htalk

Exchange messages between local agents, including Codex, Claude Code and OpenCode sessions. htalk saves requests and replies in a shared SQLite inbox and can notify the recipient through its client. Either participant can ask, answer now or return later.

For Linux and mutually trusted sessions. Mailbox processes share one trusted OS account on the mailbox host; a remote client can connect through restricted SSH. Peer names identify routes, not authenticated users. The package name is harness-talk; the command is htalk.

The roadmap tracks the planned releases and their completion criteria.

Beyond the three native clients:

• Session receivers for Pi, Oh My Pi, Hermes, OpenClaw, Agent Zero, Cline, OpenHands, Copilot CLI and Gemini CLI read a shared htalk watch stream. Kilo uses the OpenCode adapter. The installed wheel supplies htalk; the receivers are files of the matching source archive or checkout.
• Managed sessions for Goose, Letta Code and the Antigravity SDK share owner-task binding, same-session restart and explicit recovery. They do not attach to existing TUI or IDE windows.
• The MCP mailbox tool gives a client with MCP support the same message commands over stdio. Tool access and automatic session notification are documented separately for each harness.
• Two devices can share one mailbox through the MCP-over-SSH route, with a Codex receiver for a worker on another device and a Linux user service for a reverse SSH route.

Each release has its upgrade note under docs/releases. 0.13.0 changes error codes, the keys of a successful result and the text of the help, once; the mailbox stays on schema 3 and needs no migration. Stop mailbox writers and receivers before updating the CLI and their matching adapter files.

Install and share a database

With uv:

uv tool install harness-talk
export HTALK_DB=/absolute/shared/directory/mail.sqlite3

Or install with python -m pip install harness-talk (Python 3.11+). htalk is a Rust executable, distributed as a Python package. Linux wheels for x86-64 and ARM64 with glibc 2.28+ include the compiled executable and SQLite; installing a matching wheel needs no Rust compiler. The installed htalk runs without Python. Source installs require Rust 1.88+, a C compiler and a linker; see development.

Set the same HTALK_DB in both sessions. Both must be able to run htalk and write the database directory. Only peer add creates the file; other commands report database_not_found for a wrong path. --db PATH overrides the environment; storage defaults are documented separately.

A mailbox on schema 1 or 2 is backed up and upgraded to schema 3 by the first command that opens it; see migration and recovery. Update every htalk installation that uses the mailbox.

Errors carry fixed codes, listed in the reference. Parse JSON fields and codes, not prose: a file or database failure of the mailbox itself is answered in the system's words.

Find and register the participants

htalk peer discover

Use the exact session IDs and workspaces in its JSON result. Discovery only finds addresses; it does not register peers, open the inbox or start clients. Check sources before treating an empty result as absence. See client setup for discovery coverage and connection details.

For example, register a Codex builder and a Claude reviewer:

htalk peer add builder --harness codex --session CODEX_UUID \
  --workspace /absolute/builder
htalk peer add reviewer --harness claude --session CLAUDE_UUID \
  --workspace /absolute/reviewer
htalk peer list

Replace those IDs and paths with the discovered addresses. For an app-server address, keep its --socket PATH. For OpenCode, follow the server setup.

When a registered session is no longer used, htalk peer retire NAME hides it from peer list and refuses new requests to or from it. Its saved questions can still be answered.

Any agent that can run the command can instead use a pull peer:

htalk peer add helper --harness generic --delivery pull
htalk --as helper inbox

--harness is a label such as generic, hermes or openclaw; a label does not install an integration. Pull peers need no native session or workspace and reject address flags. Messages to them are saved with notification_detail: pull_only and exit 0, without a notification attempt. The agent must run inbox to get its work. It can send and reply to native peers normally. peer check reports the delivery mode and does not establish that a pull agent is running.

To find owner-published profiles through a pinned SSH IPv4 address, LAN/Wi-Fi, an active Bluetooth PAN or private Tailscale, use the opt-in profile catalogue. Known devices are authenticated through pinned SSH; choose a profile by name and use its checked MCP connection. Reachable profiles keep runtime_status: unknown until separate runtime evidence exists.

Ask, answer and recover

From the builder's session, check the recipient in the same execution scope that will send:

htalk peer check reviewer
htalk --as builder send reviewer --message 'Which case needs another test?' --wait 45

From the reviewer's session, read the request, acknowledge it and answer. Use the request ID returned by inbox. A registered Claude Code session such as this reviewer may also omit --as; see peer selection.

htalk --as reviewer inbox
htalk --as reviewer ack REQUEST_UUID
htalk --as reviewer reply REQUEST_UUID --message 'Test recovery after interruption.'

The builder can retrieve a delayed answer and then acknowledge its separate ID:

htalk --as builder show REQUEST_UUID
htalk --as builder ack REPLY_UUID

An acknowledgment records the recipient's declaration of reading and removes that message's pending Codex notice when possible. A question stays open until answered. When the builder's wait records an answer before the final notification check, htalk skips that notice. An already accepted notice may still arrive. Sending a notification does not prove that the recipient read it. The result includes submission. One whose notice failed or is uncertain adds next_action and copyable recovery commands.

After interrupted or uncertain delivery, use show, wait, inbox or sent. sent lists your newest outgoing messages first, 20 at a time with summarized texts; next_page is the command that continues the list. Never send the same question again under a new ID to retry a notification. Use --no-notify when the recipient will poll its inbox.

Reuse a request ID

For automation, generate and save a UUID before the first send, for example with uuidgen. Replace REQUEST_UUID below with that saved value. Keep the same database, sender, recipient and message text when retrying after an interruption:

htalk --as builder send reviewer --id REQUEST_UUID \
  --message 'Which case needs another test?'

# An identical retry uses the same saved UUID.
htalk --as builder send reviewer --id REQUEST_UUID \
  --message 'Which case needs another test?'

If the request is already saved, the retry returns it with created: false and makes no new notification attempt. Its exit code can be 0 even when the original notification is submission_unknown. Check the saved result:

htalk --as builder show REQUEST_UUID

Reusing that ID with different text returns message_id_conflict, exits with code 2 and preserves the original request:

htalk --as builder send reviewer --id REQUEST_UUID \
  --message 'A different question?'

Help and details

htalk --help
htalk peer add --help
htalk send --help

• Reference: message states, recovery, database selection and exit codes.
• Codex and Claude adapters, including tested versions and discovery limits.
• OpenCode setup, including how a notification can start a turn in an existing session.

htalk does not launch interactive clients or create sessions. Incoming peer messages do not grant permission to act.

Report bugs and suggestions through GitHub issues, with versions, command, expected result and actual result. Remove private conversation text and credentials. See contributing, releases and the MIT license.