mcpField
Q
quorum

quorum

Remote
github.com/terradev-cloudAI & Agents0Updated 5w ago
Repository

About quorum

Quorum is an MCP server for multi-agent proposals, shared deliberation, voting, and consensus outcomes. It suits orchestrators that need configurable social-choice algorithms and an immutable, attested record of each decision.

Key features

  1. 01Proposal creation and shared blackboard
  2. 02Append-only context with ordered history
  3. 03Multiple voting and aggregation algorithms
  4. 04Quorum and deadline-based resolution
  5. 05Attested, tamper-evident decision history
  6. 06Telinea trace emission

Use cases

  1. 01Collecting agent votes on a proposal
  2. 02Sharing analysis before agents vote
  3. 03Applying a selected consensus algorithm
  4. 04Reviewing an auditable decision record

MCP support

Quorum exposes proposal, write, read, vote, resolve, history, and account-registration tools to MCP clients. It attests proposal events and can emit traces to Telinea.

Configuration

Remote servers need no local install. Pick your client, then copy the config or install in one click.

Endpointhttps://quorum-mcp.terradev.cloud/mcp
Copy install command
Claude CodeTerminal
Cursor.cursor/mcp.json
VS Code.vscode/mcp.json
Claude Desktopclaude_desktop_config.json
Windsurf~/.codeium/windsurf/mcp_config.json

Run commands in a terminal; paste JSON into the client’s MCP config and restart. Servers that need sign-in will prompt on first connect.

Tools

7
ParameterTypeRequiredDescription
api_keystringYesYour Quorum API key -- chosen by you, presented on propose, never stored.
telinea_keystringYesYour Telinea API key -- stored encrypted, used to authenticate span pushes.

Details

From the project README

Quorum

Attested multi-agent consensus with a shared blackboard — an MCP server.

An orchestrator opens a proposal, participating agents write context to a shared blackboard and vote with reasoning, Quorum applies a principled aggregation algorithm from the social choice literature, and every step — proposal, writes, votes, outcome — is attested via Stamp and traced to Telinea. The record is immutable, tamper-evident, and independently verifiable.

Hosted endpoint (no sign-in, public): https://quorum-mcp.terradev.cloud/mcp


Why

Majority vote among LLM agents fails on disputed questions because agents trained on overlapping data share systematic blind spots — they are not the independent voters the Condorcet Jury Theorem requires. And when agents do decide together, the decision process — who voted, what they said, what context they worked from — is invisible.

Quorum is the complete primitive: algorithm plus state plus provenance.

Tools

  • register — one-time account setup: bind your api_key (a Quorum key you choose) to your telinea_key. The Telinea key is stored encrypted under a key derived from your Quorum key (AES-256-GCM, HKDF) — the Quorum key itself is never stored. Every span for your proposals is then pushed with your Telinea key, standard OTLP bearer auth.

  • propose — open a proposal: name (becomes the proposal id), question, voters (expected identities), deadline_minutes. Optional: api_key (from register — binds the proposal to your Telinea account for span streaming; omit for anonymous use), namespace (tenant prefix — the id becomes ns:name, so different teams can reuse names without colliding), options (default ["yes","no"]), algorithm (default approval), quorum (min % of voters, default 100), threshold (supermajority share, default 66.67), description. Returns the proposal id, creation attestation id, deadline, and blackboard URI.

  • write — append key/value to the proposal's blackboard, with optional author. Writes are appended, never replaced — the full ordered history is preserved and attested.

  • read — read the blackboard: every write in order with timestamp, author, and attestation id. Optional key filter. The deliberation record before voting.

  • vote — submit a ballot from an expected voter identity, with optional reasoning (stored in the attested record, never affects the outcome). Ballot shape depends on the algorithm — see below. Duplicate votes and post-deadline votes are rejected with structured errors.

  • resolve — apply the algorithm and attest the outcome. Callable once quorum is met (or after the deadline); also fires automatically when the deadline passes with quorum. Idempotent. Condorcet cycles and missed supermajority thresholds return co_winners with status unresolved — genuine ambiguity is surfaced, never hidden.

  • history — the complete attested record: creation, every write, every vote with reasoning, the outcome. Optional filter: all | writes | votes | outcome. The compliance artifact.

Algorithms

AlgorithmBallotUse when
pluralityoption stringBinary decisions only
approvaloption or list of optionsDefault. No strategic incentive; weak/multiple preferences
bordaranking of all options3–5 options; surfaces the broadly-acceptable choice
condorcetranking or single optionRational preference aggregation; cycles returned as co_winners
opinion_pool{option: probability} summing to 1Genuine probabilistic beliefs; returns full distribution
supermajorityoption stringHigh-stakes decisions needing >50% (set threshold)

The blackboard

LLM consensus failures are often information failures, not algorithm failures — agents vote in isolation. The blackboard is shared working memory scoped to one proposal: agents write analysis, read each other's findings, then vote informed. Write-first, read-second ordering and immutable pre-deadline votes are the conformity-bias defenses.

Tracing

Every proposal is one Telinea trace. The root span (proposal) opens at propose and closes at resolve or deadline expiry — its duration is the deliberation time. Each write, vote, and the resolution is a child span emitted in real time with its Stamp attestation id. Root status: OK decided, UNRESOLVED ambiguous, ERROR expired without quorum.

Span ids are deterministic (uuid5 over proposal id + event seq), so replays and restarts stay idempotent. Push is fire-and-forget and fail-safe — telemetry can never break a vote.

Auth is per-account: spans are pushed with the Telinea key registered via register, so ingest attributes them to the correct account. The key is stored encrypted (AES-256-GCM under HKDF of your Quorum key — never stored); a proposal-scoped re-encryption under the server-held QUORUM_DATA_KEY lets spans authenticate after restarts without re-presenting the Quorum key. A database dump exposes only ciphertext.

Config: TELINEA_INGEST_URL (default https://ingest.terradev.cloud/v1/traces), TELINEA_WORKSPACE_ID, TELINEA_PROJECT_ID, TELINEA_DISABLED=1, QUORUM_DATA_KEY, QUORUM_ACCOUNTS (account store path, default ~/.quorum/accounts.json).

HTTP transport

EndpointMethodDescription
/mcpPOSTJSON-RPC 2.0 — single or batch; notifications → 202
/mcpGETSSE keep-alive channel
/GET/POSTService identity / MCP alias
/healthGET{"status":"ok"}
/v1/infoGETSelf-describing service info (tools, auth, spans, docs)
/.well-known/agent.jsonGETAgent card
/.well-known/oauth-protected-resourceGETRFC 9728 — public, no auth

Concurrency capped at 100 simultaneous POST /mcp via asyncio.Semaphore.

Rate limiting is per-proposal: write and vote are capped at QUORUM_PROPOSAL_RATE events/min per proposal id (default 120), so one hot proposal can't starve the rest. The edge per-IP zone in Caddy is a coarse backstop only.

Attestation is queued and coalesced: concurrent events share one NTP sample per drain batch instead of one query per event — each event still gets its own Stamp record (unique id, own payload hash), so throughput isn't bound by NTP round-trips.

State

Event-sourced: one append-only JSONL file per proposal under ~/.quorum/proposals (override QUORUM_DATA_DIR). Namespaced proposals live in per-namespace subdirectories (proposals/<ns>/<name>.jsonl). The log is the history — history is a filtered read of the same file the tools append to. Nothing is rewritten or deleted.

Self-hosting

git clone https://github.com/theoddden/quorum-mcp.git
cd quorum-mcp
docker compose up -d --build

Caddy handles TLS once DNS points at the host. See deploy/Caddyfile.

License

Copyright 2026 theoddden. Licensed under the Apache License, Version 2.0.

Comments

Markdown supported

New to MCP?

From what the protocol is to installing your first server, these answers cover most first questions.

An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.