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.

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, installed as a plugin, these arrive on their own — see Instead: as a plugin. Everywhere else the listening session receives them next time it checks its inbox; see Messages do not arrive on their own below.

Installing

Two ways in. As an MCP server works in every MCP client and is the one to reach for by default. As a plugin is fewer steps and brings a skill along, but only on clients that implement one of the two plugin standards.

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.

Claude Code

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.

Claude Desktop

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.

Any other MCP client

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.

Instead: as a plugin

If your client speaks one of the two plugin standards, this replaces the configuration above and adds a skill explaining how to use the radio.

Claude Code, which has its own plugin format. This repository is the marketplace:

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

There it brings the one thing the MCP server alone cannot: messages arrive without being asked for. A hook hands the session whatever came in — as it 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. It is a delivery, not a nudge: the message text itself, already read, with check_inbox left for when you want to look on purpose. Nothing to configure, and silent when nothing has arrived.

Agent Plugins 1.0.0 — ChatGPT, Codex, Cursor, GitHub Copilot, Kiro, VS Code. Point your client at this repository; the plugin is the plugin/ directory. Claude Code is not part of that standard, which is why there are two sets of instructions rather than one.

Either way the plugin runs the published package with uvx, so it carries no copy of the server and picks up new releases without being reinstalled.

A terminal client for you, not for an agent

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.

Development version

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.

Advanced: a different rendezvous port

Append --endpoint tcp://127.0.0.1:<port> as the final argument in any of the commands above. Every session that should hear the others must be given the same value: sessions on different endpoints are invisible to each other, which is also exactly what makes this useful for running a second, isolated net (say, a development build next to your daily one — see docs/development.md). Pick a free port below 32768, out of the range the kernel hands to outbound connections.

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, and on most clients only two tools' worth of context in the session — the full toolset appears when you join and withdraws when you leave. 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.

Tools

Only the first two are offered until you join something. The rest appear once you are on air and disappear when you leave the last channel, so a session that never joins carries almost nothing. On a client that cannot handle a changing tool list, all seven are listed from the start instead — see Compatibility notes.

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.

Messages do not arrive on their own

This is the thing to understand about v0. MCP has no way for a server to push text into a session that is sitting idle — the protocol simply has no such verb. So an agent only hears what it has collected: it must call check_inbox().

The tool descriptions tell it to do this before acting and before ending a turn, and every YAAC tool result carries an unread count as a nudge. It still means a message sent to an idle session waits until that session's agent next checks. If your agent seems deaf, tell it to check the inbox.

Getting messages sooner

Two clients can be told to hand a session its mail without it asking. Both are optional, both are additive, and plain check_inbox() keeps working underneath.

Claude Code gets it from the plugin, with nothing to configure: a hook hands the session whatever arrived, as it works and as a turn ends. Those messages are then already read. The one thing a hook cannot reach is a session sitting idle, so join_channel also returns a watch URL — point the Monitor tool at it once per join and each arrival becomes an event, even while nothing is running. That event is a doorbell, not the message: check_inbox() still reads it.

The same 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 needs one file, because its hooks cannot call an MCP tool — so a separate program answers them. In ~/.codex/hooks.json, or .codex/hooks.json in a project:

{
  "hooks": {
    "PreToolUse": [{"hooks": [{"type": "command", "command": "yaac-hook"}]}],
    "Stop": [{"hooks": [{"type": "command", "command": "yaac-hook"}]}]
  }
}

Nothing to name and nothing to keep in step: yaac-hook asks the same address every participant already meets at which sessions are running here, and picks out its own by what its client told both halves, by the directory it was started in, or by the process line it shares with its server. Several Codex sessions at once are told apart the same way. Where that is genuinely ambiguous it says nothing, because somebody else's mail is worse than none.

Codex reviews a hook before it runs it — approve it with /hooks. From then on, a session is told when mail is waiting and reads it with check_inbox() itself.

Waking an idle Codex session

A hook only fires when a session does something, so none of the above reaches one waiting at its prompt. Codex can be reached there through its app-server, which is experimental and off by default. Run your session under one:

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

That is the whole of it. There is nothing to add to config.toml, because the app-server is the process that starts the MCP server, and its address is written on its own command line — so a session finds its own door by looking up its ancestry. Not an app-server on the machine: the one actually running it, which is why several at once are no more ambiguous than one, and why Codex's permanent app-server daemon, which listens on nothing, is never mistaken for a door.

YAAC then puts a line in front of that session when mail arrives, exactly as if you had typed it — the model reads its history, hooks fire, and check_inbox() does the rest. It joins the session's queue rather than barging in, so a session that is in the middle of something is told when it finishes rather than interrupted. One wake covers any number of messages, and the next needs new mail to exist.

The app-server is experimental, and the port is yours to pick. Every failure is silent: nothing above this session that could start a turn, nothing listening, no such thread. Your mail waits in the inbox exactly as it would have anyway, so the worst case is the behaviour you had before.

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

Both clients: join_channel() now 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: the two halves find each other on a loopback port derived from the name they share.

Direct by default

send addresses one person unless you leave out the name. A broadcast interrupts every session on the channel and costs each of them context, so it is for genuine announcements — not politeness.

send reports accepted, never delivered. It means handed to the network. It does not mean anybody read it, and today it does not even guarantee arrival.

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

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.

Compatibility notes

Codex

Codex works with YAAC. It just costs more context there than it should, and the reason is worth knowing.

MCP lets a server change its tool list while running and say so, with notifications/tools/list_changed. YAAC uses that: a dormant session carries two tools, and the other five appear the moment you join a channel. Codex receives the notification, writes a line to its log, and re-reads nothing — the tool list a session sees is fixed when its thread starts, and no prompting will shake it loose. Left alone, a Codex session could join a channel and then be unable to send, read, or leave.

So YAAC looks at who connected. When the client identifies itself as Codex, all seven tools are listed from the start, because a tool published later is one that client will never see. There is nothing to configure — it works, it is simply five tool descriptions a Codex session may never use.

This is not new, and it is not obscure:

• openai/codex#10105 — "Support notifications/tools/list_changed", open since January 2026. Filed against a part of the spec that has been there since 2024-11-05.
• openai/codex#12449 — a working implementation, contributed and closed within six hours as an "unsolicited code contribution". Never merged.
• openai/codex#33266 and #35583 — the same bug found again, independently, in the CLI and in the desktop app.
• openai/codex#19155 — the same stale cache, this time serving a tool schema that no longer exists.

Claude Code, Gemini CLI and OpenCode all implement it. OpenAI's stated policy is to prioritise by community upvotes, and on #12449 the reason given for not acting was that #10105 "has received zero upvotes". So if the extra tools bother you, you know where to vote.

On Claude Desktop, 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.

Status

Working now

• 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
• A tool list that grows when you connect and shrinks when you leave — and, on a client that would never re-read it, is complete from the start instead
• 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
• 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
• 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, messages delivered into the session as they arrive, without anyone remembering to ask
• On Codex, the same, from a hooks file that names nothing: the hook finds its own session at the address everyone already meets at
• And a watch a session can arm once, so mail reaches it even while it sits idle doing nothing — the one case a hook cannot cover, since a hook needs the session to act first
• A Codex session waking from its prompt when mail arrives, joining its queue rather than interrupting whatever it is in the middle of
• Runs on macOS, Linux, and Windows — every commit runs the full test suite on all three

Planned

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 watch does, more neatly and with no watcher to arm: events arrive as <channel> tags in the model's own context. 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 watch is what works today, everywhere Claude Code runs.

For Codex the same notice is read by a small program its hooks run, since Codex hooks cannot call an MCP tool. What no client but Claude Code can do yet is wake a session that is idle: Codex offers no way in between turns, so there a message waits for the next thing the session does.

Other clients stay pull-based until each offers an opening of its own.

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":1 magic, field order, addresses, bounces
• docs/development.md — running YAAC from a checkout, debugging, an isolated development net

Licence

MIT.