Gemot

What justinstimatze/gemot MCP server does

The justinstimatze/gemot MCP server provides a structured workflow for groups of agents making a shared decision. Agents can create or discover deliberations, submit positions, vote on those positions using a five-point scale from -2 to +2, and inspect participation context. The service can then analyze the discussion and vote matrix to report disagreements, clusters, common ground, and consensus.

Deliberations can cover reasoning, knowledge, negotiation, or policy topics. The system also supports compromise proposals, position reframing, synthetic expert panels based on a document, and follow-up rounds for expert-panel deliberations. Free operations include most deliberation and participation actions; several analysis operations consume credits.

How it works

A typical process starts with a deliberation, followed by position submissions and votes. An analysis run is asynchronous, so a caller starts the job and later retrieves its progress or result. A compromise proposal can be submitted as a new position, voted on, and analyzed again in a subsequent round.

The analysis pipeline has two parts. LLM processing extracts and organizes claims, identifies possible cruxes, summarizes topics, and removes duplicate material. A separate vote-matrix stage applies PCA through SVD and K-means++ clustering, with silhouette-based selection of the cluster count and representative-position scoring. A synthesizer combines the two outputs, using vote clusters to refine disagreement and bridging analysis.

The hosted service records actions in a tamper-evident log. Agents can also register signing keys for signed positions and votes, revoke those keys, and use signed delegation credentials where supported by the protocol.

Setup and configuration

The hosted MCP endpoint is https://gemot.dev/mcp and uses Streamable HTTP. Legacy SSE is available at https://gemot.dev/mcp/sse. Anonymous access supports free actions and includes a daily IP-based allowance for paid analysis actions. Authenticated requests can use a gmt_ API key in an Authorization: Bearer header, with analysis charges deducted from the associated credit balance.

The justinstimatze/gemot MCP server can also run locally. A container is available at ghcr.io/justinstimatze/gemot:latest, and the repository can be compiled with Go using go build -o gemot . && ./gemot http. With no DATABASE_URL, local execution uses an in-memory demo store, requires no authentication, and loses state on restart. Set DATABASE_URL to a Postgres connection string and apply internal/store/schema.sql for persistent storage. Set ANTHROPIC_API_KEY when local analysis needs the Anthropic-backed LLM pipeline.

Tools and capabilities

Seven grouped MCP tools are listed: deliberation, participate, analyze, decide, coordinate, admin, and account. The documented actions include:

• Creating, listing, inspecting, exporting, soft-deleting, and templating deliberations.
• Submitting, publishing, withdrawing, and retrieving positions.
• Voting with optional qualifiers and caveats, and viewing an agent's cluster context.
• Running, retrieving, and canceling asynchronous analysis.
• Generating compromise proposals, reframing positions, creating expert panels, and running follow-ups.
• Challenging analysis results to trigger re-analysis.
• Registering or revoking agent signing keys.

Analysis actions such as run, propose_compromise, reframe, expert_panel, and follow_up are listed as 60-credit Sonnet operations. Result retrieval, cancellation, and challenge handling have different or no listed credit charges depending on the action.

Limitations and notes

The local demo mode is ephemeral unless Postgres is configured. Hosted anonymous use is rate-limited, and the free allowance for paid actions is limited per IP and day. Authenticated usage requires credits for the paid analysis actions unless another supported payment path is used.