Agent Mesh MCP server

Read from the package source without running it. The installed server may list more.

• ack_message

  Acknowledge a message, removing it from the agent's inbox. Writes an 'ack' receipt (per-recipient — a broadcast is acked independently by each recipient).

• ask_peer

  Open a bounded, two-agent Discussion: sends the root question and (optionally) explicitly reserves one peer attempt, then waits until the conversation deadline for a settled answer. Message arrival never starts an agent by itself — wake_peer:true is the explicit, budgeted authority to run the peer once.

• attach_agent

  Attach an agent to a running fleet. It banks complete only when result_contract is ok; refused, blocked, artifact_missing, invalid, or absent banks failed.

• cast_vote

  Cast a vote on an open ratification. approve=true records approval, approve=false rejection. Re-casting CHANGES your effective vote (each change appends a new sequenced receipt — history is never rewritten); repeating your current vote is a no-op.

• collect_results

  Get fleet outputs and loss tally. Only result_contract ok can bank complete; refused, blocked, artifact_missing, invalid, or absent banks failed. Declared outcome, not quality. Check lost first.

• compile_route_candidates

  Offline projection of caller-supplied route-candidate snapshots. Does not persist, rank, execute, authorize, wake, or contact providers.

• fleet_status

  Check fleet and agent status.

• get_build_identity

  Full BuildIdentityReport for the running install: package name/version, source_commit, entrypoint_count, per-entrypoint SHA-256 map, entrypoints_match_runtime bit. Diagnostic access to the install's identity.

• get_discussion

  Read-only: derive and return a Discussion's full state — transcript, attempts, budget, and fail-closed status (invalid > closed > deadman > expired > exhausted > active > open) — from durable messages and receipts. Remains useful after every inbox entry has been acknowledged.

• get_health

  Fleet health + liveness. Pass verbosity="summary" for a smaller routine probe (entrypoints map omitted); default "full" is the full BuildIdentityReport. Use `get_build_identity` for diagnostics.

• get_inbox

  Get messages in an agent's inbox, optionally since a timestamp.

• get_receipts

  Get the full receipt trail for a message: who acked, who annotated, when. Answers 'who saw this and who acted on it'.

• get_work_receipt

  Read back a work receipt by composite key. Returns the exact stored bytes plus

• list_agents

  List all available premade agents from .opencode/agents/ directories.

• list_fleet_templates

  List all saved fleet templates, sorted by name.

• list_fleets

  List all fleets with summaries (agent count, status, completion).

• open_ratification

  Open a quorum vote ('council') over the fleet. Broadcasts a proposal; peers vote with cast_vote. Ratifies when approvals reach the quorum and every required signoff approves. Returns the proposal message_id.

• ping

  Minimal liveness check. Returns { status: 'ok', timestamp }.

• plan_speculative_backlog

  Pure, caller-approved speculative backlog projection. Does not persist, execute, authorize, wake agents, contact providers, poll, allocate capacity, schedule, spend, send, or publish.

• receiptwrite action

  Write a non-consuming receipt on a message — the audit primitive. Use actions like 'seen', 'r-ack' (approve), 'retracted'. Unlike ack_message, the message stays in the inbox. One receipt per (message, agent, action); repeat calls are idempotent.

• recommend_route

  Advisory-only ranking over caller-supplied sanitized task traits and candidate snapshots. Does not persist, execute, authorize, wake agents, or contact providers.

• record_routing_outcome

  Record whether a routed task succeeded or failed. Future route_work calls for the same agent weight their score by accumulated outcomes (Wilson-style). NOTE: outcomes are currently accumulated PER AGENT, not per capability — capability_key is recorded for forward compatibility but does not yet scope the penalty, so a failure at one capability lowers the agent's score for all of them. Feedback is i

• record_work_receipt

  Record a first-class external work receipt (Hermes Kanban result contract).

• register_capability

  Register an agent's capabilities (role, skills, model) for routing.

• reply_discussionwrite action

  Submit the one reply an active wake attempt is authorized to produce. Never launches an agent. Rejects a second reply for the same attempt, a stale head, or a reply past the conversation/attempt deadline. close:true makes the conversation terminal; exhausting the last turn without close returns exhausted.

• route_work

  Route a work description to the best-matching registered agents by keyword + role/skill overlap scoring (with synonym expansion), weighted by routing feedback (success/fail history). top_n controls how many matches to return (default 1, max = fleet size).

• save_fleet_template

  Save a named fleet template (set of agent specs) for reuse. Names: lowercase letters, numbers, dashes, underscores.

• send_messagewrite action

  Send a P2P message from one agent to another within the same fleet. Set to_agent_id to "*" to broadcast to every other agent in the fleet (each recipient acks independently; see get_receipts).

• send_messageswrite action

  Send a batch of P2P messages in ONE ledger transaction — use instead of repeated send_message calls for bulk fan-out (much faster: the recipient inbox is updated once per batch, not once per message). Atomic: one invalid message rejects the whole batch. Max 1000 messages per call.

• set_fleet_timeout

  Set a per-fleet timeout override (in milliseconds). Agents exceeding this are auto-failed.

• spawn_fleet

  Spawn parallel agents. Returns fleet_id. An agent banks complete only when result_contract is ok; refused, blocked, artifact_missing, invalid, or absent banks failed.

• spawn_from_template

  Return a fleet spec from a saved template, ready to pass to spawn_fleet.

• subscribe_events

  Subscribe to the unified fleet-wide event stream via Server-Sent Events (SSE). Returns a stream URL that emits every ledger event (messages, receipts, ratifications, spawns, completions, discussions) as it is appended. Keep-alive :hb comment frames are sent every 30s. Optional fleet_id filter narrows to events carrying that fleet_id. If the operator set MESHFLEET_AUTH_TOKEN, requests to the stream

• subscribe_inbox

  Subscribe to an agent's inbox via Server-Sent Events (SSE). Returns a stream URL that the agent opens to receive real-time push of incoming P2P messages. Falls back to polling get_inbox if SSE is unreachable. If the operator set MESHFLEET_AUTH_TOKEN, requests to the stream must carry it (Authorization: Bearer, or ?token=).

• sweep_ratifications

  Evaluate every open ratification now and persist any that reached a terminal state (deadline expiry, silent-approval, unreachable quorum). The server also sweeps automatically every AGENT_MESH_RATIFY_SWEEP_MS (default 60s).

• tally_ratification

  Read the live vote tally and current status (open / ratified / rejected / expired) of a ratification, and persist the status if it has reached a terminal state.

• verify_ledger

  Audit the ledger's internal consistency: every receipt points at a real message and honors the idempotency key, acknowledged flags are supported by ack receipts, inboxes hold no consumed or dangling messages, and ratification tallies (quorum, signoffs, vote polarity, terminal status) recompute from the receipts. Read-only. Returns ok, error/warning counts, per-finding detail, and a `scope` object

• verify_ledger_v2

  Versioned verifier output read from a dedicated read-only file snapshot; the handler performs no ledger writes. Normal parent-server startup recovery or migration may initialize or change the configured ledger before tool dispatch. Returns the unchanged internal-consistency report inside meshfleet.verify/v2 with an unsigned-snapshot evidence scope; it does not establish authorship, snapshot integr

• verify_ledger_v3

  Opt-in verifier output read from a dedicated read-only file snapshot; the handler performs no ledger writes. Returns a detached meshfleet.verify/v3 report with one local consistency band per finding, derived only from its severity. Those labels are not provenance or confidence and do not establish authenticity, completeness, tamper evidence, authorship, snapshot integrity, content binding, externa

• wake_agent

  The sole general-purpose Discussion run trigger. Atomically reserves one bounded attempt for a resident participant who is the current recipient of the canonical head, then launches it. Never waits for the reply. Duplicate concurrent calls return the existing active attempt or turn_already_active — they never launch twice.

• Code scan82 source files scanned25/25
• –Live reliabilityno gateway calls yet and no remote to proben/a
• –Tool poisoningtools not inspected (local package is not executed); not countedn/a
• Auth qualitylocal package, no credentials required12/15
• Maintenancelast push 2 days ago15/15
• Maintainer identityregistry namespace matches repository owner; GitHub account older than a year8/10

Safety score

92/100

higher than 90% of 4,460 AI and agents servers

Tools

40

the typical AI and agents server has 6

Tools that change things

10%

typical: 0%. Fewer write tools means less an agent can break

GitHub stars

1

more than 50% of 2,711 AI and agents servers

Downloads a week

205

more than 75% of 1,538 AI and agents servers with a package

Agent Mesh: common questions

Is Agent Mesh MCP server safe?

Yes, by our scan: it is graded A (92/100). Read the Agent Mesh safety report

How do I install Agent Mesh?

It runs on your machine. Copy the Claude Code, Cursor, VS Code or Claude Desktop config from the install section.

Does Agent Mesh need an API key?

Not as far as the registry entry and our scan can tell: no credentials are declared or required.

Is Agent Mesh maintained?

The last commit was 36 days ago (2026-09-29). The latest release is v0.22.0.

What can I use instead of Agent Mesh?

Servers from other publishers that do the same job: PaperBanana MCP server, Emilia Protocol MCP server and Agoragentic Agent OS MCP server.

How does Agent Mesh compare with other AI and agents MCP servers?

safety score 92/100 (higher than 90% of 4,460 AI and agents servers); tools 40 (the typical AI and agents server has 6); tools that change things 10% (typical: 0%. Fewer write tools means less an agent can break). Full comparison.