magi-councilCAUTION
MAGI council of three AI heads (Melchior/Balthasar/Casper) that deliberate, vote and execute plans — local-first multi-agent deliberation with a memory graph, worktree-isolated production mode and auto-merge. Evangelion-style web UI; swappable providers (kimi/codex/claude/Ollama).
Overview
From the repository's own README, as read at the audited commit. Badges and raw HTML are left out.
[](https://github.com/Adrian-Sandwich/magi-council/actions/workflows/ci.yml)
A MAGI council of three heads (Melchior / Balthasar / Casper — the persona lives in the seat, the provider is swappable: kimi, codex, claude, Ollama, any CLI or OpenAI-compatible API) that deliberate on your questions and vote on decisions, with a ruling, minority report, mind changes between rounds and human arbitration. With memory: a local knowledge graph that ingests your agent sessions and decisions and injects them into every deliberation. And in production mode: the council approves a plan, an executor implements it on a dedicated branch, the council reviews the diff and — on unanimous approval — it merges itself.
The UI is a MAGI-style web app (design based on TomaszRewak/MAGI) with a single text box, like a CLI. Runs on Windows, macOS and Linux.
What you need to get started (from scratch, on any machine)
- Python 3.14+.
- A reachable Postgres (any standard install; the system default is
dbname=debate host=localhost, port 5432, user = your OS user). If you have none, on Windows the repo ships a portable postmaster in experiments/pg and bin\start-magi.bat starts it.
- At least one head (with no heads the system runs "degraded" and only
you can close decisions):
- kimi (or another MCP-capable agent CLI: claude, ...) — investigates
the repo with tools and votes via MCP;
- codex or another plain-text CLI — votes with the journal inlined in
the prompt (journal: "inline" mode);
- API heads:
type: "api"withprovideranthropic (Messages
API, claude-*), openai or moonshot (kimi, OpenAI-compatible), the key in the environm
960e72a2079eOBSERVED · 2026-10-12Exposed tools (7)
5 read · 2 write · 0 destructive.
| Tool | Risk | Description |
|---|---|---|
cast_position | read | Voto estructurado de una cabeza: publica el razonamiento en el journal |
get_decision | read | El dossier de una decisión: estado, ronda, ruling, confidence, |
list_threads | read | Lista los threads con cantidad de mensajes y timestamp del último. |
post_message | write | Publica un mensaje en el thread y devuelve su id. |
read_thread | read | Lee mensajes de un thread con id > since_id, ordenados por id. |
start_decision | write | Abre una decisión MAGI: las cabezas la investigan, debaten en el |
wait_messages | read | Espera mensajes nuevos (id > since_id) hasta timeout_secs. |
Trust audit
CAUTIONgrade B · trust 89/100 Install with care. The audit found things worth knowing before you trust its output.
| Layer | What it checks | Result |
|---|---|---|
| L0 | Provenance & inventory | PASS |
| L1 | Static analysis of the code | WARN |
| L2 | Instruction surface (what it tells the agent) | PASS |
| L3 | Class-specific surface | PASS |
| L4 | Behavioural (sandbox) | SKIPPED |
What the source does
- Filesystem
- none-observed
- Network
- declared (6 observation(s))
- Shell
- declared (1 observation(s))
- Dependencies
- pinned
- Secrets in source
- none-found
Findings (7)
__import__(module)
print(f"[magi_ui] MAGI System en http://127.0.0.1:{PORT} — token de sesión: {TOKEN}")UI = "http://127.0.0.1:8051"
Then open **http://127.0.0.1:8051**.
"base_url": "http://127.0.0.1:11434/v1", "timeout_secs": 600 }
┌─────────────────────────────── http://127.0.0.1:8051
.venv/bin/python magi_ui.py # http://127.0.0.1:8051
Gates applied: no_behavioural_pass.
960e72a2079efull audit observations/trust-audit/mcp-server/adrian-sandwich__magi-council.json · Report an issue / request a re-scanAudit history
Every audit this server has had. A grade with a past is a grade somebody is still checking.
| Date | Source | Verdict | Grade | Score | Change |
|---|---|---|---|---|---|
| 2026-10-12 | 960e72a2079e | CAUTION | B | 89 | first audit |
Questions
What is the magi-council MCP server?
MAGI council of three AI heads (Melchior/Balthasar/Casper) that deliberate, vote and execute plans — local-first multi-agent deliberation with a memory graph, worktree-isolated production mode and auto-merge. Evangelion-style web UI; swappable providers (kimi/codex/claude/Ollama).
What tools does magi-council expose?
7 in total: 5 read-only, 2 that write, and 0 that can delete or overwrite. Every one is listed on this page with its risk.
Is magi-council safe to connect to an agent?
With care. The audit graded it B (89/100) and found 7 things worth knowing before you trust this server, listed below with the exact line each was found on.
What credentials does magi-council need?
No credential environment variables were found in its source, so it appears to need none.
How current is this page?
The grade is for one exact copy of the source (960e72a2079e), read on 2026-10-12. The repository is watched and re-audited when it changes.