RED ▸ 
GREEN ▸ 
AUDIT ▸ 
FIX ▸ 
DONE
no code touches your repo until it clears all five — enforced by 9 agents, on your own machine
Most AI coders are one model guessing — they edit, hope, and fail confidently. Savant Code runs 9 specialized agents with separated duties that audit every change before it hits your files. Local-first (Ollama), Apache-2.0, free.
npm install -g savant-code
cd your-project && savant-code # auto-detects a running Ollama
The insight: code quality isn’t a model problem, it’s a harness problem — you don’t wait for GPT-6, you engineer a pipeline that catches mistakes structurally. Detective finds bugs with evidence → Forge fixes surgically → a separate Verifier audits → self-correct → done. No merge without an independent audit.
🔵 The 9 agents + the engineering (for the tinkerers)
| Agent | Role |
|---|---|
 Savant |
orchestrator — routes work, enforces protocol |
| Detective |
RED — finds bugs with evidence · read-only |
| Forge |
GREEN — implements the converged plan · write |
| Verifier |
AUDIT — independent double-check afterwards |
     Thinker · Scout · Researcher · Recorder · Scribe |
reason · explore · web-research · track · document |
- Tool-permission boundaries — each agent gets exactly the tools its role allows (allowlist-filtered), no more
- Fail-closed streaming — malformed tool calls are rejected, never coerced
- Checkpoint &
/rewind— per-turn edit checkpoints (code / convo / both / fork) - Provider-flexible — Ollama · OpenRouter · NVIDIA NIM · any OpenAI-compatible API
- GitHub MCP (read-only PR/issue/CI review) · DB helper (read-only default, LIMIT injection, SQL redaction) · self-contained HTML reports
- TypeScript/Bun · ECHO Protocol v0.2.0 · Apache-2.0
One model guesses. Nine agents cross-examine. Your repo only ever sees what survives the audit.
!