3code

The economical coding agent.

Get more done with less!

Make your AI plan last longer, lower your token bill, program locally faster. Enjoy yourself more with instant startup and a calm interface.

→ 3code.capocasa.dev

---

Why

A coding agent that works with any OpenAI-compatible endpoint. Bring your own provider! Free-tier and local models work, and subscriptions will last much longer.

Quickstart

# osx/linux
curl -fsSL https://3code.capocasa.dev/install | sh

# windows
irm https://3code.capocasa.dev/install.ps1 | iex

3code

First run walks you through adding a provider (name, URL, API key, models) and verifies with a test call.

Manual

The full documentation is available

https://3code.capocasa.dev/docs

Build from source

nimble install https://github.com/capocasa/3code

Requires Nim >= 2.0 and curl on PATH.

Changelog

0.5.0 - Windows support, new providers, big stability push

• Windows support. MSYS2 bash is the supported shell; the bash tool, session locking, and color palette all work on Windows. Linux CI is split into amd64 and arm64; macOS gets its own fast build+publish workflow.
• New providers and models. Ollama, Eurouter, Lyceum, Regolo, and TensorX in the provider catalog, alongside MiniMax (M3, M2.7), Tencent Hunyuan (hy3 prompt family), Longcat, GLM-4.7-Flash (free z.ai MoE), and Novita. Refreshed DeepSeek, gpt-oss, Kimi, and GLM reasoning prompts and per-(provider, model) context windows.
• Stable and reliable streaming. Network-quiet hangs and truncated SSE streams now time out and retry instead of freezing forever. Empty model replies are recovered via finish_reason-aware turn handling. Bounded streaming recv makes the quiet-network timeout actually fire, including at provider-connect time, where verifyProfile now uses the bounded streamhttp client instead of the unbounded httpclient that could deadlock. ctrl-c cancels mid-stream, mid-tool, and during provider connect, with no leftover freeze or stale echo. The caret no longer flickers when the streaming repaint races the input thread; assistant prose and the bash tool viewport carry their inter-item gap during live streaming; and a spurious timing line no longer prints on a mid-stream interrupt.
• Session resume and locking. O(1) resume via a per-cwd session index; resume replays the full session into the scrollback. Stale session locks are reclaimed automatically and locking is atomic on Windows. A prompt draft survives an unexpected shutdown.
• Terminal rendering fixes. Redraw the fat prompt on resize without stacking chrome or drifting the prompt. Single-GUI-thread ownership of the composite frame; spinner and bar-tick merged into one renderer (kills a thread leak that froze the bottom row). Tool banners, plan glyphs, and receipts now share one byte-path renderer across live streaming and session replay, with the old alternate renderers and dead plan code removed. Correct display width for CJK, emoji, and combining marks; unicode (UTF-8) input in the editor. Light/dark color mode auto-detected via OSC 11 background query, with [settings] mode and [colors] config overrides.
• Bash tool. Native timeout (no GNU timeout dependency, default 120s, ceiling 600s) and native computeDiff (no external diff). The model is told its own timeout. File contents shown for write-tool display.
• Robustness. Sanitize wire body so invalid UTF-8 can't brick a session; guard computeDiff against binary content; exit gracefully if the working directory is deleted mid-session; no silent exit on a broken stdout mid-turn. Network quiet timeout tightened from 180s to 45s.
• Packaging. --version/3code -v reports build provenance; nightly builds carry branch+commit in the version string.

0.4.0 - error icons for failed tool calls, pin bar+prompt to bottom during scrolling, suppress raw JSON on malformed tool args

0.3.5 - $/r/w tool bullets, bright cyan receipts, bar ticks during tool execution

0.3.4 - initial docs site, -i/--interactive flag, streaming ping test

0.3.3 - icon-based tool banners, history fixes, display polish

0.3.2 - native read command, binary guard for bash, deepseek in known-good

0.3.1 - update_plan tool, gpt-oss reasoning tuning

0.3.0 - --good subcommand, ctrl-c cancel during stream, linux-arm64 builds

0.2.7 - inline receipts, gpt-oss grounding prompts

0.2.5 - token bar with cache indicator, skill autoloader, web-research trigger

0.2.1 - per-model tool dispatch, multiline input, markdown table fit

0.2.0 - initial public release

Contributing

Patches welcome at github.com/capocasa/3code.

• Known-good provider/model pairs with test results
• Bug fixes with a clear reproduction case

Bug reports welcome, but make sure you give enough specific information to reproduce the issue.

License

MIT.