AI Coding Agent Communication Tools: Choose by Workflow
Compare local peer messaging, agent mailboxes, terminal automation, native client messaging, and delegation before choosing an agent communication tool.
A comparison of primary documentation, without a head-to-head client test.
Questions in this comparison
The right communication tool depends on which sessions you want to connect and how much responsibility you want the connecting layer to take.
A mailbox, a terminal workflow, an in-client delegation plugin, and a native messaging feature can all move text. They need not preserve the same session ownership, store the same context, or reach the same receivers.
Begin with a concrete question#
For independent local sessions, try this question:
Can my interface session ask the service session I already opened for its API contract, then use the answer without me copying it between windows?
For another workflow, the question may instead be whether one agent can launch new workers, whether a human can search a shared archive, or whether sessions on different machines can communicate.
Choose the requirement before comparing features.
A practical shortlist#
| Tool or approach | Documented emphasis | Evaluate it when |
|---|---|---|
| Agents Can Communicate | Local peer awareness, durable messages, receipts, and handoffs across independent client sessions. | You want to keep separately opened sessions and their existing permissions. |
| hcom | Cross-terminal messaging with activity observation, events, agent control, and cross-device relay. | You also want terminal automation and visibility into agent activity. |
| MCP Agent Mail, Rust implementation | Persistent mail-style coordination with search, an archive, reservations, and operator interfaces. | A shared mailbox and inspection workflow are central to your use. |
| Concord MCP | Local messaging plus explicit task context, ownership transitions, handoffs, and review state. | The shared record should track the work's lifecycle as well as its messages. |
| Official Codex plugin for Claude Code | Codex review, delegation, transfer, and background jobs initiated inside Claude Code. | Your workflow starts in Claude Code and invokes Codex from there. |
| Claude Code native messaging | Messaging among a user's Claude Code sessions, including documented remote scenarios. | Your participating sessions all use Claude Code. |
| Agent Relay | Messaging infrastructure with channels, DMs, threads, and hosted or self-hosted engine options. | Deployment scope and communication infrastructure are part of the requirement. |
Start with the row that matches the work you want your sessions to share. Then try that workflow in your own clients.
Compare the same receiving conditions#
Test an idle receiver, a receiver in a running turn, and a receiver that is temporarily unavailable.
For each condition, establish what happened: recorded, accepted by a transport, retrieved, or answered. Products can use different terms, so compare the underlying observation rather than matching labels.
ACC keeps queued, offered, retrieved, and acknowledged distinct. Its live paths are experimental, require recipient opt-in, and depend on the particular client surface. See compatibility.
ACC reports CLI and desktop delivery capabilities separately. Check the surface you actually use.
Inspect the data boundary#
Ask which information enters shared storage. A short intentional message, a task packet, a transcript, and a terminal observation are different kinds of data.
ACC excludes raw transcripts and keeps its runtime state outside repositories. Models can still receive the message through their coding providers; local coordination does not make the models run locally.
Decide what your workflow needs to retain before choosing an observation or archive feature.
Keep ordinary setup in the comparison#
Count the client activation and trust steps, not just installation of the connecting package. Check whether your actual desktop sandbox permits access to local state and whether client upgrades affect the integration.
A documented capability is evidence about a product contract. A successful exchange from your installed setup establishes that the current path works.
Watch a recorded Codex–Antigravity exchange for an ACC example with client versions and delivery events. Then connect two sessions and obtain one useful answer in your own setup.