How to Connect AI Coding Agents Across Clients
Connect independent AI coding agents with ACC. Set up Claude Code and Codex or another supported pair, exchange messages, and verify the first reply.
In this guide
Two coding sessions can work on the same feature while knowing different things. One changes an API. Another builds the screen that calls it. A useful connection lets the second session ask about the interface and bring the answer back into its own work.
Agents Can Communicate, or ACC, lets AI coding agents communicate across independently opened local sessions. You choose the clients, start the sessions, and decide their work. ACC supplies local discovery, messages, replies, and handoffs.
The useful first result is one question answered by the right peer. Start there before dividing a whole project among agents.
Choose the sessions you actually use#
ACC has adapters for Antigravity, Claude Code, Codex, Gemini CLI, Grok Build, and Kimi Code. Their delivery modes differ. A generic MCP client can participate through an inbox that the agent checks.
Pick two clients you already have installed on the same machine, under the same operating-system user. For example, use Antigravity for the interface and Kimi Code for a service change. Claude Code and Codex are another combination, rather than a required pair.
Check client compatibility before expecting an idle session to wake. An integration, a next-turn message, and experimental live delivery are different capabilities.
Install ACC and activate the integrations#
ACC requires Node.js 24 or newer. It supports macOS, Linux, and Windows 10/11.
npm install -g agents-can-communicate
acc install
Follow the activation instructions for the clients the installer finds. Restart them from the project directory. Complete the client's own hook or plugin trust review when requested.
For the first exchange, the default delivery behavior is sufficient if you are present to give the receiver its next turn or ask it to check its inbox. Experimental live delivery is a separate opt-in that can start a model turn and spend tokens.
Use Claude Code and Codex together#
For this pair, activate both the Claude Code integration and the Codex integration, then open the clients normally in the same project or in worktrees of the same repository.
Give each session its own task. For example, ask Claude Code to implement a change and ask Codex to review the identified revision. Tell the implementer to send its review request through ACC and the reviewer to return findings in the same conversation. Use each receiver's supported delivery mode; installation alone does not make an idle session wake.
This connects sessions you run independently. For Codex delegation from inside Claude Code, see the official Codex plugin comparison.
The same connection workflow applies to other supported combinations. Use the Antigravity, Gemini CLI, Kimi Code, or Grok Build inbox instructions for the relevant client.
Give each session a concrete part of one feature#
Use a small feature with an obvious shared dependency:
| Session | Task |
|---|---|
| Service session | Add account registration and define its request, response, and error cases. |
| Interface session | Build the registration form against the service contract. |
Give each session its own ordinary task. Supported hooks introduce peer awareness and install coordination instructions. Models still decide when a peer is relevant.
If you want an explicit first exchange, tell the interface session:
Ask the service session through ACC for the registration request, response, and validation errors. Use its answer when implementing the form. Keep your work within the interface.
Tell the service session:
Answer the interface session's ACC question with the current contract. Identify anything still undecided.
These are example prompts, not a transcript of a measured run.
Verify discovery before assuming a message arrived#
From the project directory, inspect the roster:
acc status --json
Look for both participants and their work. If several sessions use the same client, the agent must select an exact participant from the roster. A client name alone cannot choose safely between them.
If a peer is missing, check the launch directory, hook trust, integration activation, and current permissions. Git worktrees of one repository share an ACC workspace; unrelated directories do not automatically share one.
The setup documentation explains client-specific activation. A diagnostic report should describe your actual session rather than assume that installation made it ready.
Look for a substantive answer#
ACC records a message before attempting faster delivery. The receipt can then advance through queued, offered, retrieved, and acknowledged.
An offer means the transport accepted the message. It does not establish that the model considered the question. The result you want is an answer that names the agreed contract, the remaining uncertainty, and the artifact to check.
Each receiver keeps its own permissions and instructions. A peer request supplies context; it does not grant new authority.
Use the connection for the next dependency#
Once the first exchange works, use the same channel for independent review, coordination across worktrees, or a handoff to another client.
ACC does not merge conversation histories or choose a lead agent. Its role is to let the sessions you already chose communicate while you retain control of the work.
Sources#
Reviewed against the documentation and recorded evidence for ACC 0.11.1.