Karajan Code
Code with several AIs, with guarantees.
The governor of coding agents: your agent writes; Karajan enforces TDD-first, cross-AI review on every commit and git gates. A false green is impossible.
governs claude-code · codex · copilot · antigravity · kimi · qwen · opencode · aider · gemini — one writes, a different one reviews, a third arbitrates
v4 · the Karajan Environment
Paste it into your AI assistant (Claude Code, Codex, Cursor, VS Code). It installs Karajan, prepares the project and opens a single window with the board and your agent inside — you never touch the terminal. It only stops if a step needs your permission.
Paste it into your AI agent (Claude Code, Codex, Cursor, VS Code). It checks your machine, installs what is missing, and stops to wait for you whenever a step needs your permission — you stay in control.
Want to see what gets installed, or set it up yourself first? Requirements and manual install.
Already installed? To start a project the Karajan way, paste brief.md into your agent. Not sure how to describe what you want? Read how to write a spec.
The method
The Karajan method, in seven pieces
Mandatory cross-review
A different AI reviews every change, bound to the exact diff hash. Change a line and skip the review: the system rejects the commit, even if the agent tries to bypass it.
read → 02Rules as mechanism
Norms live in git guards and project policy, not the prompt. It passes or the commit is refused. Security is non-negotiable.
read → 03The Sentinel
A watchman inside the session blocks the forbidden action the moment it happens — and links you to the rule.
read → 04Claims with a source
Every fact the agent publishes is checked against what its own session ran. Disproven by its own terminal? Rejected.
read → 05Hash-chained ledger
Every command, verdict and exception is sealed, each entry carrying the fingerprint of the last. The past can't be rewritten unseen.
read → 06Trust expires
The Steward re-verifies the project's guarantees on a schedule; stale evidence stops counting. Projects rot when nobody looks.
read → 07Project memory (RAG)
The agent queries the indexed code before it assumes — with a sensitivity policy, and secrets excluded from the index.
read →Mechanism where there was opinion. Evidence where there was a claim.
The family
One method, one family of tools
Around Karajan Code sits a family that shares the same method — context, vigilance, governance and the ground they all stand on.
Want to know more?
- Nine agents under one government — Claude Code, Codex, GitHub Copilot, Antigravity, Kimi Code, Qwen, OpenCode, Aider, Gemini. Any of them is the brain and absorbs the roles (triage, planner, architect, tester, security, audit) via
kj brief; when the reviewer runs out of quota, the panel fails over with a loud notice — never silence, never self-review. See the method → - How the government works: three tiers — at tool time the Sentinel denies the violation BEFORE the damage; at commit time no diff enters without a cross-AI verdict bound to its sha256 (the guarantee floor on any host); at merge time CI re-checks the same policy, so a tampered local hook changes nothing. Security findings are never overridable — not even by arbitration. See the gates →
- Cross-AI review on every commit — a different AI approves each diff; a third one arbitrates disputes; security findings nobody overrides. See the gates →
- Headless pipeline for CI — the classic multiagent loop, unattended, same gates. See headless mode →
- 6,000+ tests. 90+ releases. Open source (AGPL-3.0). See the repo →