Agent Wire
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Agent Wiresend reviewer a message: interface is ready"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Agent Wire
Let Codex and Claude Code talk to each other.
Agent Wire gives coding agents a common address book, inbox, and reply API. It runs locally, stores messages in SQLite, and delivers them through each runtime's own session interface. Both agents use the same MCP tools.
flowchart LR
Codex <-->|MCP tools| Broker[Agent Wire · local broker]
Claude[Claude Code] <-->|MCP tools| Broker
Broker -->|App-server tool output| Codex
Broker -->|Session inbox| ClaudeAlpha software for Linux and macOS with Python 3.11+. Native adapters currently target Codex app-server 0.159.x and Claude Code 2.1.280. Other versions fail explicitly. Windows, remote hosts, Claude Channels, and automatic runtime launching are outside this first release.
Quick start
Install from this repository (there is no PyPI release yet):
python3 -m venv .venv
.venv/bin/python -m pip install 'git+https://github.com/kevinmanase/agent-wire.git'
.venv/bin/agent-wire serveKeep the broker running. In another terminal, discover native sessions:
agent-wire discoverUse the session IDs and socket paths returned by discovery. Enroll each session you want to participate, with a distinct name:
agent-wire register --name coder --runtime codex \
--session CODEX_THREAD_ID --socket /absolute/path/to/app-server-control.sock
agent-wire register --name reviewer --runtime claude \
--session CLAUDE_SESSION_ID --socket /absolute/path/to/claude-inbox.sockEach command returns an identity_file. It contains a private session
credential. Use the appropriate file for the sender:
agent-wire send --identity /path/to/coder-identity.json \
--to reviewer --body 'The interface is ready. Please report any mismatches.'
agent-wire status --identity /path/to/coder-identity.json MESSAGE_IDThe CLI prints JSON. queued and submitted do not mean the recipient read
the message. acknowledged means it explicitly acknowledged receipt. A reply
records reply_id; it does not prove the requested work is complete.
Use an absolute installed executable path if agent-wire is not on your PATH.
All commands accept --state /private/directory before the subcommand.
Related MCP server: set-agent-comm
MCP tools in both agents
Configure a stdio MCP server that runs:
agent-wire mcp --identity /absolute/path/to/this-session-identity.jsonTool | Purpose |
| Find explicitly enrolled recipients |
| Send text; set |
| Read pending messages, optionally after a cursor |
| Acknowledge a received message |
| Inspect a message you sent or received |
A bound MCP server belongs to one native conversation. Do not share that
configuration between unrelated Codex threads: an app-scoped MCP process may
serve several conversations. For shared configurations, run agent-wire mcp
without --identity and enroll through the optional SessionStart hook; each
tool call supplies that conversation's session_handle.
See setup examples for both clients, hook registration, permission behavior, and a model-free two-mailbox walkthrough.
Delivery behavior
Messages persist across broker restarts. Recipients explicitly enroll; discovery alone does not enroll them.
Codex messages arrive as tool output. Claude messages use its peer inbox and keep its inbound controls.
Native adapters never resume old conversations or launch new agents. Re-enrollment retires the previous credential and address.
A transport failure after writing begins becomes
unknown. It is not automatically retried.Reuse an idempotency key for a retry of the same send. Expiry, bounded inboxes, rate limits, and reply-hop limits constrain loops.
Agent Wire does not execute message bodies, forward permission approvals, or clear conversations.
The local OS user is the trust boundary. This is not a sandbox between mutually hostile agents running as the same user. State and the broker socket are private to that user; session credentials prevent accidental sender confusion and scope message access. See security and protocol and adapter notes.
Development and contributions
git clone https://github.com/kevinmanase/agent-wire.git
cd agent-wire
python3 -m venv .venv
.venv/bin/python -m pip install -e '.[dev]'
.venv/bin/python -m pytest
.venv/bin/ruff check .
.venv/bin/ruff format --check .Tests use isolated state and fake runtime sockets; no accounts, API keys, or model calls are required. Live runtime checks are separate, explicit exercises. See initial verification for the tested native paths and a successful live Codex → Claude → Codex exchange. Contributions to adapters, cross-platform support, and delivery semantics are welcome. Start with CONTRIBUTING.md.
License
Copyright © 2026 Kevin Manase and Agent Wire contributors.
GNU AGPL version 3 only. You may use Agent Wire commercially. Covered redistribution and modified network-served versions carry reciprocal source-sharing obligations under the license. The license does not require a pull request to this repository; upstream contributions are warmly encouraged.
This server cannot be deployed
Maintenance
Related MCP Connectors
Email inboxes for AI agents: send, receive, reply, search, and manage threaded email over MCP.
Hosted MCP messaging across owners, tools, and machines, with readable transcripts.
Email infrastructure for AI agents — send, receive, search, and reply to email over MCP.
Messaging and inboxes for AI agents: register, send signed messages, check your inbox, find agents.
Related MCP Servers
- AlicenseBqualityAmaintenanceEnables durable message passing between Codex project agents with MCP tools and automatic wake-up delivery, storing messages as JSON on disk.23MIT
- AlicenseNot gradedqualityCmaintenanceEnables local agent-to-agent messaging between Claude Code sessions via file-based channels, with a registry, MCP tools and CLI for sending, reading, and tracking messages.2MIT
- AlicenseNot gradedqualityAmaintenanceEnables AI agents to message each other by @nickname via an MCP server, with contacts, presence, and durable delivery across local and remote agents.3Apache 2.0
- AlicenseNot gradedqualityAmaintenanceEnables agents to communicate asynchronously via private inboxes and to ask a human a question and block until answered. Provides MCP tools for sending messages, checking unread messages, and listing peers.Apache 2.0