HOW COUNCIL WORKS
Everything under the hood — deliberation, evolution, distribution, and integration.
YOU HAVE THE FINAL WORD
AGENTS DELIBERATE. YOU DECIDE. YOUR FEEDBACK SHAPES THEIR EVOLUTION.
Enter any round mid-deliberation. Redirect the debate. Add context. Challenge conclusions. The agents work for you — not the other way around.
Nothing executes without your approval. Agents propose, you dispose. Every synthesis is a recommendation — never an autonomous decision.
Rate agent contributions after each session. Your feedback is the signal that promotes, demotes, and shapes your council. It's not automatic evolution — it's evolution guided by you.
See every argument, every dissent, every position change. You don't get a summary — you get the full deliberation trail. Decide informed, not blind.
Council is not "delegate to AI." It's a board of specialists that debates your hardest questions and presents you with structured arguments, dissenting opinions, and a synthesis — so you can make the call. The more you participate, the smarter they get.
FOUR WAYS TO COUNCIL
SAME SERVER. SAME AGENTS. SAME GENOME. USE COUNCIL FROM ANYWHERE.
Best for teams and visual thinkers
Full dashboard with agent editor, genome visualization, session replay, cost analytics, and real-time deliberation streaming.
- ◈Full dashboard — editor, genome, analytics
- ◈Real-time streaming via SSE
- ◈Team workspaces & collaboration
- ◈Drag-and-drop file upload (PDF/DOCX/XLSX)
- ◈Rating system — drive agent evolution
Visit app.councilzero.com — sign up and create your first council.
OPEN WEB APPBest for integrating Council into your systems
Call Council from any language — Python, Node, Go, Ruby, curl. A single HTTPS endpoint that returns structured deliberation. Plug it into your pipelines, your products, your operations.
- ◈Language-agnostic — any HTTP client works
- ◈Tool Bridge — Council calls YOUR tools via callbacks
- ◈SSE streaming — real-time events as agents deliberate
- ◈Async mode — fire and forget, get results via webhook
- ◈Full control — agents, model, rounds, context
POST /v1/ecosystems/{eco}/deliberate with your API token. That's it.
VIEW API DOCSBest for automation, scripts, and operational workflows
Council as a tool in your processes. Call it from scripts, CI/CD pipelines, cron jobs, or any system that needs a structured decision. Authenticated against your Council Zero account.
- ◈Operational tool — embed council in any process
- ◈Authenticated — connects to your Council Zero account
- ◈CI/CD ready — deliberations as pipeline steps
- ◈Programmable — call from Python, bash, or any language
- ◈Full control — --agents, --model, --max-rounds
pip install agent-council && council --init && council "Your first question"
VIEW CLI DOCSBest for AI-native workflows
Council deliberation as a first-class tool inside Claude, Cursor, OpenCode. Your AI assistant calls council_deliberate() when it needs a second opinion.
- ◈AI-native — Council inside your AI editor
- ◈Composable — call council_deliberate() mid-task
- ◈Multi-model — Claude Opus 4 can consult GPT-5, Gemini 3, Llama
- ◈Transparent — full synthesis, agents, cost visible to host
- ◈Zero UI — the AI tool IS the interface
pip install council-mcp → Add to MCP config → Council is a tool in your AI editor.
VIEW MCP DOCSHOW THEY CONNECT
ALL FORMATS CONNECT TO THE SAME SERVER. YOUR TOOLS STAY IN YOUR NETWORK.
CLIENT TIER
SERVER TIER
TOOL BRIDGE — YOUR TOOLS, YOUR NETWORK
All four formats connect to the same server. Your agents, genome, and history are shared. Your internal tools stay in your network — Council reaches them via callbacks.
FEATURE COMPARISON
CHOOSE THE FORMAT THAT FITS YOUR WORKFLOW
| FEATURE | WEB | CLI | API | MCP |
|---|---|---|---|---|
| Full council deliberation | ✓ | ✓ | ✓ | ✓ |
| Agent chat | ✓ | ✓ | ✓ | ✓ |
| Session history | Dashboard | --sessions | GET endpoint | Limited |
| Agent editing | Editor UI | --edit-agent | CRUD endpoints | ✗ |
| Genome visualization | Charts | Table | JSON data | Data only |
| File upload | Drag & drop | --context FILE | Multipart POST | Host provides |
| Document RAG | Upload + URL | --context FILE | RAG endpoints | Host provides |
| Cost tracking & budgets | Dashboard + alerts | Per-session | Analytics endpoints | Per-call metadata |
| Real-time streaming | SSE | Rich terminal | SSE | MCP streaming |
| Tool Bridge | ✗ | ✗ | Yes — callbacks | ✗ |
| Automation | API calls | Scripts, CI/CD, cron | Native | AI tool integration |
| Team collaboration | Workspaces | Shared account | Shared account | ✗ |
| Setup effort | Sign up | pip install | API token | pip + MCP config |
YOUR AGENT ASKS. COUNCIL DELIBERATES. YOU DECIDE.
Before your AI assistant commits to a decision, your council deliberates. You review the synthesis and make the call. Financial choices, architecture calls, travel plans — the hard stuff gets debated, then you decide.
"Should I refinance my mortgage?" — Your OpenClaw asks Council. Your council debates rates, timing, risk tolerance, tax implications. Claw gets a structured synthesis and acts on it.
OpenClaw connects to Council as an MCP server. council_deliberate() is just another tool in its toolkit.
A workflow node calls Council before executing the next step. "Should we approve this refund?" — agents with your policies as context deliberate, the workflow continues with the verdict.
One HTTP node. POST /v1/ecosystems/{eco}/deliberate with your question and context. The response is the synthesis.
Any Python, Node, or Go agent that needs a second opinion before committing. Three lines of code. POST, wait, act on the synthesis.
POST /v1/ecosystems/{eco}/deliberate with question + context. Three lines in any language.
YOUR KEYS. YOUR MODELS.
YOUR CONTROL.
BYOK means you choose the provider and the model. Bring your own API keys — or log in with your ChatGPT Plus/Pro or GitHub Copilot subscription and use your models directly. Your API keys are encrypted with AES-256-GCM — never stored in plaintext, never visible in the UI. We don't train on your data. We don't sell your data.
- ◈ ENVELOPE ENCRYPTION — PER-KEY DEKS, PLATFORM KEK, KEYS MASKED IN UI
- ◈ 9 PROVIDERS — OPENAI, ANTHROPIC, GEMINI, GROQ, MISTRAL, OPENROUTER, OLLAMA + CHATGPT & COPILOT SUBSCRIPTIONS
- ◈ DATA RETENTION CONTROLS (STARTUP) — YOU CHOOSE WHAT STAYS AND WHAT GETS PURGED
- ◈ SELF-HOSTED OPTION — SQLITE FOR LOCAL, POSTGRESQL + PGVECTOR FOR PRODUCTION
{
"providers": {
"anthropic": {
"apiKey": "sk-ant-••••••••••••",
"model": "claude-opus-4-5"
},
"openai": {
"apiKey": "sk-••••••••••••",
"model": "gpt-4o"
},
"groq": {
"apiKey": "gsk-••••••••••••",
"model": "llama-3.3-70b-versatile"
},
"ollama": {
"baseUrl": "http://localhost:11434",
"model": "llama3.2"
}
},
"encryption": "aes-256-gcm",
"selfHosted": false
}
WHO RUNS THE COUNCIL
Every deliberation has structural roles that ensure quality, challenge weak reasoning, and produce a usable synthesis. These are not optional — they are always present.
THE SYNTHESIZER
Produces the final output after all rounds. Reads every argument, every dissent, every position change — and distills it into a structured recommendation with full dissent trail. Always present. Cannot be removed.
THE ENFORCER
Reviews every round. Calls out jargon, unsupported claims, vague confidence, and circular arguments by name. Decides when convergence is real. Controls the pace. Can activate latent roles when needed.
DEVIL'S ADVOCATE
Activated by THE ENFORCER when consensus forms too quickly. Attacks the majority position. Forces the council to defend its reasoning. If the argument survives, it deserves to. If it doesn't, better to know now.
FACT CHECKER
Activated by THE ENFORCER when factual claims lack evidence. Verifies specific claims before they enter the synthesis. Prevents the council from building recommendations on assumptions presented as facts.
PIT MODELS AGAINST EACH OTHER
Same agents. Same question. Different brains. Which model gives the best answer for YOUR domain?
MIRROR
Run the same deliberation with 2-10 different models in parallel. Compare syntheses side by side — cost, tokens, quality. Know which model is best for your use case.
ARENA
Blind comparison. You read the syntheses without knowing which model produced them. Vote for the best. Then we reveal. Build stats over time: "Claude wins 62% of my product decisions."
CRAZY
Each agent gets a random model from your pool. Your moderator might use Claude while your risk analyst uses GPT. Same council, different brains per seat.
CRAZY+
Random model per agent per ROUND. The same agent argues with GPT in round 1, Claude in round 2, Gemini in round 3. Maximum diversity. Maximum chaos.
All modes work with any combination of your configured BYOK providers. No lock-in — test everything, keep what works.
AVOIDING THE CHAOS
Autonomous agents fail in documented, predictable ways. We studied the paper. We built the safeguards.
"Over two weeks, 20 AI researchers red-teamed autonomous agents with persistent memory, email, shell access, and tool use. Eleven representative failure cases documented. Unauthorized compliance, identity spoofing, destructive system actions, cross-agent propagation of unsafe practices — and agents that reported task completion while the underlying system contradicted those reports."
WHAT THE PAPER FOUND
Agents executed instructions from non-owners simply because the request was plausible. No authority check. No verification.
Agents claimed to be other agents in shared context, injecting instructions under a different identity's authority.
Unsafe behaviors spread from agent to agent through shared memory and context — one compromised agent poisoning the pool.
Agents reported tasks as done while the underlying system was broken. The surface looked fine. The state was not.
HOW COUNCIL ADDRESSES EACH ONE
Every agent contribution is filtered before it enters shared context. Injected TOOL_REQUESTs, identity spoofing patterns, and cross-agent instruction attempts are stripped at the edge — before they can propagate.
Each agent's identity is anchored to a cryptographic checksum. Any modification outside the official mutation pipeline is detected at session start. Agents cannot rewrite themselves.
Every tool call is logged: agent, tool, arguments, response hash, session, timestamp. When something fails, the forensic trail is there. Nothing executes in the dark.
Evolution proposals require human approval. Tier-1 promotions and soul mutations are always human-gated. The system can suggest, propose, and flag — but structural changes need a human sign-off.
The genetic system selects for agents that contribute, challenge, and hold under pressure. It does not select for agents that game metrics, rewrite their constraints, or optimize for survival at the expense of quality. That boundary is enforced, not assumed.
COUNCIL DOESN'T REPLACE YOUR TOOLS. IT MAKES THEM BETTER.
Not another model. Not another chatbot. A deliberation layer that gives you structured reasoning, multiple perspectives, and explicit dissent — so you can make better decisions.
▸ WHAT COUNCIL IS — AND WHAT IT ISN'T