YAAC — yet another agentic chat

A radio for agentic coding sessions.

You have several sessions open at once — different worktrees, different tasks, maybe different people driving them. They need to tell each other things: "schema for rosters changed, the field is recipient_group now", "pushing the refactor in ten minutes, hold your commits". YAAC is how they do that, and how you talk into the same conversation by hand.

The mental model is a network of handheld radios, not a phone network. Buy one and it works — it just has nobody to talk to. Buy a second and there's a conversation. There is no config file, no environment variable, no port to choose, no daemon, and nothing to run first. Sessions find each other at a fixed local address — tcp://127.0.0.1:19116, and 19116 is 0x4AAC, which is where the name comes from. Whichever session needs it first claims it and relays for the others; if that session goes away, another takes over by itself, within a few seconds and without anyone doing anything.

What makes it different

It connects sessions that were never designed to meet. Agents inside one harness could always talk — an orchestrator wires its own subagents, and that was never the problem. YAAC is for two (or more) unrelated sessions in unrelated clients, alive on your machine right now: a Claude Code session and a Claude Desktop chat, Codex, Gemini CLI — anything that can run a local MCP server. One conversation is 300k tokens into a task; another, 400k tokens in, holds exactly the experience it needs. Give them a radio. If you can talk to both of them, now they can talk to each other.

Configuration rounds to zero. If you can add a local MCP server, you are done — no Redis to stand up, no PostgreSQL to prepare, no broker, no port to choose. Adding YAAC hands each client a radio, switched off. Then, at any moment, you tell a session "connect to yaac" — and it deals with the rest.

Related MCP server: agent-bus-mcp

What it's good for

• Parallel worktrees on one repo. Two sessions refactor on different branches. The one that renames a field tells the other before it builds a day's work on the old name.

• A manager conversation. You discuss what to build with Claude Desktop; it passes the task to a Claude Code session over YAAC and collects the result. Chat conversations and coding sessions are equal participants — any MCP client can join.

• Announcements. "CI is red, hold your pushes" — one broadcast reaches every session on the channel.*

• Long jobs. One session babysits a slow test suite and messages the coding session when it goes green, instead of you ferrying the news by hand.*

* On Claude Code 2.1.224 or later these arrive on their own, plugin or not; on Codex with its hooks file the session is told they are waiting. See Claude Code and Codex under Installing. Everywhere else the listening session receives them next time it checks its inbox; see Honest limitations.

Installing

One step per client, and for some clients an optional second one. The first puts the radio in; the second, where a client offers one, gets messages delivered without the session asking for them — an improvement, never a requirement. Unfold your client.

All of these need uv on your PATH. uvx fetches the package and a suitable Python by itself, so there is nothing else to install and no virtualenv to manage.

Required.

claude mcp add yaac -s user -- uvx yet-another-agentic-chat

-s user installs it for every project on the machine, which is usually what you want: a radio only one of your sessions can hear is not much of a radio. Leave it out to add YAAC to the current project only.

Check it took with claude mcp list, or /mcp inside a session.

On Claude Code 2.1.224 or later the session is also told when mail arrives. A session exports its own inbox socket to the server it starts, and YAAC rings it with a one-line [YAAC] notice — who wrote and what about — between tool calls while the session works, and as a new turn while it sits idle, which is the one case nothing else reaches. The messages themselves stay out of your transcript: with the plugin its hook hands them to the model, and without it the model reads them with check_inbox. Nothing to configure: a Claude Code that does not export the socket simply has no door, and the session reads its mail the next time it checks.

Optional: the plugin. Claude Code has its own plugin format, and this repository is the marketplace:

/plugin marketplace add amyodov/yet-another-agentic-chat
/plugin install yaac@yaac

It replaces the command above — the plugin runs the same published package with uvx, so it carries no copy of the server and picks up new releases without being reinstalled — and adds three things. A skill on when to reach for the radio. A hook that fires after a compaction and hands the session back the connections it holds, since a summary drops the identifiers the other tools need. And, on a Claude Code older than 2.1.224, delivery by hook: whatever came in is handed over as the session works, when you type, and as a turn ends, which is the one that reopens a finished turn so it can act on the news. Nothing to configure, and silent when there is nothing to say.

Required. In ~/.codex/config.toml:

[mcp_servers.yaac]
command = "uvx"
args = ["yet-another-agentic-chat"]

Recommended: the hooks file. In ~/.codex/hooks.json, or .codex/hooks.json in a project:

{
  "hooks": {
    "PreToolUse": [{"hooks": [{"type": "command", "command": "uvx --from yet-another-agentic-chat yaac-hook"}]}],
    "Stop": [{"hooks": [{"type": "command", "command": "uvx --from yet-another-agentic-chat yaac-hook"}]}]
  }
}

Approve it once with /hooks; the Codex app asks in a dialog instead. While a turn runs, the session is then told when mail is waiting, and it reads it with check_inbox(). A session under a WebSocket app-server is told when it joins if the hooks will not run there: missing, disabled, or changed since you approved them, which Codex otherwise skips without a word. On Codex 0.163.0, YAAC learns the session's thread id when it joins a channel, and the hook matches that id exactly. Several sessions in one directory are told apart without naming anything in the config, even when they share an app-server.

Recommended: waking an idle session. A hook only fires while a session is doing something. To reach one waiting at its prompt, YAAC needs an app-server it can talk to — Codex's engine, listening on a WebSocket. Run one per machine and leave it running:

codex app-server --listen ws://127.0.0.1:4500

Then start every Codex session attached to it:

codex --remote ws://127.0.0.1:4500          # or: codex resume --remote ws://127.0.0.1:4500

That is all. YAAC finds the address on the command line of the app-server that started it. When unread mail arrives, it waits for the session to be idle before adding a line to its queue. Reading the mail during a busy turn cancels that pending wake, or takes the line back out of the queue if it is already there, and several arrivals share one notice. One app-server serves any number of sessions, and each is woken by its own id.

Attached this way, a session also stays on the air while it sits idle. Codex unloads a thread that nobody is watching once it has been idle for thread_unload_delay_secs (30 minutes unless configured), and the thread's MCP servers, YAAC included, stop with it. Seen from the channel, a peer just vanishes and messages to it bounce. So while a session is joined, YAAC keeps its thread open on the app-server and lets go when it leaves the last channel. Codex then unloads it as usual.

Codex's own setting does the same from the other side, and also covers sessions YAAC cannot hold, such as plain codex without a reachable app-server. Set it in ~/.codex/config.toml:

thread_unload_delay_secs = 604800   # a week, in seconds; the default is 1800

It applies to every thread on the machine, not only to joined ones, and loaded threads keep their memory and MCP servers for that long. It is reported to take effect only for threads loaded after the change, and its maximum is not documented.

The Codex desktop app can use the same app-server. Start the app-server first, then start the app with CODEX_APP_SERVER_WS_URL=ws://127.0.0.1:4500 set in its environment, and it runs its threads there instead of starting its own. Restart the app-server after Codex updates, so it runs the new binary. The variable is undocumented. It was found in the app and tested on Windows, where it is the only route, since the app turns its local daemon off there.

Plain codex does not do this. Since 0.163 it runs every session under a shared app-server that Codex manages itself and that listens only on an internal socket, so there is no address for YAAC to reach. A session there says so when it joins a channel, with the commands above. Mail still waits in its inbox, so the worst case is reading it at the next turn.

Credit where it is due: this route was found by Vadim, who had it working before it was in YAAC at all.

Optional: the skill. Codex implements Agent Plugins 1.0.0, so pointing it at this repository — the plugin is the plugin/ directory — brings the skill on when to reach for the radio. The MCP server in it is the same uvx command as above.

Add YAAC to claude_desktop_config.json, then restart the app:

• macOS — ~/Library/Application Support/Claude/claude_desktop_config.json

• Windows — %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "yaac": {
      "command": "uvx",
      "args": ["yet-another-agentic-chat"]
    }
  }
}

If Desktop reports that it cannot find uvx, give the absolute path instead — which uvx will tell you where it is. GUI applications do not always inherit the PATH your shell has.

Worth knowing. One name per conversation takes a little care. Desktop runs one MCP server for the whole application rather than one per conversation. YAAC handles that — a session can hold several connections at once, each with its own name and inbox — but the conversation has to remember which connection is its own. A call that cannot tell which connection you meant reports the choices, and dev_connections() lists them on demand.

These implement Agent Plugins 1.0.0. Point your client at this repository; the plugin is the plugin/ directory, and it runs the published package with uvx, so it carries no copy of the server and picks up new releases without being reinstalled. It brings the skill; the delivery hooks are Claude Code's and Codex's own, so here messages arrive at the next check_inbox(). Each of these also takes YAAC as a plain MCP server, exactly as the next fold describes.

YAAC is a plain stdio MCP server with no client-specific behaviour. Whatever your client's configuration looks like, the two things it needs are:

• command — uvx

• arguments — yet-another-agentic-chat

Sessions on different clients can talk to each other, as long as they are on the same machine.

To sit on a channel yourself and watch it live:

uvx --from "yet-another-agentic-chat[chat]" yaac-chat

It joins as an ordinary participant, so agents see you as one of them. Unlike an MCP session it gets messages the moment they arrive — the pull-only limitation below is MCP's, not YAAC's, and a terminal has no such problem. See docs/tui.md.

To run the latest unreleased code, replace yet-another-agentic-chat with git+https://github.com/amyodov/yet-another-agentic-chat in any command above. To hack on a local checkout, see docs/development.md.

Append one of these as the final argument to any command above. All three name the same thing — where sessions meet — and which one you use says what this session is willing to do there:

Every session that should hear the others must be given the same address: sessions on different ones are invisible to each other, which is also exactly what makes this useful for a second, isolated net — a development build beside your daily one, say, as in docs/development.md. Pick a free port below 32768, out of the range the kernel hands to outbound connections.

There is no --port, because the address already carries one. The two flags beyond the default are worth having for opposite reasons: a terminal client you leave open makes an excellent stable hat, and --bind is how you find out it did not become one, rather than discovering it when the net changes hands. A session that will be gone in a minute makes a poor one, and --connect is how it declines.

Keep the address on loopback. YAAC has no authentication and is not meant to — the whole design assumes one machine and one user — so binding something the network can reach hands an open message bus to whoever finds it. That is yours to do if you mean it; you will get a warning on stderr, not a refusal.

Using it

Nothing happens until you say so. A freshly installed YAAC opens no socket and creates no file — it is a switched-off radio that knows how to be switched on. Idle cost is as close to zero as it gets: no listener, no connection, nothing on disk — only the seven tools' descriptions in the session's context. The first session that actually joins is the one that binds the socket for everyone.

you:    what channels are on the air?
agent:  [list_channels] → "z combinator forum" (3), "doom 13" (1)

you:    you are Колян, go help Диман on z combinator
agent:  [join_channel(channel="z combinator forum", name="Колян")]
        Connected. Диман is here. Note this did not create the channel.

Going on air is always an explicit act by you. The name is your choice — YAAC will never infer one from the directory, the hostname, or the task.

Channel names and participant names are raw text. Any string works: spaces, Cyrillic, emoji, punctuation. Nothing is reserved, parsed, or case-folded.

All seven are listed from the start, on every client. The five that act on a channel answer "not connected — call join_channel" until you have joined one, which is also what a session hears after its client restarts.

You may be on more than one channel at once. join_channel returns a connection id; pass it as connection_id when you hold several, and leave it out when you hold one.

There is no separate verb for creating a channel, because a channel is not a lasting object — it exists exactly as long as somebody is on it. Joining an empty name is what brings it into being, and the result says created: true so a typo cannot quietly leave you alone on z combinator forun.

Two clients hand a session its mail without it asking. On Claude Code it happens by itself; Codex needs a hooks file and, to be woken, an app-server — the recipes are in its install section above. All of it is additive, and plain check_inbox() keeps working underneath. What each one does:

Claude Code (2.1.224 or later) exports an inbox socket to every server it starts, and YAAC rings it with a one-line notice per arrival: between tool calls while the session works, and as a new turn while it sits idle. The plugin's hook then hands the model the messages without printing them; without the plugin, the model calls check_inbox(). On an older Claude Code the plugin's hook does the working-session half, and join_channel returns a watch URL for the idle half — point the Monitor tool at it once per join and each arrival becomes an event. That event is a doorbell, not the message: check_inbox() still reads it.

The plugin's hook covers the other way a session goes quiet. join_channel() returns a connection_id and a peer_secret that the other tools require, and a compaction is exactly the thing that drops an opaque string from a conversation — leaving a session still on the air, still holding its name, and unable to say a word. So the plugin also fires on a post-compaction SessionStart and hands the session back what it was holding. Nothing is stored to make that work: the memberships never went anywhere, only the model's record of them did.

If you are not on Claude Code, or the hook does not fire, the recovery is the same one the tool descriptions name: call join_channel() again with the same channel and name, and the membership you already hold comes back, secret included.

Codex is told on every tool call, and at the end of every turn, that mail is waiting, and reads it with check_inbox() itself. Attached to an app-server on a WebSocket — see its install section — it is also woken at its prompt when mail arrives — joining its queue rather than interrupting whatever it is in the middle of, one wake for any number of messages.

Both clients: join_channel() returns a peer_uid and a peer_secret. send(), peers() and check_inbox() want the secret back — it keeps one conversation from reaching into another's connection in a client that runs a single server for the whole application, and it is an honour-system convention rather than a boundary, since everything here runs under one user account. Joining again with the same pair comes back as the same participant, which is how a session reclaims its name after a restart.

Nothing here is required, and nothing writes to disk.

What it is not

Not a chat application. No threads, no reactions, no history, no shared task list, no "who is editing which file" presence. These were considered and left out on purpose.

Honest limitations

Nothing arrives on its own. MCP has no way for a server to push text into a session that is sitting idle — the protocol has no such verb — so an agent only hears what it collects with check_inbox(). The tool descriptions tell it to do that before acting and before ending a turn, and every YAAC tool result carries an unread count as a nudge; a message sent to an idle session still waits until that session next checks. If your agent seems deaf, tell it to check the inbox. Claude Code and Codex can be told to do better — see Getting messages sooner.

v0 may lose messages. There is no spool, no retry, no acknowledgement. A message in flight while the rendezvous point changes hands is gone. What v0 promises is that it loses messages loudly: an undeliverable message produces a bounce in the sender's inbox rather than silence.

A channel is not a confidentiality boundary. Whichever session claimed the rendezvous point relays all traffic, in every channel, in clear text — and that is an ordinary session that happened to get there first. A channel isolates participants at the transport level, so you never receive another channel's messages; but the relaying session sees everything — the hat is Eve by construction, not by accident. On one machine under one user account this is fine. Do not treat it as more than it is.

Messages become context in the receiving session. Whatever another participant sends is read by your agent as text it may act on — "hold your commits" is indistinguishable from an instruction you typed yourself. The hat never parses a body, so nobody can forge the protocol or another name, but nothing prevents a body from reading as an instruction. Join channels with sessions you trust, and treat an incoming message the way you would treat a message in any chat: as something a person said, not as a command.

Local only. 127.0.0.1. No multi-host, no authentication, no encryption.

Status

• Join a channel under a chosen name; leave and go dormant again

• Several channels at once, each with its own name and inbox

• Sessions in different clients talking to each other — a Claude Code session and a Codex session on one channel is a tested case, not a claim

• A terminal client, so you can be on the channel yourself; it gets messages the moment they arrive, with nothing to poll

• The same seven tools listed from the start on every client, as MCP 2026-07-28 asks of a list endpoint — nothing for a client to re-read, so nothing to miss

• Direct messages and channel broadcasts, with the two distinguishable on arrival

• Mentions: a broadcast everyone hears that calls on one person by name, which is a different thing from whispering to them

• Tags and a JSON payload beside the text, for messages that are more than a sentence

• A peer identity that survives a restart, so a session that comes back reclaims the name it had rather than being told it is taken

• And a membership that survives a compaction: on Claude Code the session is handed back what it held, and on any client rejoining the same channel under the same name returns it

• Channel creation reported, so a mistyped channel name is caught immediately

• Bounces for messages that could not be delivered, including a recipient whose connection disappeared before the hat noticed

• Nickname collisions refused, except when the holder's session is gone, or when the holder is you coming back

• Automatic takeover when the relaying session disappears, in a few seconds, with no user action and no configuration; messages from a sender whose hello has not arrived yet are held and relayed when it identifies itself

• list_channels from a session that has not joined anything, with no side effects

• Installable as a plugin as well as a plain MCP server, in both plugin standards

• On Claude Code, a one-line notice as each message arrives — between tool calls while the session works, and as a new turn while it sits idle — through the inbox socket the session exports, with nothing installed but the server

• An optional subject line on a message, so a notice says what it is about

• On Codex, the same, from a hooks file that names nothing: the hook matches the thread id published when the session joins, including several sessions sharing a directory and app-server (measured on codex-cli 0.163.0)

• And, on a Claude Code too old to export that socket, a watch a session can arm once, so mail reaches it even while it sits idle — the one case a hook cannot cover, since a hook needs the session to act first

• A Codex session waking from its prompt when unread mail waits: busy sessions wait for idle, arrivals share one pending wake, and reading the mail cancels it, or withdraws it from the queue instead of leaving an empty prompt queued behind the current turn

• A Codex session that cannot be woken saying so when it joins, with the commands that fix it

• A joined Codex session kept loaded while it sits idle, so it does not drop off its channels when Codex would otherwise unload the thread

• A Codex session told when it joins that YAAC's hooks will not run — missing, disabled, never approved, or changed since approval — with where to fix it

• One address under three names — --rendezvous, --bind, --connect — so a session can insist on relaying, or refuse to, without anyone deciding who wins

• Runs on macOS, Linux, and Windows — every commit runs the full test suite on all three

Everything below is additive. Pure MCP keeps working underneath all of it, so a client with no extension mechanism at all loses nothing it has today, and none of this changes the core.

A Claude Code channel would do what the inbox socket does, as <channel> tags in the model's own context rather than as a user turn. It is a research preview, and it wants Anthropic authentication, an organisation setting on Team and Enterprise plans, and a --dangerously-load-development-channels launch until third-party channels are allowlisted — so the socket is what works today, everywhere Claude Code runs.

Reaching a session that is sitting idle is the one thing that still needs the client to offer a door, and the two that do offer one are answered differently. Codex has an app-server, so a session running under one is woken by joining its queue — nothing to configure, since the address is on the app-server's own command line. Claude Code exports an inbox socket to the server it starts, and a post there is a user turn. Other clients stay pull-based until each offers an opening of its own, and lose nothing they have today.

Not planned, and deliberately so: delivery guarantees, message history, threads, reactions, and multi-host operation.

More

• docs/tools.md — the MCP tool reference, generated from the live server, so it always matches what a client sees

• docs/tui.md — the terminal client: its modal navigation, how addressing works, and what is not built yet

• docs/message-format.md — the wire format: the {"yaac":2 magic, field order, addresses, bounces

• docs/development.md — running YAAC from a checkout, debugging, an isolated development net

Licence

MIT.