OmniCouncil – Zero-API-cost local MoA desktop app for Mac

OmniCouncil, a local-first Mixture-of-Agents desktop app for macOS, drives the official claude, codex and agy CLIs non-interactively using the user's own subscription sign-ins, so no separate API keys or per-token billing are required. The app runs workers in parallel in Judge Mode, where a Judge model receives anonymized and shuffled answers and returns a structured verdict with a consensus score of 1.0, 0.5 or 0, and triggers a second review by a model from a different vendor on low confidence or zero consensus. A persistent warm process pool skipped cold starts and was 31–36% faster end to end in the developer's tests, and the app warns users to check each provider's terms because requests count against their plan's usage limits.

A local-first Mixture-of-Agents desktop app for macOS. Ask Claude, Gemini and Codex at once. They answer independently or debate, and a Judge model returns one verdict with a measured consensus score. Important OmniCouncil drives the vendors' official CLIs claude , codex , agy non-interactively, signed in with your own accounts. Requests count against your plan's usage limits, and each provider's terms for this kind of use differ and can change. Check the terms of every provider you connect before using OmniCouncil. Different models get different things wrong. Asking several models and having a strong model judge the results can catch mistakes that one model alone would miss — and makes disagreement visible instead of hiding it behind one confident answer. - 🔌 No separate API keys or per-token billing. Every model call goes through an official CLI using the subscription sign-in you already have. API keys are removed from each CLI's environment env unset , so a call is never silently billed to an API account instead. - ⚖️ Two modes: Judge and Co-work - Judge Mode: workers answer in parallel without seeing each other. The Judge receives the answers anonymized and shuffled , and returns a structured verdict with a consensus score 1.0 🟢 / 0.5 🟡 / 0 🔴 and a confidence level. Low confidence or zero consensus triggers a second review by a model from a different vendor . - Co-work Mode: a multi-round discussion. Each round, every worker sees the other anonymous participants' answers and revises, adds to or keeps its own. If the Leader still finds a dispute, it issues guidance and the discussion gets extra rounds. - 🖥️ Native macOS GUI. A dark, minimal PySide6 app: the final answer is the main body of each reply, with every agent's raw answer, timing and the Judge's analysis folded away until you open them. Saved history, follow-up questions with context, attachments and voice notes, an English/Chinese interface and a global hotkey ⌘⇧J . | Warm process pool | Persistent CLI processes skip cold starts 31–36% faster end to end in our tests . Sessions are reset between runs, so questions never share context. | | Honest consensus | With only one usable answer, the UI says "Insufficient quorum · not cross-validated" instead of showing a perfect score. | | Prompt-injection hygiene | Worker answers and attachment text are passed to the Judge as length-capped, clearly delimited untrusted data . | | Multimodal input | Images, PDFs, text and voice notes. The Leader turns them into text first; workers only see plain text. | | Per-worker models | Switch any worker between fast and heavy models e.g. Haiku ↔ Opus from a dropdown. | | Account & usage dashboard | Sign-in status, 5-hour / 7-day usage, rate-limit alerts, and a log-in / switch-account button per provider. | | Terminal CLI | Everything also runs from the terminal omnicouncil ask … . | | | Three chat tabs | OmniCouncil | |---|---|---| | Ask every model | Copy-paste into each tab | One question, sent to all in parallel | | Compare answers | You read and judge them yourself | A Judge compares them, blind and shuffled | | See disagreement | Easy to miss | Measured consensus score, flagged in the UI | | Second opinion | Ask another model yourself | Automatic, from a different vendor, on low confidence or no consensus | | One model is down or rate-limited | That tab just fails | Isolated: the others continue, the failure is shown | | Follow-ups and history | Separate per tab | One thread, saved locally, reopened exactly | | Attachments | Upload to each tab | Parsed once by the Leader, shared as text | llm-council https://github.com/karpathy/llm-council popularized this idea: models answer, review each other's answers anonymously, and a Chairman model writes the final answer. OmniCouncil takes a different route on a few points: | | llm-council | OmniCouncil | |---|---|---| | Model access | OpenRouter API key, pay-per-token credits | Official CLIs with your existing subscriptions | | Interface | Local web app FastAPI + React | Native macOS desktop app + terminal CLI | | Review step | Every model ranks the others anonymized | One Judge scores consensus anonymized, shuffled ; a different-vendor reviewer on low confidence or no consensus | | Discussion | — | Co-work mode: multi-round debate with dispute guidance | | Speed | API calls | Warm pool of persistent CLI processes | | Also | — | Attachments, voice, history, usage dashboard, i18n, tests + CI | | Status | Stated by the author: "99% vibe coded", no support | Maintained; contributions welcome | php flowchart LR U User -- |question · files · voice| GUI "Desktop GUI<br/ omnicouncil.gui<br/ PySide6 + qasync" T Terminal -- TCLI "Terminal CLI<br/ omnicouncil.cli<br/ Typer + Rich" GUI -- E TCLI -- E subgraph E "Engine output-free, event-driven " direction TB PRE "multimodal<br/ Leader reads attachments" CTX "orchestrate<br/ context injection" ORC{"Mode"} PRE -- CTX -- ORC end ORC -- |Judge: parallel| W ORC -- |Co-work: N rounds,<br/ anonymous peers| W subgraph W "CLI Workers · subscription logins" direction TB W1 "claude" W2 "agy · Gemini" W3 "codex" end POOL "pool<br/ persistent processes" -.- W1 W -- |answers| BL "blind + bounded<br/ untrusted data" BL -- J "Judge / Leader<br/ JSON verdict + consensus" J -- |low confidence<br/ or no consensus| R "Second review other vendor <br/ or extra Co-work round" R -- V J -- V "Final verdict" V -- GUI V -- DB "SQLite<br/ data/history.sqlite" Every model call is a local subprocess, argv only never through a shell , with its own timeout. Workers that fail for example at a usage limit are dropped from later rounds without blocking the others. Closing the app terminates every child process. Does asking several models and judging the results actually help? A first pilot on 20 randomly sampled MMLU-Pro https://huggingface.co/datasets/TIGER-Lab/MMLU-Pro questions 10-option multiple choice, graded by exact match; workers: Claude Code default model, Gemini 3.8 Flash, Codex gpt-5.6-sol ; Judge: Claude Opus : | System | Accuracy | 95% CI | Median time / question | Model calls / question | |---|---|---|---|---| | Claude best single worker | 19/20 = 95% | 76–99% | 4.0 s | 1 | | Claude Opus alone the Judge model, no council | 18/20 = 90% | 70–97% | 4.8 s | 1 | | Majority vote of the three workers | 17/20 = 85% | 64–95% | 15.9 s | 3 | | Judge mode | 20/20 = 100% | 84–100% | 26.1 s | 4.1 | | Co-work mode first 10 questions | 9/10 = 90% | 60–98% | 65.7 s | 9.9 | What it suggests: - The council fixed both questions the Judge model got wrong on its own. On one, all three workers agreed on the right answer and the Judge followed them. On the other, no worker was right and they disagreed completely consensus 0 ; that triggered the different-vendor second review, which reached the correct answer on its own. - The Judge beat a plain majority vote 3 times and never overturned a correct majority. Once it sided with a correct minority answer against two agreeing workers. - It costs time and calls: ~4 model calls and ~26 s per question versus one call and a few seconds. Co-work was slower still and no more accurate here. This is a pilot, not proof: with 20 questions every confidence interval overlaps, and one sample is no substitute for hundreds of questions across domains. Full results: evals/pilot-2026-10/ https://github.com/OmniCouncil/OmniCouncil/blob/main/evals/pilot-2026-10 . Reproduce or extend it uses your subscription quota : python evals/run eval.py --n 200 --cowork 50 python evals/run eval.py --augment evals/results/<run /results.jsonl add a "Judge model alone" baseline - macOS the GUI, global hotkey and voice recording use macOS APIs; tested on macOS 15 . - Python 3.9+ 3.11+ recommended — 3.9 is past end-of-life . - At least one of these CLIs , installed and signed in with your subscription:Provider CLI Sign in Anthropic Claude Code https://claude.com/claude-code · claude claude auth login OpenAI Codex CLI https://github.com/openai/codex · codex codex login Google Antigravity https://antigravity.google CLI · agy run agy once and follow the promptsAgents whose CLI is missing are skipped automatically. Two signed-in CLIs are enough to compare answers. With pipx recommended — isolated, adds the omnicouncil and omnicouncil-gui commands : pipx install git+https://github.com/OmniCouncil/OmniCouncil.git omnicouncil-gui From a clone for development : git clone https://github.com/OmniCouncil/OmniCouncil.git cd OmniCouncil python3 -m venv .venv && source .venv/bin/activate pip install -e ". dev " python gui.py or: omnicouncil-gui As a macOS app built locally with Briefcase https://briefcase.readthedocs.io/ : pip install briefcase briefcase build macOS → build/omnicouncil/macos/app/OmniCouncil.app ~465 MB briefcase package macOS --adhoc-sign → a .dmg in dist/ The app is ad-hoc signed, not notarized: the first time, right-click it and choose Open . Launched from Finder, it picks up your login shell's PATH , so it finds the CLIs just like a terminal would. On first launch, config.json is created from the bundled template. Installed via pipx, OmniCouncil keeps its files in ~/Library/Application Support/OmniCouncil ; run from a clone, it uses ./config.json and ./data . OMNICOUNCIL HOME overrides both. - Type a question and press Enter Shift+Enter for a new line . - The settings panel chooses the mode, the workers, each worker's model, the Leader, and the second-review / extend-discussion option. - 📎 attaches files and 🎤 records a voice note; the Leader reads them before the workers start. - Click Process on any reply to see each agent's raw answer, timing, the Judge's analysis and which anonymous label was which model. - ⌘⇧J brings the window to the front from anywhere. omnicouncil ask "Which is larger, 9.11 or 9.9?" Judge Mode omnicouncil ask -m cowork -r 3 "Is a tomato a fruit?" Co-work Mode, 3 rounds omnicouncil ask -l "Gemini 3.1 Pro" "..." pick the Leader omnicouncil ask --model Claude=haiku -v "..." override a model, print commands omnicouncil ask -f invoice.png "What is the total due?" attachments omnicouncil config --show agents, models, install status From a clone without installing, use python main.py instead of omnicouncil . 1. Every enabled worker answers on its own, in parallel. 2. The Judge receives the answers anonymized "Assistant A/B/…" and in random order , each inside <answer tags that the prompt declares untrusted. It returns JSON: {"consensus score": 1.0 | 0.5 | 0 | null, "confidence": "...", "analysis": "...", "final answer": "..."} Score Meaning 1.0 🟢Every agent's core answer is the same 0.5 🟡A majority agrees, but at least one answer is clearly different 0 🔴No majority; the answers all differ — Only one answer: insufficient quorum, not cross-validated Malformed output is still parsed: from code fences, by matching braces, or with regex fallbacks. 3. If confidence is Low , or the answers show no consensus score 0 , a reviewer from a different vendor is picked automatically, and its verdict replaces the first. 1. Round 1: independent answers. 2. Rounds 2–N: each worker gets the question, its own previous answer and the other participants' answers under anonymous, per-run-stable labels, and revises, adds to or keeps its answer. 3. Leader summary blind, like the Judge . If confidence is still Low or there is no consensus, the Leader's dispute guidance becomes the focus of another round, up to cowork max rounds . config.json controls everything. Main fields: | Field | Description | |---|---| | workers .command | The CLI argument list; {prompt} marks where the prompt goes | | workers .available models / selected model / model flag | Model dropdown options; the flag is inserted right before the prompt | | workers .env unset | Environment variables removed for this CLI only e.g. ANTHROPIC API KEY , so Claude uses your subscription | | .web search | Let the agent search the web default on . Claude gets only the read-only WebSearch / WebFetch tools, pre-approved; Codex gets --search . agy has no safe switch opening pages needs a permission that non-interactive runs deny | | .is persistent | Run the agent in the warm pool supported: claude ; agy optional | | leaders .file flag / file types | How the Leader receives attachments and which kinds it can read | | mode | judge or cowork | | cowork rounds / cowork max rounds / cowork extend | Co-work round settings | | review on low confidence | Second review in Judge Mode | | hotkey | Global hotkey, e.g. cmd+shift+j or option+space | | language | en or zh | Custom prompts: every prompt is an editable file in omnicouncil/prompts/ https://github.com/OmniCouncil/OmniCouncil/blob/main/omnicouncil/prompts <name .<lang .md . To change one, copy it to <OMNICOUNCIL HOME /prompts/ with the same name and edit the copy; it overrides the built-in version. - OmniCouncil has no backend and stores its own data locally — history, attachments, recordings and the usage cache live in data/ or ~/Library/Application Support/OmniCouncil/data . - Your prompts, conversation context and attachment contents are still sent to the model providers you select , through their CLIs, under those providers' privacy terms. Local storage does not mean the data never leaves your machine. - No credential files are read: account status comes from each CLI's own status command. - Claude-based workers and Leaders get no file or memory tools — only the read-only web tools when web search is on. Only attachment pre-processing enables the Read tool, limited to the attachment folder. - With web search on, search queries are sent to the provider's search service as well. See SECURITY.md https://github.com/OmniCouncil/OmniCouncil/blob/main/SECURITY.md for the threat model and how to report a vulnerability. - Microphone voice notes : the first recording asks for permission for the app you launch OmniCouncil from Terminal, iTerm, VS Code… . If it doesn't ask, or recordings are silent: System Settings → Privacy & Security → Microphone , enable that app, and restart it. - Global hotkey: uses Carbon RegisterEventHotKey , which needs no Accessibility permission. If the combination is taken, choose another in config.json . pip install -e ". dev " pytest ~100 tests, all with fake CLIs — no model calls, no quota python -m pyflakes omnicouncil tests omnicouncil/ orchestrate.py Judge mode, blind answer formatting, context injection, reviewer choice cowork.py Co-work mode multimodal.py attachment pre-processing pool.py warm pool stream-json persistent processes runner.py one model call: pool or one-shot subprocess, structured-output adapters parsing.py verdict parsing with fallbacks prompts.py prompt loading prompts/ .md, user overrides config.py config loading, validation, migration storage.py SQLite history schema v2 accounts.py account status, usage limits, login helpers cli.py terminal CLI · cli render.py rich rendering gui.py desktop app · hotkey.py global hotkey · i18n.py UI strings evals/run eval.py accuracy / latency / cost evaluation tests/ pytest suite See CONTRIBUTING.md https://github.com/OmniCouncil/OmniCouncil/blob/main/CONTRIBUTING.md before opening a pull request. - Depends on the CLIs: their flags and output formats can change between versions; usage counts against your plan's limits. Respect each provider's terms of service. - Uneven pool support: codex has no stable persistent mode, so it always runs one-shot. - macOS only for now. MIT https://github.com/OmniCouncil/OmniCouncil/blob/main/LICENSE © 2026 OmniCouncil