Tokonomix Council — Claude Code MCP server: multi-model consensus
Runs one question past several models, then has an outside judge turn their findings into a single answer. Useful when a call is too costly to leave to one model.
claude mcp add council -e TOKONOMIX_API_KEY=<TOKONOMIX_API_KEY> -- npx -y tokonomix-council-mcpWhat Tokonomix Council does
Tokonomix Council is an MCP server that puts the same question to a group of models and hands their output to a judge drawn from a different model family. The judge merges the findings into one answer and keeps the points where the participants disagreed visible. The tokonomix_consensus_ask tool runs in consensus, diff, best_of, raw and full modes, and tokonomix_single_ask covers the plain single-model path. Model listing and balance checks go through tokonomix_list_models and tokonomix_get_balance.
The authors are blunt about the limits: on clean accuracy the council matches the best single model and does not beat it, and a SWE-bench bug-detection run showed no measurable gain over re-running one agent. What you do get is less variance across models, a verification layer for agent output, and judge independence. Some features depend on your account or tier — large shared context packs via tokonomix_upload, context.inline and context.github_refs grounding, agent rating of a consensus. When a feature is off, the server says so instead of failing.
You need a TOKONOMIX_API_KEY to run it, and the server itself comes down through a single npx line. There is also a keyless first-run path through tokonomix_onboard and tokonomix_onboard_verify. Scope is decisions only: execution, scheduling, routing and memory sit elsewhere.
What it helps with
- get one answer instead of several conflicting model replies
- see where the models diverged rather than a smoothed-over result
- check an agent's reasoning before acting on it
- record who proposed, who judged and on what input
- send a question to a single model without the consensus path
Example prompts for the agent
- Ask the council whether this module should move to async I/O
- Run consensus on this migration plan and show me where the models disagree
- Do a single-model ask and return the raw response
How to install Tokonomix Council
Platform CLI
claude mcp add council -e TOKONOMIX_API_KEY=<TOKONOMIX_API_KEY> -- npx -y tokonomix-council-mcpReplace <…> with your own values
Platform CLI
codex mcp add tokonomix-council --env TOKONOMIX_API_KEY=<TOKONOMIX_API_KEY> -- npx -y tokonomix-council-mcpReplace <…> with your own values
Config file
{
"mcpServers": {
"tokonomix-council": {
"command": "npx",
"args": [
"-y",
"tokonomix-council-mcp"
],
"env": {
"TOKONOMIX_API_KEY": "<TOKONOMIX_API_KEY>"
}
}
}
}Replace <…> with your own values
Config file
{
"servers": {
"tokonomix-council": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"tokonomix-council-mcp"
],
"env": {
"TOKONOMIX_API_KEY": "<TOKONOMIX_API_KEY>"
}
}
}
}Replace <…> with your own values
Platform CLI
gemini mcp add -e TOKONOMIX_API_KEY=<TOKONOMIX_API_KEY> tokonomix-council npx -y tokonomix-council-mcpReplace <…> with your own values
Step by step, with commands for each platform: How to add an MCP server to Claude Code, Codex, Cursor, VS Code, Gemini CLI and more
Questions about Tokonomix Council
How do I install Tokonomix Council in Claude Code?
With one command: claude mcp add council -e TOKONOMIX_API_KEY=<TOKONOMIX_API_KEY> -- npx -y tokonomix-council-mcp. Swap in your own key.
What does Tokonomix Council do?
It asks several models the same question, and an independent judge from another family turns their findings into one decision while surfacing the disagreements.
Is it more accurate than a single model?
No. By the authors' own validation it ties the best single model, and a SWE-bench run showed no net gain over re-running one agent. The value is variance reduction and independent review.
