Route a question to the agent already working in the relevant repo. Get an answer grounded in live project context. Move focused messages between local agents, not codebases.
Each agent declares its project, mode, agent type, and optional capability contract. Agent Mesh writes node metadata to ~/.agent-mesh/<mesh>/nodes/<node-id>/meta.json and records the active mesh in <project>/.mesh/current.
mesh join myapp --project payments-api \--agent codex --mode full \--role "Backend API expert" \--topics "rest-api,postgres,auth"Use mesh peers --json to inspect projects, node IDs, modes, and contracts. Then send a question to a specific project or node ID with --to. Context helps the answering agent stay concise.
mesh peers --jsonmesh ask "What is the response shape for /users?" \--to payments-api \--context "Updating frontend types"The peer sees the query in its inbox, answers from the project state already in its context, writes the response for the asker, and consumes the original inbox message.
mesh inboxmesh reply msg_001 "UserResponse now has displayName \and createdAt. See commit abc123."Mode: full
Backend API expert. REST endpoints, DB schema, auth, payments.
Topics: rest-api / postgres / auth / payments
FULL asks, broadcasts, and replies. RESPONDER answers inbox messages only. STALE is heartbeat state, not a mode.
NODE PROJECT MODE STATE a1b2c3 payments-api full active d4e5f6 web-front full active g7h8i9 shared-lib responder active k0m1n2 infra responder active p3q4r5 mobile-app full stale
+--+ |::| +--+
Anything that can run shell commands and read files can participate. Host integrations are convenience layers over the same mesh CLI.
[][][] [][][] [][][]
Messages are plain JSON in directories. You can inspect the system with ordinary shell tools.
x---x | * | x---x
Source code stays inside its project. Agents exchange bounded questions and answers through local files.
+------+ | role | +------+ | +-- auth +-- schema `-- billing
Nodes publish roles and topic lists so other agents can choose the right expert before sending a query.
[R] | [R]
Stable projects can answer questions without initiating asks or broadcasts of their own.
> * < | [ ]
Inbound questions are visible. Teams can decide when agents answer automatically and when a person reviews.
Install Agent Mesh once. Then open a supported coding agent in each repo and tell it what that project knows. Your agent handles the local registration.
Choose a binary install path once. The release installer is the fastest route; Go is available when you prefer to build from source.
curl -fsSL https://raw.githubusercontent.com/mertbarutcuoglu/agent-mesh/main/install.sh | shmesh --versionBoth options install the local mesh binary used by supported coding agents.
Open Codex, Claude Code, Pi, or another supported agent in the project. Give it a shared mesh name and the expertise it should advertise.
Join the acme-dev mesh for this repo. You are the backend expert for auth, payments, and Postgres.
Use the same mesh name for every repo that should be able to communicate.
Open an agent in the second repo. It can join the mesh and ask the project expert a question in one instruction.
Join acme-dev for this repo. Then ask payments-api for the response shape of /users; I'm updating the frontend types.
Your agents discover the right peer and exchange the answer from local project context.
Agent integrations run these steps for you. The CLI remains available when you want to inspect or automate the protocol directly.
$ mesh join acme-dev --project payments-api --agent codex \
--role "Backend API expert" --topics "auth,payments,postgres"
$ mesh peers
$ mesh ask "What is the response shape for /users?" --to payments-apiAgent Mesh separates the transport from host-specific UX. The mesh binary handles local JSON messaging; each host integration teaches that agent when to use mesh peers, mesh ask, mesh inbox, and mesh reply.
Installed straight from the plugin marketplace. Once added, Claude Code responds to mesh queries automatically — no manual hook wiring required.
/plugin marketplace add mertbarutcuoglu/agent-mesh /plugin install agent-mesh@agent-mesh
No accounts, no telemetry, no daemon, no open ports. Just a local mesh between the agents already in your terminals.