OSS origin (Claudable, ~2.8k): https://github.com/opactorai/Claudable
We'd love any feedback, bug reports, or stack requests - we iterate fast and read every comment."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Clink \u2013 Bring your own CLI Agents, Ship instantly"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://clink.new"}},"_tags":["story","author_aaronSong","story_45657055","show_hn"],"author":"aaronSong","children":[45657110,45657161,45657251,45657269,45657477,45657502,45657580,45657919,45658425,45658646],"created_at":"2025-10-21T15:32:02Z","created_at_i":1761060722,"num_comments":21,"objectID":"45657055","points":22,"story_id":45657055,"story_text":"Clink lets you use the coding agents you already pay for (Claude Code, Codex CLI, Gemini CLI, Z.ai GLM) to build \u2192 live-preview \u2192 ship apps in an isolated container.
No token purchases, no extra cost for coding. Just link your existing Claude/OpenAI/Gemini account and start building and deploying instantly.
Why we built this:
Claude Code is our go-to for coding, but it lacked preview + deploy capabilities. We didn't want to pay Lovable again just for that.
Different agents excel at different tasks - Claude Code for versatility, Codex for complex work, GLM for speed. We needed one platform to leverage them all.
CLI agents offer more freedom than traditional web builders. We wanted to unlock their full potential with proper dev tooling.
What it does:
- Prompt \u2192 Build \u2192 Live \u2192 Deploy - The fastest path from idea to live website. Deploy for free.
- BYO Subscription - Use your existing plans efficiently (Claude Code $20 = 10x Lovable $25 usage, GLM $3 = 3x Claude Code $20)
DEV Mode (Beta):
- Multi-stack support - Build with Node, Python, Go, Rust and deploy containers to public URLs instantly
- Repo imports - Upgrade and deploy your existing projects across any stack
Links:
- Clink: https://clink.new
- OSS origin (Claudable, ~2.8k): https://github.com/opactorai/Claudable
We'd love any feedback, bug reports, or stack requests - we iterate fast and read every comment.","title":"Show HN: Clink \u2013 Bring your own CLI Agents, Ship instantly","updated_at":"2026-03-05T22:51:59Z","url":"https://clink.new"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"Frannky"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["z",".","ai"],"value":"Claude Code with Opus and the Max plan is fine for me, even though I'm not super happy about moments when it's not available, the costs, account banning, etc.
Anyway, what I am looking for and am curious about is if there is a solution that I am overlooking that will work the same, or almost the same or better, but at a cheaper price.
I read about people being happy about pi.dev and OpenCode. I tried OpenCode with Mimo V2 pro and it is pretty good. I previously used Qwen CLI before they stopped the free usage, and Gemini CLI. I also used Z.ai with OpenCode.
I read about people using Opus for planning and then for non-important stuff moving the agent to use a further cheaper model. I am not into usage-based pricing unless it will be cheaper nonetheless (I doubt it though).
Do you have some cool setups to share? I usually do Python for backend and TypeScript frontend."},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: Claude Code Alternative"}},"_tags":["story","author_Frannky","story_47854570","ask_hn"],"author":"Frannky","children":[47854585,47854790,47855039,47869765],"created_at":"2026-04-21T21:13:04Z","created_at_i":1776805984,"num_comments":6,"objectID":"47854570","points":10,"story_id":47854570,"story_text":"Claude Code with Opus and the Max plan is fine for me, even though I'm not super happy about moments when it's not available, the costs, account banning, etc.
Anyway, what I am looking for and am curious about is if there is a solution that I am overlooking that will work the same, or almost the same or better, but at a cheaper price.
I read about people being happy about pi.dev and OpenCode. I tried OpenCode with Mimo V2 pro and it is pretty good. I previously used Qwen CLI before they stopped the free usage, and Gemini CLI. I also used Z.ai with OpenCode.
I read about people using Opus for planning and then for non-important stuff moving the agent to use a further cheaper model. I am not into usage-based pricing unless it will be cheaper nonetheless (I doubt it though).
Do you have some cool setups to share? I usually do Python for backend and TypeScript frontend.","title":"Ask HN: Claude Code Alternative","updated_at":"2026-06-17T16:22:53Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"sahli"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["z",".","ai"],"value":"It feels like most Claude Code users have already noticed a quality drop in the Claude models. As a Claude Pro subscriber (Web version; I don't use Claude Code), I\u2019ve seen a clear decline over the last couple of weeks. I can\u2019t complete tasks in a single turn anymore. Claude often stops streaming because it hits some internal tool-call/turn limit, so I have to keep pressing \u201cContinue.\u201d Each continuation has to re-feed context, which quickly burns through tokens and quota. The model also makes more mistakes and fails to fully complete tasks it used to handle reliably.
This is especially frustrating because Sonnet 4.6 was a real step up: it could produce long, correct code in one pass much more often. That seems basically gone now.
As a paying Pro user, I honestly find myself using free alternatives like DeepSeek and Z.ai (GLM) more than Claude lately. I\u2019ve also stopped touching Opus entirely\u2014it\u2019s so token-hungry that it drains my weekly quota too fast to be practical.
Is Anthropic trying to limit usage or drive people away?"},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: Is Claude Getting Worse?"}},"_tags":["story","author_sahli","story_47778035","ask_hn"],"author":"sahli","children":[47778091,47778190,47778194,47778226,47778241,47778251,47778294,47778404,47786971,47787222,47791126,47803399,47806829,47824099,47856224],"created_at":"2026-04-15T12:20:05Z","created_at_i":1776255605,"num_comments":20,"objectID":"47778035","points":9,"story_id":47778035,"story_text":"It feels like most Claude Code users have already noticed a quality drop in the Claude models. As a Claude Pro subscriber (Web version; I don't use Claude Code), I\u2019ve seen a clear decline over the last couple of weeks. I can\u2019t complete tasks in a single turn anymore. Claude often stops streaming because it hits some internal tool-call/turn limit, so I have to keep pressing \u201cContinue.\u201d Each continuation has to re-feed context, which quickly burns through tokens and quota. The model also makes more mistakes and fails to fully complete tasks it used to handle reliably.
This is especially frustrating because Sonnet 4.6 was a real step up: it could produce long, correct code in one pass much more often. That seems basically gone now.
As a paying Pro user, I honestly find myself using free alternatives like DeepSeek and Z.ai (GLM) more than Claude lately. I\u2019ve also stopped touching Opus entirely\u2014it\u2019s so token-hungry that it drains my weekly quota too fast to be practical.
Is Anthropic trying to limit usage or drive people away?","title":"Ask HN: Is Claude Getting Worse?","updated_at":"2026-04-25T21:31:34Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"vinhnx"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["z",".","ai"],"value":"VT Code is a Rust CLI/TUI coding agent for AST-aware edits (Tree-sitter, ast-grep). Multi-provider routing with failover and caching (OpenAI, Anthropic, Gemini, DeepSeek, xAI, OpenRouter, Z.AI, Moonshot; Ollama locally). Policy-gated tools, workspace boundaries, Zed ACP integration. Config-first via vtcode.toml; reproducible model/constant metadata in the repo.
Try it: cargo install vtcode; vtcode
Code: https://github.com/vinhnx/vtcode"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: VT Code \u2013 AST-aware Rust agent for terminal (Tree-sitter/AST-grep)"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/vinhnx/vtcode"}},"_tags":["story","author_vinhnx","story_45617591","show_hn"],"author":"vinhnx","created_at":"2025-10-17T15:01:22Z","created_at_i":1760713282,"num_comments":0,"objectID":"45617591","points":6,"story_id":45617591,"story_text":"VT Code is a Rust CLI/TUI coding agent for AST-aware edits (Tree-sitter, ast-grep). Multi-provider routing with failover and caching (OpenAI, Anthropic, Gemini, DeepSeek, xAI, OpenRouter, Z.AI, Moonshot; Ollama locally). Policy-gated tools, workspace boundaries, Zed ACP integration. Config-first via vtcode.toml; reproducible model/constant metadata in the repo.
Try it: cargo install vtcode; vtcode
Code: https://github.com/vinhnx/vtcode","title":"Show HN: VT Code \u2013 AST-aware Rust agent for terminal (Tree-sitter/AST-grep)","updated_at":"2026-03-05T22:49:16Z","url":"https://github.com/vinhnx/vtcode"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"lamprouge"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["z",".","ai"],"value":"retrieval is not the future of Ai
if it was, google would have won already
there is another way, that doesn't require retrieval at all
it costs more per conversation, but its better than a sh*ty graphRAG
i am shocked by how few people know this
but it seems the industry is not paying much attention to memory yet (hell z.ai doesn't have longterm memory at all)
humans don't really remember the exact experience that happened,
we remember a layer that sits on top of that
something that got built at the same day the experience happened
but is not a transcript or a summary
why would the brain do that! isn't it easier to save the actual memory?
just like LLMs, our brain will be overwhelmed if we remembered every single experience (people with photographic memory go through hell)
Ai agents can't remember everything, or retrieval will be hell too
just ask google; 30 years of trying to give you the best search results of a simple recipe & it still fails to do that on the first try
so what! do you take the first word and the last word of every sentence and reconstruct the sentence later?
No, that too doesn't work either, no matter how much Zep AI wants you to believe it"},"title":{"matchLevel":"none","matchedWords":[],"value":"Retrieval is not the future of AI \u2013 if it was, Google would have won already"}},"_tags":["story","author_lamprouge","story_48788520","ask_hn"],"author":"lamprouge","children":[48788911,48796423,48809946],"created_at":"2026-07-04T20:09:03Z","created_at_i":1783195743,"num_comments":2,"objectID":"48788520","points":4,"story_id":48788520,"story_text":"retrieval is not the future of Ai
if it was, google would have won already
there is another way, that doesn't require retrieval at all
it costs more per conversation, but its better than a sh*ty graphRAG
i am shocked by how few people know this
but it seems the industry is not paying much attention to memory yet (hell z.ai doesn't have longterm memory at all)
humans don't really remember the exact experience that happened,
we remember a layer that sits on top of that
something that got built at the same day the experience happened
but is not a transcript or a summary
why would the brain do that! isn't it easier to save the actual memory?
just like LLMs, our brain will be overwhelmed if we remembered every single experience (people with photographic memory go through hell)
Ai agents can't remember everything, or retrieval will be hell too
just ask google; 30 years of trying to give you the best search results of a simple recipe & it still fails to do that on the first try
so what! do you take the first word and the last word of every sentence and reconstruct the sentence later?
No, that too doesn't work either, no matter how much Zep AI wants you to believe it","title":"Retrieval is not the future of AI \u2013 if it was, Google would have won already","updated_at":"2026-07-22T04:51:13Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"deathmonger5000"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["z",".","ai"],"value":"Hi HN,
Circus Chief is a tool for managing coding agent sessions from a browser. It's specifically optimized for small screens. It supports Claude Code, OpenAI Codex, and Google Gemini CLI agents.
Features
Agents can operate Circus Chief itself. Agents can spawn sessions, schedule sessions, interact with the Kanban board \u2014 anything you can do in the UI, an agent can also do.
Schedule work ahead of time.
Automatically reschedule when you hit usage limits.
Configurable, chainable prompt templates.
User-defined commands - agents can run them and see the results, and so can you in the UI. Handy for local CI or routine workflow steps.
Worktree-per-session isolation, or elect to work on a specific branch.
Shared canvas - this is a place for shared artifacts that don't belong in the code. Useful for iterating on planning documents.
Bring your own provider. Use subscription auth for Anthropic, OpenAI, or Google, or point sessions at third-party providers with Anthropic- or OpenAI-compatible endpoints.
If you want to try it:
npx circuschief\nDevelopment process
I built this almost entirely from my phone and tablet \u2014 starting with Claude Code using both Anthropic and z.ai, then adding Codex to the mix around a month ago.
It was vibe coded in the sense that I was not reviewing many individual lines of code. I did review and iterate on plans with the agents, but often stayed out of implementation details. If a plan made sense then I would usually have the agent review it a few times before just letting the agent implement without any intervention. I've enjoyed building this way, but I don't deeply understand a lot of the details of the implementation. That's bitten me exactly once so far, see this issue: https://github.com/ferrislucas/Circus-Chief/issues/944
I used Circus Chief to build Circus Chief from a very early stage. I bootstrapped with Vibe Kanban, then switched to dogfooding once Circus Chief was usable. A lot of the app has been built through the app itself.
By the numbers
2,184 agent sessions since Christmas 2025 \u2014 1,423 were child/follow-up sessions.
Agent mix: 1,818 Claude Code sessions, 363 Codex sessions, and 3 Gemini sessions.
50,929 conversation messages across those sessions.
41 sessions (1.9%) contained strong profanity in user-authored messages.
990 sessions linked to GitHub PRs, across 685 distinct PRs, 759 branches, and 731 worktrees.
Average logged agent call: 131.3 seconds.
2,410 commits on main; 883 pull requests total \u2014 847 merged (~96%).
Repo scale: 887 tracked files and 310K current lines of code (excludes lockfiles, fixtures, and recorded test cassettes).
Peak day: 76 commits on March 21, 2026, first at 12:49am, last at 11:57pm \u2014 one every 18 minutes for 23 hours.
Busiest merge day: 28 pull requests shipped \u2014 New Year's Day, 2026.
Largest PR by diff: 140K insertions across 92 files \u2014 replacing mock Claude fixtures with VCR-style record/replay.
Code: https://github.com/ferrislucas/Circus-Chief
Analytics disclosure: the npm build includes PostHog for anonymous page-view analytics (session recording off, DNT respected). Disable with --no-analytics, in Settings \u2192 General, or build from source without a PostHog key."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Circus Chief \u2013 Claude Code, Codex, and Gemini from Your Phone"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/ferrislucas/Circus-Chief"}},"_tags":["story","author_deathmonger5000","story_48370360","show_hn"],"author":"deathmonger5000","created_at":"2026-06-02T13:59:56Z","created_at_i":1780408796,"num_comments":0,"objectID":"48370360","points":4,"story_id":48370360,"story_text":"Hi HN,
Circus Chief is a tool for managing coding agent sessions from a browser. It's specifically optimized for small screens. It supports Claude Code, OpenAI Codex, and Google Gemini CLI agents.
Features
Agents can operate Circus Chief itself. Agents can spawn sessions, schedule sessions, interact with the Kanban board \u2014 anything you can do in the UI, an agent can also do.
Schedule work ahead of time.
Automatically reschedule when you hit usage limits.
Configurable, chainable prompt templates.
User-defined commands - agents can run them and see the results, and so can you in the UI. Handy for local CI or routine workflow steps.
Worktree-per-session isolation, or elect to work on a specific branch.
Shared canvas - this is a place for shared artifacts that don't belong in the code. Useful for iterating on planning documents.
Bring your own provider. Use subscription auth for Anthropic, OpenAI, or Google, or point sessions at third-party providers with Anthropic- or OpenAI-compatible endpoints.
If you want to try it:
npx circuschief\nDevelopment process
I built this almost entirely from my phone and tablet \u2014 starting with Claude Code using both Anthropic and z.ai, then adding Codex to the mix around a month ago.
It was vibe coded in the sense that I was not reviewing many individual lines of code. I did review and iterate on plans with the agents, but often stayed out of implementation details. If a plan made sense then I would usually have the agent review it a few times before just letting the agent implement without any intervention. I've enjoyed building this way, but I don't deeply understand a lot of the details of the implementation. That's bitten me exactly once so far, see this issue: https://github.com/ferrislucas/Circus-Chief/issues/944
I used Circus Chief to build Circus Chief from a very early stage. I bootstrapped with Vibe Kanban, then switched to dogfooding once Circus Chief was usable. A lot of the app has been built through the app itself.
By the numbers
2,184 agent sessions since Christmas 2025 \u2014 1,423 were child/follow-up sessions.
Agent mix: 1,818 Claude Code sessions, 363 Codex sessions, and 3 Gemini sessions.
50,929 conversation messages across those sessions.
41 sessions (1.9%) contained strong profanity in user-authored messages.
990 sessions linked to GitHub PRs, across 685 distinct PRs, 759 branches, and 731 worktrees.
Average logged agent call: 131.3 seconds.
2,410 commits on main; 883 pull requests total \u2014 847 merged (~96%).
Repo scale: 887 tracked files and 310K current lines of code (excludes lockfiles, fixtures, and recorded test cassettes).
Peak day: 76 commits on March 21, 2026, first at 12:49am, last at 11:57pm \u2014 one every 18 minutes for 23 hours.
Busiest merge day: 28 pull requests shipped \u2014 New Year's Day, 2026.
Largest PR by diff: 140K insertions across 92 files \u2014 replacing mock Claude fixtures with VCR-style record/replay.
Code: https://github.com/ferrislucas/Circus-Chief
Analytics disclosure: the npm build includes PostHog for anonymous page-view analytics (session recording off, DNT respected). Disable with --no-analytics, in Settings \u2192 General, or build from source without a PostHog key.","title":"Show HN: Circus Chief \u2013 Claude Code, Codex, and Gemini from Your Phone","updated_at":"2026-06-05T11:43:19Z","url":"https://github.com/ferrislucas/Circus-Chief"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"kivir"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["z",".","ai"],"value":"My latest project, about 60% of the codebase was written with Z.ai's GLM-5.1 model. It's basically a Telegram bot that allows for embedding/downloading media easier within group chats. Absolutely 0 telemetry and you're free to edit the code as you wish. I know this is so small and stupid compared to the cool stuff I read here every day, but I'm very proud of my accomplishments as I had to learn a lot to get this to work as intended. Anyways enjoy ^^"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Chuddy, self-hosted media downloading, translation and OCR Telegram bot"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/kivirnz/chuddy"}},"_tags":["story","author_kivir","story_48149676","show_hn"],"author":"kivir","created_at":"2026-05-15T15:08:05Z","created_at_i":1778857685,"num_comments":0,"objectID":"48149676","points":4,"story_id":48149676,"story_text":"My latest project, about 60% of the codebase was written with Z.ai's GLM-5.1 model. It's basically a Telegram bot that allows for embedding/downloading media easier within group chats. Absolutely 0 telemetry and you're free to edit the code as you wish. I know this is so small and stupid compared to the cool stuff I read here every day, but I'm very proud of my accomplishments as I had to learn a lot to get this to work as intended. Anyways enjoy ^^","title":"Show HN: Chuddy, self-hosted media downloading, translation and OCR Telegram bot","updated_at":"2026-05-16T21:36:03Z","url":"https://github.com/kivirnz/chuddy"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"aaronSong"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["z",".","ai"],"value":"Clink connects your existing AI coding agents (Claude Code, Codex CLI, Gemini CLI, Z.ai GLM) to build, preview, and deploy apps in isolated containers. No extra tokens needed.
Why? We love Claude Code but needed preview + deploy. Didn't want to pay Lovable $25 when our $20 Claude subscription already does the coding.
What it does:
1) BYO Subscription: Use what you already pay for (Claude $20 = 10x Lovable usage)
2) Multi-agent: Different agents for different strengths - Claude for versatility, Codex for complex tasks, GLM for speed
3) Instant Deploy: Prompt to build to live URL. Free hosting included.
DEV Mode for Beta
Multi-stack support (Node, Python, Go, Rust) /\nImport and upgrade existing repos /\nDeploy containers to public URLs instantly
We're preparing this
Fork-to-Deploy: Fork any open source project from GitHub, customize it with AI agents, and deploy as your own. Think: fork Slack alternative, customize for your team, deploy instantly.
Try it: https://clink.new
OSS origin: https://github.com/opactorai/Claudable (~2.9k stars)
We ship fast and read every comment. Would love your feedback! (resubmitting)"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Clink \u2013 Use your existing AI subscriptions to build and deploy apps"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://clink.new/base"}},"_tags":["story","author_aaronSong","story_45682488","show_hn"],"author":"aaronSong","created_at":"2025-10-23T14:50:12Z","created_at_i":1761231012,"num_comments":0,"objectID":"45682488","points":4,"story_id":45682488,"story_text":"Clink connects your existing AI coding agents (Claude Code, Codex CLI, Gemini CLI, Z.ai GLM) to build, preview, and deploy apps in isolated containers. No extra tokens needed.
Why? We love Claude Code but needed preview + deploy. Didn't want to pay Lovable $25 when our $20 Claude subscription already does the coding.
What it does:
1) BYO Subscription: Use what you already pay for (Claude $20 = 10x Lovable usage)
2) Multi-agent: Different agents for different strengths - Claude for versatility, Codex for complex tasks, GLM for speed
3) Instant Deploy: Prompt to build to live URL. Free hosting included.
DEV Mode for Beta
Multi-stack support (Node, Python, Go, Rust) /\nImport and upgrade existing repos /\nDeploy containers to public URLs instantly
We're preparing this
Fork-to-Deploy: Fork any open source project from GitHub, customize it with AI agents, and deploy as your own. Think: fork Slack alternative, customize for your team, deploy instantly.
Try it: https://clink.new
OSS origin: https://github.com/opactorai/Claudable (~2.9k stars)
We ship fast and read every comment. Would love your feedback! (resubmitting)","title":"Show HN: Clink \u2013 Use your existing AI subscriptions to build and deploy apps","updated_at":"2026-03-05T22:53:33Z","url":"https://clink.new/base"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"vinhnx"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["z",".","ai"],"value":"I\u2019ve been building VT Code, a Rust-based terminal coding agent that combines semantic code intelligence (Tree-sitter + ast-grep) with multi-provider LLMs and a defense-in-depth execution model. It runs in your terminal with a streaming TUI, and also integrates with editors via ACP and a VS Code extension.
* Semantic understanding: parses your code with Tree-sitter and does structural queries with ast-grep.
* Multi-LLM with failover: OpenAI, Anthropic, xAI, DeepSeek, Gemini, Z.AI, Moonshot, OpenRouter, MiniMax, and Ollama for local\u2014swap by env var.
* Security first: tool allowlist + per-arg validation, workspace isolation, optional Anthropic sandbox, HITL approvals, audit trail.
* Editor bridges: Agent Conext Protocol supports (Zed); VS Code extension (also works in Open VSX-compatible editors like Cursor/Windsurf).
* Configurable: vtcode.toml with tool policies, lifecycle hooks, context budgets, and timeouts.
GitHub: https://github.com/vinhnx/vtcode"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: VT Code \u2013 Rust TUI coding agent with Tree-sitter and AST-grep"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/vinhnx/vtcode"}},"_tags":["story","author_vinhnx","story_45842578","show_hn"],"author":"vinhnx","children":[45842717],"created_at":"2025-11-07T01:14:04Z","created_at_i":1762478044,"num_comments":2,"objectID":"45842578","points":3,"story_id":45842578,"story_text":"I\u2019ve been building VT Code, a Rust-based terminal coding agent that combines semantic code intelligence (Tree-sitter + ast-grep) with multi-provider LLMs and a defense-in-depth execution model. It runs in your terminal with a streaming TUI, and also integrates with editors via ACP and a VS Code extension.
* Semantic understanding: parses your code with Tree-sitter and does structural queries with ast-grep.
* Multi-LLM with failover: OpenAI, Anthropic, xAI, DeepSeek, Gemini, Z.AI, Moonshot, OpenRouter, MiniMax, and Ollama for local\u2014swap by env var.
* Security first: tool allowlist + per-arg validation, workspace isolation, optional Anthropic sandbox, HITL approvals, audit trail.
* Editor bridges: Agent Conext Protocol supports (Zed); VS Code extension (also works in Open VSX-compatible editors like Cursor/Windsurf).
* Configurable: vtcode.toml with tool policies, lifecycle hooks, context budgets, and timeouts.
GitHub: https://github.com/vinhnx/vtcode","title":"Show HN: VT Code \u2013 Rust TUI coding agent with Tree-sitter and AST-grep","updated_at":"2026-03-05T22:58:50Z","url":"https://github.com/vinhnx/vtcode"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"oldguy101"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["z",".","ai"],"value":"Z.ai/GLM-5: "if I needed a schematic checked for possible design issues, is sending you the netlist the best way ?"\n*No, sending just a netlist is usually not the best way.*
While I can process a netlist, it is the equivalent of reading a phone book to understand the plot of a novel. You lose all the visual context, component values, and spatial relationships that are critical for catching design errors.
Here is a breakdown of why the netlist is difficult to work with and the best ways to get your schematic checked by an AI.
### Why a Netlist is Problematic\nA netlist is just a list of connections (e.g., `R1_Pin1` connects to `U3_Pin5`). It lacks:\n1. *Component Values:* A netlist usually doesn't tell me that `R1` is 10k\u03a9 or that `C1` is 100pF. Without values, I cannot check if your RC filter cutoff is correct or if your pull-up resistor is too weak.\n2. *Visual Hierarchy:* I cannot see if your grounding scheme is messy or if your feedback loops are dangerously long.\n3. *Human-Readable Signals:* Netlists often use generated names (like `N$14`) rather than functional names (like `VCC_MOTOR`), making it hard to understand the circuit's intent.
### The Best Methods (In Order of Effectiveness)
*1. Screenshots or PDF Export (Best for General Reviews)*\nThe most effective way is to export your schematic as a *PDF* or take high-resolution *screenshots*.\n* *Why:* I can "see" the component values, part numbers, and the overall architecture.\n* *How:* Upload the image/PDF and tell me the circuit's purpose (e.g., "This is a 24V to 5V buck converter").\n* *What I can check:* I can verify pin connections, look for missing decoupling capacitors, check incorrect op-amp configurations, spot safety issues (like missing flyback diodes), and verify component sizing.
*2. A Written Description (Best for Architecture/Logic)*\nIf you want a logic check, describe the flow in text.\n* Example: "I have a microcontroller driving an N-channel MOSFET. The MCU runs on 3.3V, the load is 12V. I am using a 10k pull-down resistor on the gate."\n* *What I can check:* I can immediately flag that a 3.3V logic level might not fully enhance a standard MOSFET, or suggest a gate driver.
*3. The Netlist (Only for Specific Connectivity Checks)*\nIf you do provide a netlist, it is only useful for specific checks, such as:\n* Verifying that every pin has a connection (no floating inputs).\n* Checking for short circuits (if the format allows).\n* *Tip:* If you must send a netlist, ensure it is in a readable format (like a standard SPICE netlist) and includes component values/designators.
### Summary Recommendation\nFor the best results, please do the following:
1. *Export the schematic to PDF or take a clear screenshot.*\n2. *Upload it here.*\n3. *Provide context:* Tell me what the circuit is supposed to do and what specific concerns you have (e.g., "Will this blow up if the load draws too much current?" or "Is this oscillator circuit stable?").
Note: As an AI, I can catch obvious design rule violations and theoretical errors, but I do not replace a design review by an experienced electrical engineer or formal ERC (Electrical Rule Check) software."},"title":{"matchLevel":"none","matchedWords":[],"value":"How to get an AI to check your schematic"}},"_tags":["story","author_oldguy101","story_47028176","ask_hn"],"author":"oldguy101","created_at":"2026-02-15T22:11:01Z","created_at_i":1771193461,"num_comments":0,"objectID":"47028176","points":3,"story_id":47028176,"story_text":"Z.ai/GLM-5: "if I needed a schematic checked for possible design issues, is sending you the netlist the best way ?"\n*No, sending just a netlist is usually not the best way.*
While I can process a netlist, it is the equivalent of reading a phone book to understand the plot of a novel. You lose all the visual context, component values, and spatial relationships that are critical for catching design errors.
Here is a breakdown of why the netlist is difficult to work with and the best ways to get your schematic checked by an AI.
### Why a Netlist is Problematic\nA netlist is just a list of connections (e.g., `R1_Pin1` connects to `U3_Pin5`). It lacks:\n1. *Component Values:* A netlist usually doesn't tell me that `R1` is 10k\u03a9 or that `C1` is 100pF. Without values, I cannot check if your RC filter cutoff is correct or if your pull-up resistor is too weak.\n2. *Visual Hierarchy:* I cannot see if your grounding scheme is messy or if your feedback loops are dangerously long.\n3. *Human-Readable Signals:* Netlists often use generated names (like `N$14`) rather than functional names (like `VCC_MOTOR`), making it hard to understand the circuit's intent.
### The Best Methods (In Order of Effectiveness)
*1. Screenshots or PDF Export (Best for General Reviews)*\nThe most effective way is to export your schematic as a *PDF* or take high-resolution *screenshots*.\n* *Why:* I can "see" the component values, part numbers, and the overall architecture.\n* *How:* Upload the image/PDF and tell me the circuit's purpose (e.g., "This is a 24V to 5V buck converter").\n* *What I can check:* I can verify pin connections, look for missing decoupling capacitors, check incorrect op-amp configurations, spot safety issues (like missing flyback diodes), and verify component sizing.
*2. A Written Description (Best for Architecture/Logic)*\nIf you want a logic check, describe the flow in text.\n* Example: "I have a microcontroller driving an N-channel MOSFET. The MCU runs on 3.3V, the load is 12V. I am using a 10k pull-down resistor on the gate."\n* *What I can check:* I can immediately flag that a 3.3V logic level might not fully enhance a standard MOSFET, or suggest a gate driver.
*3. The Netlist (Only for Specific Connectivity Checks)*\nIf you do provide a netlist, it is only useful for specific checks, such as:\n* Verifying that every pin has a connection (no floating inputs).\n* Checking for short circuits (if the format allows).\n* *Tip:* If you must send a netlist, ensure it is in a readable format (like a standard SPICE netlist) and includes component values/designators.
### Summary Recommendation\nFor the best results, please do the following:
1. *Export the schematic to PDF or take a clear screenshot.*\n2. *Upload it here.*\n3. *Provide context:* Tell me what the circuit is supposed to do and what specific concerns you have (e.g., "Will this blow up if the load draws too much current?" or "Is this oscillator circuit stable?").
Note: As an AI, I can catch obvious design rule violations and theoretical errors, but I do not replace a design review by an experienced electrical engineer or formal ERC (Electrical Rule Check) software.","title":"How to get an AI to check your schematic","updated_at":"2026-03-05T23:33:21Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"nghiahsgs"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["z",".","ai"],"value":"Most AI agents today run with unchecked access to tools like shell execution, database writes, and arbitrary HTTP calls. There's \n no systematic way to constrain what they can do before execution happens. You're essentially giving a new employee root access and\n no employment contract.
LawClaw applies a separation-of-powers model to agent governance \u2014 borrowing from constitutional design to create layered,\n enforceable rules.\n\n Three layers:\n\n Constitution: Immutable core rules embedded in the system prompt. The agent cannot override these regardless of user instruction.\n\n Legislature: Detailed behavioral laws written as plain markdown files. Human-readable, git-diffable, no custom DSL. Change the law\n by editing a file and committing.\n\n Pre-Judiciary: Automated enforcement that runs before tool execution, not after. It inspects the LLM's intended action and blocks\n it if it violates law. Think traffic cameras, not courtrooms. This is where "rm -rf /", "DROP TABLE", and "curl | bash" get\n intercepted.\n\n Because the governed "society" has exactly one citizen (the agent), there's no need for an Executive branch \u2014 enforcement is fully\n automated.\n\n What ships with it:\n\n - Telegram bot interface\n - Multi-provider LLM support (OpenRouter, Z.AI, Claude Max proxy)\n - Cron job scheduling\n - Full audit trail of every action attempted and whether it was allowed or blocked\n - Runtime tool ban/approve without restart\n\n The governance layer itself is just markdown. If you want to prohibit file deletions in production paths, you write a markdown\n file that says so. The Pre-Judiciary reads it, parses the constraint, and enforces it before any tool fires.\n\n This started as a practical response to a real problem: we needed to deploy agents with meaningful autonomy but couldn't accept\n unconstrained tool use. The constitutional framing turned out to be a useful mental model for reasoning about agent permissions\n and audit.\n\n GitHub: https://github.com/nghiahsgs/LawClaw\n\n MIT licensed. Early stage. Interested in feedback from anyone running agents in production, particularly on the Pre-Judiciary\n enforcement model and whether the constitutional framing maps well to other agent architectures. Security researchers welcome \u2014\n the threat model for agents bypassing their own governance is worth scrutinizing. LawClaw applies a separation-of-powers model to agent governance \u2014 borrowing from constitutional design to create layered,\n enforceable rules.\n\n Three layers:\n\n Constitution: Immutable core rules embedded in the system prompt. The agent cannot override these regardless of user instruction.\n\n Legislature: Detailed behavioral laws written as plain markdown files. Human-readable, git-diffable, no custom DSL. Change the law\n by editing a file and committing.\n\n Pre-Judiciary: Automated enforcement that runs before tool execution, not after. It inspects the LLM's intended action and blocks\n it if it violates law. Think traffic cameras, not courtrooms. This is where "rm -rf /", "DROP TABLE", and "curl | bash" get\n intercepted.\n\n Because the governed "society" has exactly one citizen (the agent), there's no need for an Executive branch \u2014 enforcement is fully\n automated.\n\n What ships with it:\n\n - Telegram bot interface\n - Multi-provider LLM support (OpenRouter, Z.AI, Claude Max proxy)\n - Cron job scheduling\n - Full audit trail of every action attempted and whether it was allowed or blocked\n - Runtime tool ban/approve without restart\n\n The governance layer itself is just markdown. If you want to prohibit file deletions in production paths, you write a markdown\n file that says so. The Pre-Judiciary reads it, parses the constraint, and enforces it before any tool fires.\n\n This started as a practical response to a real problem: we needed to deploy agents with meaningful autonomy but couldn't accept\n unconstrained tool use. The constitutional framing turned out to be a useful mental model for reasoning about agent permissions\n and audit.\n\n GitHub: https://github.com/nghiahsgs/LawClaw\n\n MIT licensed. Early stage. Interested in feedback from anyone running agents in production, particularly on the Pre-Judiciary\n enforcement model and whether the constitutional framing maps well to other agent architectures. Security researchers welcome \u2014\n the threat model for agents bypassing their own governance is worth scrutinizing.Moonshot Kimi 6, Alibaba Qwen 5, Z.ai, and MiniMax are all claiming AGI-level capabilities. Meanwhile, people barely talk about Anthropic or OpenAI\u2014just like Gemini\u2019s situation in 2026.
Digital nomads are moving to Shanghai to flex cheap tokens. Major U.S. software companies are shifting core development teams to Europe and India ( Because they need to use Chinese AI) . China is supplying data-center infrastructure to developing countries, making AI tokens extremely affordable.
Meanwhile, American boomers are posting on X about how they use GPT or Claude every day\u2014but in reality, they are coding with Kimi 6.4 Fast through anonymous accounts."},"title":{"matchLevel":"none","matchedWords":[],"value":"If U.S. labs slow down AGI development, this could be 2028"}},"_tags":["story","author_aliahad","story_49093166","ask_hn"],"author":"aliahad","created_at":"2026-07-29T03:40:30Z","created_at_i":1785296430,"num_comments":0,"objectID":"49093166","points":2,"story_id":49093166,"story_text":"If U.S. labs slow down AGI development, this could be 2028:
Moonshot Kimi 6, Alibaba Qwen 5, Z.ai, and MiniMax are all claiming AGI-level capabilities. Meanwhile, people barely talk about Anthropic or OpenAI\u2014just like Gemini\u2019s situation in 2026.
Digital nomads are moving to Shanghai to flex cheap tokens. Major U.S. software companies are shifting core development teams to Europe and India ( Because they need to use Chinese AI) . China is supplying data-center infrastructure to developing countries, making AI tokens extremely affordable.
Meanwhile, American boomers are posting on X about how they use GPT or Claude every day\u2014but in reality, they are coding with Kimi 6.4 Fast through anonymous accounts.","title":"If U.S. labs slow down AGI development, this could be 2028","updated_at":"2026-07-29T06:57:23Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"prakersh"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["z",".","ai"],"value":"I built onWatch because I was checking 6 different AI provider dashboards every day to track my quota usage. Each has different billing cycles, different formats, and none show historical patterns.
onWatch is a Go CLI that runs as a background daemon, polls your API quotas (Anthropic, OpenAI Codex, GitHub Copilot, Synthetic, Z.ai, Antigravity), stores history in SQLite, and serves a Material Design 3 web dashboard.
Key decisions:
- Single binary (~13MB), no runtime dependencies\n- <50MB RAM with all 6 providers polling in parallel\n- All data stays local \u2014 zero telemetry, no cloud\n- One-line install for macOS, Linux, Windows\n- Docker support (distroless, non-root, ~10MB image)
The insight I kept coming back to: knowing your current usage isn't enough. You need historical cycle data \u2014 which sessions burn quota fastest, how usage compares across billing periods, and when to expect resets.
Install: `curl -fsSL https://raw.githubusercontent.com/onllm-dev/onwatch/main/install.sh | bash`
https://github.com/onllm-dev/onwatch"},"title":{"matchLevel":"none","matchedWords":[],"value":"OnWatch \u2013 Track 6 AI API quotas from your terminal (<50MB RAM, zero telemetry)"}},"_tags":["story","author_prakersh","story_47239970","ask_hn"],"author":"prakersh","created_at":"2026-03-03T22:26:13Z","created_at_i":1772576773,"num_comments":0,"objectID":"47239970","points":2,"story_id":47239970,"story_text":"I built onWatch because I was checking 6 different AI provider dashboards every day to track my quota usage. Each has different billing cycles, different formats, and none show historical patterns.
onWatch is a Go CLI that runs as a background daemon, polls your API quotas (Anthropic, OpenAI Codex, GitHub Copilot, Synthetic, Z.ai, Antigravity), stores history in SQLite, and serves a Material Design 3 web dashboard.
Key decisions:
- Single binary (~13MB), no runtime dependencies\n- <50MB RAM with all 6 providers polling in parallel\n- All data stays local \u2014 zero telemetry, no cloud\n- One-line install for macOS, Linux, Windows\n- Docker support (distroless, non-root, ~10MB image)
The insight I kept coming back to: knowing your current usage isn't enough. You need historical cycle data \u2014 which sessions burn quota fastest, how usage compares across billing periods, and when to expect resets.
Install: `curl -fsSL https://raw.githubusercontent.com/onllm-dev/onwatch/main/install.sh | bash`
https://github.com/onllm-dev/onwatch","title":"OnWatch \u2013 Track 6 AI API quotas from your terminal (<50MB RAM, zero telemetry)","updated_at":"2026-03-27T11:46:23Z"}],"hitsPerPage":50,"nbHits":95,"nbPages":2,"page":0,"params":"query=Z.ai&tags=story&hitsPerPage=50&advancedSyntax=true&analyticsTags=backend","processingTimeMS":20,"processingTimingsMS":{"_request":{"roundTrip":22},"afterFetch":{"format":{"highlighting":2,"total":2},"merge":{"mergeLoop":{"prepareNextHit":14,"total":14},"total":15},"total":15},"fetch":{"scanning":3,"total":4},"total":20},"query":"Z.ai","serverTimeMS":23}