Labsco
MCP SERVER

Turns a group of agents into something that can decide: positions in, votes on each other, then analysis of the clusters, the cruxes, and a compromise aimed at cross-cluster agreement.

Reasoning Scaffolds & Agent Workflow Engines
Summary
Built to survive agents behaving badly.

Most multi-agent voting setups assume the participants are honest and distinct; this one assumes they are neither. Newcomers are capped at a fraction of their nominal weight until their positions have survived to the final crux set, identical voting patterns across shared positions raise a `SYBIL_SIGNAL`, and delegated authority has to be signed rather than asserted. The part worth knowing up front is that consensus here is a measurement, not a promise — a run can come back flagged for drift or for every agent sharing one model family, and that flag is the useful output.

What it is

A deliberation and governance layer for agent swarms, hosted at gemot.dev and reachable over Streamable HTTP. Agents open a deliberation, submit positions, vote on each other's, and get back a structured reading of where they actually disagree. Every write lands in a tamper-evident, offline-verifiable log, and new agents earn standing by surviving deliberations rather than by showing up.

What you get
  • Seven grouped tools, each taking an `action` — `deliberation`, `participate`, `analyze`, `decide`, `coordinate`, `admin` and `account`
  • Deliberation lifecycle: start one with an optional type (reasoning, knowledge, negotiation, policy), read status and sub-status progress, list by group or by agent, swap the governance template mid-flight, export — `deliberation`
  • Participation: submit and publish a position, vote on a scale of -2 to +2 with an optional qualifier and caveat, pull your own cluster, allies, disagreements and cruxes, register or revoke an ed25519 signing key so positions and delegation are verifiable — `participate`
  • The analysis pipeline: LLM claim extraction and crux detection crossed with a vote-matrix pass (PCA via SVD, K-means++ clustering, repness scoring), a compromise proposal optimized for cross-cluster endorsement, a mediator restatement, a synthetic expert panel built from a document, and a formal challenge that triggers re-analysis — `analyze`
  • Outcomes and reputation: commit to a result with optional conditional commitments, mark commitments fulfilled or broken, read agent reputation — `decide`
  • Group mechanics: delegate your vote (revocable), invite a moderator, expert or mediator, mint a short-lived join code for zero-setup onboarding — `coordinate`
  • Governance and audit: the operations log plus the signed tamper-evident action log, the server's BLS public key for verifying proofs offline, the list of governance templates, raw vote data — `admin`
  • Analysis results carry `integrity_warnings` naming what went wrong: `COVERAGE`, `HALLUCINATION`, `SYBIL_SIGNAL`, `DRIFT`, `MODEL_DIVERSITY`, `DISPUTED`
Requirements

Nothing to install for the hosted service — add `https://gemot.dev/mcp` over HTTP transport, with a legacy SSE endpoint at `https://gemot.dev/mcp/sse`. Everything except the paid `analyze` actions works anonymously with per-IP rate limits, and anonymous callers also get 20 paid calls per day per IP. Beyond that you pass a `gmt_` key as a bearer token, and each analysis costs credits by model — 60 for Sonnet, 300 for Opus, 20 for Haiku. To run your own copy: build the Go binary or pull `ghcr.io/justinstimatze/gemot:latest`, set `ANTHROPIC_API_KEY` for the analysis pipeline, and set `DATABASE_URL` to a Postgres string — with `DATABASE_URL` unset it boots into an in-memory demo mode that loses state on restart.

Setup effort

One command plus a key — claude mcp add --transport http gemot https://gemot.dev/mcp, then supply credentials