cancho-code

An AI coding assistant that proves its work.

Most assistants hand you a transcript and ask you to read the diff. cancho-code hands you a verdict. The cancho compiler checks every piece the model writes, and a task is done only when the compiler and the task's own examples say so — never because the model says so.

[UNIT]            char_count   building
  attempt 0  refused   expected an expression, found `if`    <- the compiler's answer goes back to the model
  attempt 1  verified
[ISSUE_VERDICT]   verified     attempts 2
[PROJECT_VERDICT] done

• Describe it, get a plan, get a build — in one command. A brief becomes a plan of small units, each with a signature and examples; each unit is built and checked on its own. The functions your brief names keep their names.
• See everything. Every run keeps a journal (each prompt, reply, the compiler's full refusal, the diff between attempts, the plan stage). A read-only board shows every unit (waiting, building, verified, stuck); click one to read its whole story.
• Loops stop. Attempt ceilings and a same-refusal stop rule end a stuck run, named, instead of burning money.
• Any model. Claude (API key) or a local model through Ollama. You choose the provider and the model every time; there is no default.
• Measured (Claude Haiku 5.5; hidden tests the model never sees): 13 small libraries from hand-written plans finish 38 of 39 for $0.30 in total; from a plain-language brief alone, with the model writing the plan, 20 of 26 builds finish and 14 are fully right. Four small web services were built unit by unit; their final scores came after the guide was fixed on the failures it showed, so they show what is reachable, not a blind result. Conditions and caveats: eval/results.

Project page · Guide: run it end to end · Reference: every flag, contract and gate

Try it

Claude Code skill: copy skills/cancho-code into ~/.claude/skills/ (or your project's .claude/skills/) and the agent knows how to install cancho-code, run a build, read what happened and continue a stuck unit, with the key rules built in. Other agents can read the same file as plain instructions.

Or ask your coding agent to do it: paste docs/agent-prompt.md into Claude Code, Codex or similar. It installs everything, proves the install with a test that needs no key, and only then asks which model to use.

git clone https://github.com/alpibrusl/cancho-code.git && cd cancho-code
export CANCHO=<path-to-the-cancho-binary>      # the commit named in .cancho-rev
scripts/build-agent.sh && scripts/build-cancho-build.sh

# one brief in, a verified build out (Claude; the key is read from a file)
build/cancho-build "$CANCHO" brief.txt "$PWD/work" \
  "$PWD/scripts/cancho-agent.sh --key-file $HOME/.config/cancho-code/anthropic.key --primer $PWD/guidelines.md --max-tokens 16000" \
  --provider anthropic --model claude-haiku-5-5 --guide "$PWD/docs/plan-guide.md" --review on --review-guide "$PWD/docs/review-guide.md"

# watch or review it as a board
scripts/cancho-board.sh work && python3 -m http.server -d docs 8000     # then open /board.html

docs/build.md explains the command and the run directory, docs/journal.md what a run records, docs/dashboard.md the board. For a local model with no key, follow the guide.

More

Build command · Dashboard · Design · Plans · Projects · Epic and working rules

EUPL-1.2.