Tokonomix Council

What Tokonomix Council MCP server does

Tokonomix Council MCP server provides a decision-review layer for MCP clients. It sends the same question to multiple independent proposer models, then uses a separate judge to reconcile their findings. The result can include a synthesized answer, differences between responses, or the underlying model outputs, depending on the selected mode.

The service is intended for consequential decisions where exposing disagreement and checking claims against source material matter. Examples include security analysis, software migration decisions, GDPR interpretation, legal-clause review, and other work where a confidently incorrect single-model answer could be costly. It is not an execution engine, scheduler, router, memory system, browser automation tool, or general-purpose chat service.

How it works

A request can be evaluated by two to six frontier models from different providers. Their responses are produced independently, after which a cross-family judge assesses and synthesizes the findings. This structure is designed to reduce dependence on whichever single model happened to be selected and to preserve disagreements that a single answer would hide.

Grounding is available through inline context and GitHub references, and vision input can be supplied with an ask. Eligible accounts can also use uploaded shared context packs through session or handle references. These mechanisms let the council inspect an artifact rather than relying only on a prompt. The README describes grounding-gate behavior as a gated rollout, so availability can depend on the account or platform state.

Setup and configuration

The package is published as tokonomix-council-mcp, and the project documents installation through a single npx line. It works with Claude Code, Cursor, Cline, Continue, and Zed. The material provided does not specify required environment variables or a complete client configuration block, so those values should be taken from the repository or package documentation rather than inferred here.

Keyless first-run onboarding is available through tokonomix_onboard and tokonomix_onboard_verify. Some features are account- or platform-gated, and unavailable gated tools return a not-enabled response rather than behaving as if the feature were active.

Tools and capabilities

• tokonomix_consensus_ask with consensus, diff, best_of, raw, and full modes.
• tokonomix_single_ask for a single-model passthrough.
• tokonomix_list_models and tokonomix_get_balance for account and model information.
• tokonomix_skill_version and tokonomix_get_skill for skill metadata.
• tokonomix_rate_consensus for agent ratings and feedback, where enabled.
• tokonomix_relay_human_feedback for a human-verdict channel, where enabled.
• Vision input, inline or GitHub-reference grounding, and account-gated context uploads.

Limitations and notes

The Tokonomix Council MCP server does not claim higher clean-task accuracy than the best individual model, and the supplied material reports no measurable net improvement in one SWE-bench bug-detection run. Consensus can add cost and noise on routine or already-easy work. Agreement is not proof of correctness, especially when models share blind spots, so grounding the request in the actual file, diff, specification, or logs remains important.

Reranking and a web-search plugin are listed as planned rather than live. Context uploads, grounding gates, ratings, and human feedback may be restricted by account or platform tier. Council handles decisions; execution, scheduling, routing, durable memory, and verification of generated media belong to other components in the wider ecosystem.