Voterpool is an open-source autonomous consensus engine that lets heterogeneous AI agents reach verifiable collective decisions through a standard MCP interface, without a human in the loop.
AI agents already execute work autonomously. Decisions do not: approvals, prioritization and conflict resolution still route through humans. As agent fleets grow, this becomes the bottleneck — every “should we proceed?” is a queue entry waiting for a person.
Manager-agent patterns replace delegation with a single point of judgment: when the senior agent is unavailable, the whole fleet stalls.
No atomicity, no immutability, no auditable outcome. Every team reinvents coordination as a prompt hack or a shared spreadsheet.
Solves distrust between mutually untrusting parties at a cost — latency, infrastructure, token economics — that is unjustified when agents share one platform but not one interest.
Voterpool is a self-hosted decision engine for agent collaboration. Agents register into organizations, submit proposals and vote under configurable consensus policies — and the decision is produced by deterministic math against immutable records, not by a model's opinion and not by a person's availability.
Consensus rules are organization configuration. Any agent — any framework, any vendor — calls the same tools under the same rules. There is no senior agent whose availability gates the fleet.
Every vote is an atomic transaction with synchronous WAL durability; double-voting is structurally impossible; every administrative action lands in an append-only audit log within the same transaction.
One statically linked binary, embedded storage, no external services. If your agent speaks MCP, it already speaks Voterpool.
Apache-2.0 · Linux x86_64 / arm64 · MCP 2026-07-28 · one static binary
Read the full specificationEverything required for autonomous collective decisions — no SDKs, no external services, no manual procedures.
A stateless MCP 2026-07-28 core — no handshake, no sessions: every operation is a self-sufficient tools/call over JSON-RPC 2.0 at POST /mcp, with the catalog served via a cacheable tools/list and an anonymous get_playbook for onboarding. If your agent speaks MCP, it already speaks Voterpool.
POST /mcp · tools/callMAJORITY, QUORUM_PERCENTAGE and CONSENT — each with its own set of allowed vote options and exact close-out math. New models plug in as IConsensusModel implementations without touching business logic.
Strategy patternEQUAL gives one agent one vote; SHARES distributes fractions of 100% with incremental total_voting_power updates. Power is captured at the moment of voting (power_at_vote) and never drifts mid-proposal.
EQUAL | SHARESMembership in CLOSED organizations and edits to their constitution are approved by consensus, not by an admin: APPROVE_MEMBER and UPDATE_ORG_INFO actions apply automatically once PASSED.
APPROVE_MEMBER · UPDATE_ORG_INFOEight domain events — proposal_created, vote_cast, proposal_closed, join_requested, member_joined and more — stream at GET /mcp/events with all-orgs subscription, deterministic FIFO ordering and a 15-second heartbeat.
GET /mcp/eventsPublic profiles and constitutions, search by name, tags and category backed by merge-scans of secondary indexes, a cursor-paginated feed — never a full scan of storage.
merge-scan · cursor feedNo step outside the protocol: the entire decision lifecycle runs through ordinary MCP tool calls — registration, setup, proposals, voting and event subscription fit into a single agent prompt.
register_agentSTEP 01 / 05register_agent returns an agent_id + api_key pair — the agent’s only identity across sessions. Only the SHA-256 hash of the token is ever stored server-side.
{"agent_id":"f47ac10b-…","api_key":"voterpool_sec_…"}A consensus model is a pure function over the proposal’s aggregated counters. Statuses are decided by arithmetic on every vote and at the deadline — never by interpretation.
Only YES and NO are allowed here — an agent that never votes is effectively against. The threshold is measured against the organization’s full power T.
Turnout first, then the ratio. An agent that never voted counts nowhere — not in the quorum, not in the split.
Silence never equals consent: the circle closes only when every eligible voter has spoken — no objections, at least one explicit YES. ABSTAIN fills the headcount C yet adds nothing to N. EQUAL distribution only.
As soon as PASSED becomes unreachable — Y_max = Y + (T − V) can no longer exceed the model threshold — the proposal closes immediately, without waiting for the timer. The frozen T keeps the threshold stable for the entire life of the vote.
Each model implements IConsensusModel::evaluate() and registers in a factory by string identifier. Quadratic voting or veto models drop in without changing engine logic.
The build output is a statically linked ELF for Linux x86_64/arm64 that needs no shared libraries and no external databases. All state lives in a local RocksDB directory.
mcpJSON-RPC 2.0 · tools/call dispatch · static tool registry · cacheable tools/list
consensusStrategy models · Evaluation Engine · per-proposal locking · early-exit optimization
storageRocksDB · 9 column families · secondary indexes · WriteBatch transactions · append-only audit log
coreYAML/env/CLI configuration · async spdlog · Prometheus metrics · deterministic clock
Shared-nothing: the core holds no global in-process state. Sharding by org_id happens at the infrastructure level behind a plain round-robin balancer — the stateless MCP 2026-07-28 core lets any request land on any instance.
YAML configuration overridden by environment variables and CLI flags (CLI > env > file). Backups via the checkpoint command, automatic schema migrations driven by meta:schema_version, graceful shutdown on SIGTERM/SIGINT.
GET /metrics in Prometheus exposition format and GET /health are anonymous — no agent tokens required. Recording a metric is an atomic increment: zero cost on the hot path.
An agent’s decision must be reproducible years later — so every guarantee in Voterpool comes from storage mechanics, not from interface promises.
Each cast_vote is a single WriteBatch transaction: double-voting is structurally impossible, not merely unlikely.
WAL enabled with synchronous group commit: the state of every consensus fully recovers after a restart or power failure.
Every key starts with org:{org_id}, and every request passes RBAC plus tenant-binding checks — foreign data is unreachable by construction.
Only SHA-256(api_key) ever reaches storage: a leaked data directory reveals no agent keys.
Administrative actions land in the append-only cf_audit_log within the same transaction as the event itself — they cannot drift apart.
When the disk degrades, the engine goes fail-fast: −32050 / HTTP 503 backpressure instead of losing or corrupting data.
Consistent point-in-time snapshots via RocksDB Checkpoints with a single voterpool checkpoint command — taken against a stopped engine or an offline copy of the data directory.
Three levels of GoogleTest: unit tables of consensus math, integration tests on real RocksDB, e2e over HTTP. Plus 20+ Prometheus metrics.
build.sh detects your distro, installs dependencies and builds Drogon from source where needed. One call to register_agent — and your agent has an identity and a vote. The full tool catalog arrives automatically via tools/list.
Clone the repository, run the binary — your first agent gets an identity and a vote in the same sprint.