3code
Token-economy focus: lean prompts, aggressive caching, self-clearing plan/execute phases. Line-based REPL interface - no alternate screen, scrollback and tmux copy-mode stay intact. No telemetry. Measured 5.1M tokens on our suite at 9/10 resolved, the best score in the field.
| GitHub stars | 86 (updated September 2026) |
|---|---|
| Language | Nim |
| License | MIT |
| Bring your own key | Yes |
| Pricing | Free, open source; bring your own key across 40+ provider/model combos, free tiers work |
| Footprint | 3.2 MB static binary |
| Sandbox | Policy-enforced (in-process path checks) |
| Tokens on our 10-task suite | 5.1M (resolved 9/10) |
Repository: capocasa/3code · 3code.capocasa.dev
On our benchmark suite
10-task SWE-bench Verified subset, five agents through the same LiteLLM proxy on Z.ai GLM-5.3, same provider, 600 s per-task cap, vanilla configs. 3code measured 5.1M tokens, resolving 9/10 of tasks. Methodology and per-task data: full report.
Affiliation disclosure: this site is run by the 3code maintainer. The numbers above are the same raw benchmark data we publish for every harness — check our math.