The economical coding agent.
Does 5× more work for the same tokens. Your subscription lasts longer. Free-tier models actually work.
Claude Code, Cursor, Copilot — they all optimize for capability. Nobody optimizes for cost. 3code is the only coding agent that treats your token budget as a first-class constraint.
That means your $20 subscription lasts a month instead of a week. It means you can use DeepSeek V4 or GLM-5.2 free tier and still get real work done. It means you don't have to think twice before asking a question.
Works with any OpenAI-compatible endpoint. Bring your own provider — free tiers, flat-rate coding plans, subscriptions, and local servers all work.
SWE-bench Verified, 10-task subset — five agents through the same LiteLLM proxy on Z.ai GLM-5.3, same provider, same 600s per-task cap, vanilla configs. Every difference in the results is the harness, not the model:
| agent | total tokens | % of median | resolved |
|---|---|---|---|
| 3code | 4,209,360 | 88.3% | 7/10 |
| pi | 4,641,357 | 97.4% | 6/10 |
| zcode | 4,766,000 | 100.0% | 6/10 |
| hermes | 6,072,610 | 127.4% | 6/10 |
| opencode | 6,771,747 | 142.1% | 6/10 |
3code used 9% fewer tokens than pi and resolved one more task. Output tokens are the starkest gap: 71,751 for 3code vs 117,707 for pi — a 1.6× difference, and on a pay-per-token API that's money on every single turn.
An earlier round on GLM-5.2 against opencode alone: 3code used 75% fewer tokens — 4× savings — and resolved a task opencode failed. Details in the blog post. Full per-task data and methodology → 3code.capocasa.dev/swe
- Chunked mode — constantly extracts relevant context and discards what's stale. Keeps the agent sharp without carrying dead weight, and improves results because of intelligent discarding.
- Aggressive caching — covers all bases; every cache hit is money saved.
- Context compaction — supersede-aware: later writes elide stale reads, shrinking context automatically.
- Self-clearing execution — plan/execute skill resets context between phases for larger tasks. No context bloat.
We eat our own dog food — 3code is now written entirely with 3code, using GLM 5.2 on Z.ai's free tier.
# macOS / Linux
curl -fsSL https://3code.capocasa.dev/install | sh
# Windows (PowerShell)
irm https://3code.capocasa.dev/install.ps1 | iex
# Termux on Android arm64
pkg install curl && curl -fsSL https://3code.capocasa.dev/install | sh
The Termux build runs on your phone: the installer detects $PREFIX and drops the binary in $PREFIX/bin. Android has no OS sandbox, but the in-process path checks still apply. Details in the manual.
- Get an API key from build.nvidia.com. NVIDIA Build has a free tier with several known-good coding models, no payment required, which makes it the recommended way to try 3code for the first time.
- Run
3codein a project directory. With no provider configured, the setup wizard starts by itself. - At the first prompt, enter
nvidiaand paste your key. - Pick a model from the list.
glm-5.3-flashanddeepseek-v4-flashare solid defaults. - Type a prompt:
❯ Write a Hello World in Nim and run it
That's it. When the free tier runs out, run :provider add to stack
another. The provider guide
covers free tiers, flat-rate coding plans, subscription logins, and local
servers.
- 1.6 MB binary — single executable, no runtime dependencies
- Cross-platform — Linux x86-64/arm64 · macOS universal · Windows · Termux (Android arm64)
- No daemon, no web UI — run it, use it, done
- Instant startup — loads and responds instantly; your agent should never keep you waiting
- 40 known-good combos — validated provider + model pairings, just works out of the box
- Loop guard — detects runaway autonomous edits, halts at configurable thresholds
- Session persistence — human-readable
.3logformat; resume any past session - Native web search — built-in, no curl dependency
- No telemetry — sessions stay local, nothing phoned home
- MIT license — do whatever you want with it
3code is also a Nim library: the same agent the CLI runs, embeddable in your own program with the terminal replaced by return values and callbacks. Sandbox, tool calls, session persistence, all of it. Build a web frontend, a chat bot, a CI runner that fixes its own failures, an IDE plugin — the agent loop, tool use, and sandboxing are done; you bring the interface.
The programming manual, with examples and API details, lives in the docs.
Patches welcome! Open an issue, send a PR, or just try the bleeding edge and report back:
# macOS / Linux
curl -fsSL https://3code.capocasa.dev/main/install | sh
# Windows (PowerShell)
irm https://3code.capocasa.dev/main/install.ps1 | iex
MIT.