llm-council
/llm-council
Zadaje pytanie kilku modelom LLM różnych dostawców przez OpenRouter, a model przewodniczący pisze ostateczną odpowiedź.
- Repozytorium
- llm-council ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 13.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 3
- Skrypty
- 1
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 3
Brak sygnałów. Repozytorium zmienione 13.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- 1 skrypt w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: scripts/council.mjs
Co robi i kiedy
Zadaje pytanie kilku modelom LLM różnych dostawców przez OpenRouter, a model przewodniczący pisze ostateczną odpowiedź.
Wyzwalacze z opisu
Oryginalny opis (EN)
Ask a council of LLMs from different vendors (by default GPT-5.6 Terra, Claude Fable 5, Gemini 3.1 Pro and Grok 4.5) through OpenRouter. Each model answers in parallel, then reviews and ranks the other answers anonymously, and a chairman model writes the final answer. Use when the user asks for the "llm council", "ask the council of LLMs", "ask the other models", "what do other LLMs think", "a second opinion from several models or vendors", "compare GPT, Gemini, Grok and Claude", or wants an important decision checked by models from different companies. It spends the user's OpenRouter credits.
Jak działa
- Purpose
- Instructions
- How to run
- What to do with the result
- Errors
- Config
- Test without spending credits
- Prerequisites
- Limitations
- Troubleshooting
Podgląd SKILL.md
---
name: llm-council
description: Ask a council of LLMs from different vendors (by default GPT-5.6 Terra, Claude Fable 5, Gemini 3.1 Pro and Grok 4.5) through OpenRouter. Each model answers in parallel, then reviews and ranks the other answers anonymously, and a chairman model writes the final answer. Use when the user asks for the "llm council", "ask the council of LLMs", "ask the other models", "what do other LLMs think", "a second opinion from several models or vendors", "compare GPT, Gemini, Grok and Claude", or wants an important decision checked by models from different companies. It spends the user's OpenRouter credits.
version: "0.3.6"
metadata:
author: Lucas Saldanha Werneck <lucas@2olhares.com>
tags: [council, llm, openrouter]
---
# LLM Council
## Purpose
Ask the same question to several LLMs from different vendors, let them rank each other's answers anonymously, and return a chairman's final answer with the ranking, cost and transcript path.
Runs the three stages of Andrej Karpathy's llm-council by calling the OpenRouter API directly, with no web app.
The script `scripts/council.mjs` does all of it. Don't rebuild the flow by hand.
## Instructions
1. Write the user's question, unchanged, to a temporary file.
2. Run `council.mjs --question-file <file>` in the background (see How to run).
3. Show the chairman's final answer, the ranking table, the cost and the transcript path.
4. On an error, follow the exit-code list under Errors.
## How to run
1. Write the user's question, in full and unchanged, to a new temporary text file with the Write tool. Don't summarize it.
2. Run the script with that file:
```bash
node "${CLAUDE_PLUGIN_ROOT}/skills/llm-council/scripts/council.mjs" --question-file "/path/to/question.txt"
```
- Never put the question text inside the shell command. Questions often contain `$(...)`, backticks or quotes, and the shell would run or change them. The file path is the only thing that goes in the command.
- `--question-file -` reads the question from stdin instead.
- Progress goes to stderr. The final result goes to stdout as markdown.
- Run it with the Bash tool's `run_in_background: true` and wait for the completion notification (don't poll, don't sleep). A run usually takes 1 to 3 minutes, but three rounds of calls of up to 240 s each, plus retries and a fallback, can pass the Bash tool's 10-minute foreground limit. The transcript is saved after every stage, so a run that is stopped still keeps the answers already paid for.
- Each run spends the user's OpenRouter credits. If the question looks trivial, ask the user before running.
## What to do with the result
1. Show the user the **Council final answer** (the chairman's synthesis) in full.
2. Show the **ranking table** (how the models rated each other's anonymous answers) and the line under it that says how many reviews gave a valid ranking. A review counts only if it ranks every answer exactly once. If none did, the script says the ranking is unavailable; say so too, don't invent an order.
If the final answer ends with a warning that it was cut off at the output token limit, tell the user.
3. Give the cost of the run and the path of the full transcript. Transcripts (every answer and every review) are saved in `~/.config/llm-council/sessions/`, or in `$LLM_COUNCIL_DIR/sessions/` if that variable is set.
4. If the user wants one model's own answer, read the saved transcript. Don't run the council again.
## Errors
- **Exit code 2 (no API key):** tell the user to set `OPENROUTER_API_KEY` in the environment Claude Code runs in, or to save it in `~/.config/llm-council/.env` as `OPENROUTER_API_KEY=...` with `chmod 600`. Keys are created at https://openrouter.ai/keys and need prepaid credits. Never ask the user to paste the key into the chat.
- **Exit code 3 (fewer than 2 members answered):** usually no credits left or an OpenRouter outage. The user can check the balance at https://openrouter.ai/credits. The answer that did come back, and each member's error, are in the partial transcript named on stderr.
- **Exit code 5 (chairman failed):** there is no final answer, but the ranking and the transcript are delivered. Say there is no synthesis, show the ranking, and point to the transcript for the individual answers.
- **Exit code 1:** bad config, missing question or unreadable question file. The message says which.
- **Exit code 4 (cannot save transcripts):** the sessions folder can't be created or written. Nothing was sent to OpenRouter and no credits were spent. Tell the user the folder from the message, or suggest setting `LLM_COUNCIL_DIR` to a writable folder.
- **"Transcript not saved" in the result:** the run finished and the answer and cost are valid, but the transcript file could not be written at the end. Tell the user; individual answers can't be read later for that run.
- **Model gone (404, or 400 "not a valid model"):** the script already tries the `fallback` model when one is set. If a model was removed from OpenRouter, change it in the user config (below). Rate limits (429) and server errors (5xx) are retried a few times before a member is dropped.
## Config
Defaults live in `${CLAUDE_PLUGIN_ROOT}/skills/llm-council/config.json`. Don't edit that file: plugin updates replace it.
To change models, write `~/.config/llm-council/config.json` with only the keys to override, for example:
```json
{ "council": [ { "model": "openai/gpt-5.6-terra" }, { "model": "deepseek/deepseek-v4-pro" } ] }
```
Or set `LLM_COUNCIL_MODELS` (comma-separated model ids) and `LLM_COUNCIL_CHAIRMAN`. Valid ids are listed at https://openrouter.ai/models.
`node "${CLAUDE_PLUGIN_ROOT}/skills/llm-council/scripts/council.mjs" --show-config` prints the config in effect.
## Test without spending credits
`--dry-run` fakes the answers and runs the whole pipeline:
```bash
node "${CLAUDE_PLUGIN_ROOT}/skills/llm-council/scripts/council.mjs" --dry-run --question-file "/path/to/question.txt"
```
## Prerequisites
- Node.js 18 or newer and an OpenRouter API key with prepaid credits (`OPENROUTER_API_KEY` in the environment, or in `~/.config/llm-council/.env`).
- Models and the chairman come from the default config; `~/.config/llm-council/config.json` overrides them.
## Limitations
- Each run spends the user's OpenRouter credits and takes 1 to 3 minutes, sometimes longer.
- A review counts toward the ranking only if it ranks every answer exactly once.
## Troubleshooting
- Error exit code 2: no API key. Exit code 3: fewer than 2 members answered (often no credits). Exit code 5: the chairman failed, so only the ranking is delivered.
- The full list of errors and fixes is under Errors.
## Examples
- User: "ask the llm council whether we should use Postgres or SQLite for this app". Claude writes the question to a file, runs the council in the background, then shows the final answer, the ranking table, the cost and the transcript path.
- User: "what do other LLMs think about this plan?". Same flow; the plan goes in the question file.
- User: "ask the council, but dry run first". Claude adds `--dry-run`: the pipeline runs with fake answers and no credits are spent.
Pliki w paczce (3)
- SKILL.md 7 KB
- config.json 1 KB
- scripts/council.mjs 18 KB
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · lucas-saldanha-werneck/llm-council | 100 | MIT |
| kirodotdev/kirocrew | 100 | Apache-2.0 |
| jjmartres/ai-coding-agents | 100 | MIT |
| UncleBenisimo/claude-kit-designeri | 100 | MIT |
| ray-aqno/toto-wolff | 100 | MIT |
| JZKK720/cubecloud-skills-bundle-kit | 100 | MIT |
| vishalsachdev/claude-skills | 100 | MIT |
| vishalsachdev/claude-code-skills | 100 | MIT |
| autohandai/community-skills | 100 | Apache-2.0 |
Migawka bazy z 13.11.2026 · wersja paczki 2f9569fbebc2 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100cross-review
Przegląda plan lub decyzję modelami zewnętrznymi: wysyła ten sam tekst równolegle do rady modeli i do Codexa, a potem łączy wyniki w jedną tabelę. - 100openrouter-cli
Pyta inne modele przez OpenRouter o radę, przegląda plany i kod oraz pokazuje ceny, wydatki i saldo konta. - 100council
Prowadzi ślepą radę na decyzję, dokument albo raport faktów: Claude, GPT i Gemini odpowiadają niezależnie, a przewodniczący Claude ocenia je bez wiedzy o autorze. - 100council
Zadaje to samo pytanie trzem modelom z różnych rodzin równolegle i tylko do odczytu, a potem ocenia odpowiedzi bez wiedzy o autorze. - 91ask-agents
Pyta wskazanego przez użytkownika agenta (Claude, Codex, Gemini lub Grok) i zwraca jego odpowiedź. - 95ask
Zadaje to samo pytanie kilku modelom naraz i zestawia odpowiedzi obok siebie, bez oceniania i bez konsensusu.