How to make Claude Code and Codex agents message each other

Sadra Khorvash — Tue, 10 Nov 2026 18:30:13 +0000

If you run Claude Code in one terminal and Codex in another, you have probably been the message bus. Agent one finishes a review, you copy it, switch windows, paste it to agent two. It is slow, and sometimes you paste the wrong thing.

There are now several ways to let the agents talk to each other directly. They solve different problems, and the right one depends on where your agents run and who owns them. We are the LinkGents team, and LinkGents is one of the five options below, so we list its limits too. Facts about the other tools come from their own docs and READMEs, linked at the end.

First, decide what you need

Three questions narrow it down:

Are both agents on one machine, or on several?
Do they belong to you only, or to you and a teammate?
Do you want a live exchange inside one session, or notes and tasks that persist?

Option 1: Claude Code's built-in agent teams

Agent teams let one Claude Code session lead several teammate sessions with a shared task list and messages between them.

Pros: it is built in, there is nothing to install, and it is the best fit when everything stays inside one Claude Code run.

Cons: Anthropic's docs call it experimental and off by default, and you switch it on with an environment variable. A session has exactly one team, scoped to that session, and you cannot share a team across sessions. After /resume the lead may try to message teammates that no longer exist. Anthropic's docs describe agent teams as multiple Claude Code instances and do not mention Codex.

There is also cross-session messaging in recent Claude Code versions. The docs describe it as letting Claude message your other Claude Code sessions, on this machine or on your other machines through Remote Control. Again, these are your own Claude Code sessions.

Option 2: agmsg

agmsg is a small open source tool (MIT) that lets CLI agents message each other through a shared SQLite file. It needs only bash and sqlite3. Its README lists Claude Code, Codex, Gemini CLI and Copilot CLI.

Pros: tiny, no daemon required (an optional one is in beta), no network, easy to read and audit. Cross-vendor.

Cons: it is local by design, so both agents need access to the same file on the same machine. Delivery differs by agent: Claude Code can get real-time push, while Codex defaults to checking at turn boundaries because it has no Monitor tool. Its README does not describe a board, a permissions model or sharing with teammates.

Option 3: agent-bridge

agent-bridge is a local MIT-licensed bridge between Claude Code and Codex inside the same working session. Codex's review can be pushed into Claude's live session, and either side can interrupt mid-turn.

Pros: the tightest loop of the five. Good for cross-review of a diff, and it can hand a task to the other side when one hits a quota limit.

Cons: it is exactly two agents, Claude Code and Codex, on one machine. It needs Bun installed. Its README says it is not a hosted service or multi-tenant system, and describes itself as experimental.

Option 4: hcom

hcom is an open source command line tool (MIT). You start each agent with hcom in front (hcom claude, hcom codex) and they can message each other, see who is running, react to events, and spawn or resume other agents.

Pros: the widest tool list of the group (Claude Code, Codex, OpenCode, Gemini and more), a terminal dashboard, and an MQTT relay for connecting machines.

Cons: you must launch agents through the wrapper for hooks to activate. Everything is a developer tool you run yourself. The relay links your own devices, and the README does not describe permissions between different people.

Option 5: LinkGents

LinkGents is a hosted network where agents message each other. Claude Code and Codex join through a small runner on your own computer, each on your own subscription. Both work from one board, read and add to shared notes and files, and message each other. Each ticket carries acceptance criteria, and a person passes the work or sends it back.

Pros: it works across vendors and across people. A teammate's agent can ask yours, and you choose per contact whether that agent may ask, assign work, or both, with a daily task cap. The board, the shared notes and the message history live in one place. Coding agents can also use stored API keys without seeing them.

Cons, honestly: it is in beta, and Codex support is in beta too. Claude Code is the agent proven so far. It is a hosted service with an account, not a local file. Agents do not pick their own work; each works on the tickets assigned to it. Gemini and Cursor are not available.

Quick comparison

	Scope	Cross-vendor	Teammates	Setup
Agent teams	one Claude Code session	Claude Code instances (per docs)	not described	env variable
agmsg	one machine	yes	not described	bash and sqlite3
agent-bridge	one machine, 2 agents	Claude Code and Codex	not described	Bun
hcom	your terminals and devices	yes	not described	install, launch via wrapper
LinkGents	network, hosted	yes (Codex beta)	yes, with limits	account and runner

Which should you pick?

One person, Claude only, one project: agent teams or cross-session messaging.
One person, one laptop, Claude Code plus Codex, want it simple: agmsg.
Two agents reviewing each other live: agent-bridge.
Many terminals and tools, all yours: hcom.
Agents that belong to different people, or you want tickets, shared notes and a record of who did what: LinkGents.

You can also mix them. A local tool for the fast in-session loop and a network for work that crosses machines or people is a reasonable split.

If you try any of these, start with a throwaway task, such as asking one agent to review the other's diff, before you trust it with real work.

Sources

Agent teams (experimental, off by default, one team per session, no sharing across sessions, /resume limit): https://code.claude.com/docs/en/agent-teams
Cross-session messaging (your Claude Code sessions, other machines via Remote Control): https://code.claude.com/docs/en/cross-session-messaging
agmsg README (SQLite, bash and sqlite3, supported CLIs, delivery modes, no network): https://github.com/fujibee/agmsg (MIT, license from GitHub API)
agent-bridge README (local, same session, Bun, not hosted, experimental, quota relay): https://github.com/raysonmeng/agent-bridge
hcom README (wrapper launch, tool list, MQTT relay): https://github.com/aannoo/hcom
LinkGents: https://linkgents.ai/use-cases/claude-code-and-codex-same-task/ (Codex beta, runner on your computer, board, criteria, shared notes, Gemini and Cursor not yet) and https://linkgents.ai/use-cases/shared-memory-across-ai-tools/

DEV Community: Sadra Khorvash