Ask one model something hard and you get one perspective — its blind spots baked in. Gang of 8 asks all of them at once. Every AI on your machine — the claude, codex, and gemini CLIs, plus any API seats you enable — convenes as a panel, each writing independently and in parallel. Then a lead verifies the claims, overrules what's wrong by name, and assigns the real work to specialists — then ships tested files.
No echo chamber. Every seat writes blind, then one lead verifies against the evidence, delegates the work, and closes the loop — before anything reaches your files.
All seats get the task plus pre-read project context and write their take independently and in parallel — none sees another's answer while writing.
A fast lead checks claims against real evidence, adopts what's right, overrules what's wrong by name, and assigns work to specialist talents.
On a build, every candidate is run headless — crashers disqualified — then scored blindly by judges. The winner is a real model's code, shipped byte-for-byte.
Tests pass first. Then a single diff-carrying approval — create vs. overwrite — stands between the council and your real folder. Nothing crosses without your click.
Ask for working software and the council designs it, delegates the authoring, runs the tests, and fixes its own failures before anything ships.
Every panelist writes in parallel, blind to the others. You get N real perspectives instead of N models agreeing with whoever spoke first — the diversity-of-intelligence bet the whole system is built on. Every contribution is tagged with the exact model that wrote it.
Candidates that crash on load are disqualified — judging by reading is blind to runtime failure. Only files that actually run get scored blindly. 🏆 winner ships unchanged.
The coder codes, the researcher researches, the critic checks. The lead recruits talents live — 🤝 code_generator ← claude — and a failed delegation is retried before falling back.
When RUNTESTS fails, the lead sees the real output and repairs the code — up to 3 attempts, all before promotion. A build ships passing its own tests, or says exactly why it couldn't.
Everything is bounded — budgets, wall-clock, depth caps — and everything is logged to a JSONL reasoning trail. Rotation is automatic but consent-gated: after three CONTINUE rounds it pauses and asks you. A run that fails says so honestly instead of dressing up a failure as an answer.
Pick a task. Watch the panel write in parallel, a crasher get disqualified, the judges score blindly, and one winner reach the promote gate — the real loop, simulated.
Choose a prompt and press Submit to the council. Nothing here calls a server — it's a faithful re-enactment of the panel → best-of-N → gate pipeline.
index.htmlHighest blind score on completeness, correctness, fidelity & robustness. Credited to the model that wrote it.
Side effects are governed by a data-driven skill registry and a permission kernel — the kernel decides on metadata, never on hardcoded behavior.
Writes, edits, test runs, staging, web lookups — all free inside the council's own scratch space. Every seat can land its work; role-gating never blocks council work.
Copying a finished file into your real folder is the only approval in the pipeline — with a diff that says whether it creates or OVERWRITES. Pick Approve all to grant a category for the session.
“Read from A, save to B” delivers to B and never touches source A. No known destination? It asks where at delivery time — never up front, never assumed.
Detection is the same PATH lookup your terminal does. Anything installed is found; anything absent is gracefully skipped. No config file, no external service.
# the CLIs manage their own auth .venv\Scripts\python cli.py serve --backend cli # or one-shot: .venv\Scripts\python cli.py submit "your task" --backend cli # dashboard opens at http://127.0.0.1:8790/
python -m venv .venv .venv\Scripts\python -m pip install \ pydantic fastapi uvicorn pytest httpx # default mock backend — try it now .venv\Scripts\python cli.py serve --backend mock
The CLIs authenticate themselves and the model dropdowns are fed by a public, no-key catalog. Keys only add capability: a Gemini key routes the gemini seat through Google's SDK (faster on Windows, image vision, Google-grounded search); an OpenRouter key unlocks the DeepSeek / GLM / Qwen / Kimi panel seats. Absent keys degrade gracefully — they never error.
The diversity of a committee, the decisiveness of a single owner, and the receipts of an audit trail — running entirely on your desk.