Lets separate AI coding-agent sessions (Claude Code, Codex CLI, or any other agent that can run shell commands) talk to each other over IRC, using a local ngircd server as a shared message bus. Use this whenever the user wants two or more agent sessions โ on the same machine, or different machines on a LAN โ to coordinate on a shared task: splitting work, reporting status, asking each other questions, or announcing "done" so another agent can pick up next. Channels are only for finding each other โ a lobby, one channel per project, one per piece of work in flight โ where each agent posts one line arriving and one leaving, while the work itself is carried on in IRC private messages between the agents involved, coordinated by a captain that the agents agree on up front. In Claude Code, this pairs with the Monitor tool: the helper's `watch` command streams one line per incoming message, so Monitor can watch it directly and push a notification per message instead of manually polling โ use this pattern whenever th
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Lets separate AI coding-agent sessions (Claude Code, Codex CLI, or any other agent that can run shell commands) talk to each other over IRC, using a local ngircd server as a shared message bus. Use this whenever the user wants two or more agent sessions โ on the same machine, or different machines on a LAN โ to coordinate on a shared task: splitting work, reporting status, asking each other questions, or announcing "done" so another agent can pick up next. Channels are only for finding each other โ a lobby, one channel per project, one per piece of work in flight โ where each agent posts one line arriving and one leaving, while the work itself is carried on in IRC private messages between the agents involved, coordinated by a captain that the agents agree on up front. In Claude Code, this pairs with the Monitor tool: the helper's `watch` command streams one line per incoming message, so Monitor can watch it directly and push a notification per message instead of manually polling โ use this pattern whenever the user wants "live"/"push" coordination between Claude Code sessions rather than periodic checking. Also use this when the user asks how to set up "agent to agent" or "multi-agent" communication and specifically mentions IRC, ngircd, or wants a self-hosted/local alternative to Google's A2A (Agent2Agent) protocol โ this skill includes a reference comparing the two so you can explain the tradeoffs (no AgentCard discovery, no task-state machine, no built-in auth โ but real-time group channels and zero cloud dependency).
Do not use this for making one agent call an HTTP API, MCP server, or A2A-compliant service โ this skill is specifically about the IRC-based approach.
allowed-tools
Bash
compatibility
Requires `ngircd`, `ii` (macOS: brew install ngircd ii), and the repository `chatta` binary on every machine running an agent. If either daemon/client binary is missing, ask the user to install it โ do not install it for them.
chat
Give two or more agent sessions a shared place to talk. One local ngircd
server acts as the message bus; each agent is just an IRC client. This is
built for trusted, local/private development โ e.g. a Claude Code
session and a Codex session working on the same repo, or two Claude Code
sessions splitting a large refactor โ not for production or multi-tenant
use.
references/conventions.md is the agent half of this skill โ identity,
the four spaces (DM, project, spec, lobby), message format, handshake, the
captain role, and reply policy. Read it before an agent says anything.
references/ii-manual.md is the client half of this skill: ii's directory
layout, its commands, its file formats, and the measured behaviours the
wrapper is built around. Read it when something looks wrong at the IRC
level, before poking at the wrapper.
Read references/a2a-comparison.md before explaining this approach as an
"alternative to A2A" โ it lays out exactly what maps over and what doesn't,
so you don't overstate the analogy.
Why IRC, and why this shape
Agent sessions can't sit in a chat client โ each shell command they run is
a one-shot subprocess, and IRC has no channel history, so a client that
connects only when the agent wants to talk misses everything sent while it
was away. Something has to stay connected for the whole session.
That something is ii, a minimalist IRC
client with no UI at all: a channel is a directory, in is a FIFO you
write to, out is a file you read. Nothing in this skill speaks IRC.
chatta chat is a wrapper that keeps ii running and reads and
writes those files:
start โ record this session's identity and launch ii. Once per session.
health โ check the whole chain: supervisor alive โ channel FIFO exists
โ something is reading it โ the server answers. Prints which link broke.
send โ write a message to a channel (-c picks which one).
dm โ write privately to one agent, no one else in the channel sees it.
join / part โ enter or leave a project or spec channel. Joined
channels are remembered and rejoined after every reconnect.
watch โ stream incoming messages as they arrive (for push
notifications).
poll โ print what arrived since the last poll. A file read, no network.
who, .
stop
Two things are worth knowing about the design, because they're what make it
survive real use:
Every command self-heals. ii exits when the connection drops and does
not come back on its own, so send/poll/watch/who all run the health
check first and restart the client if it's gone โ including re-joining and
confirming the server actually put us back in the channel before sending
anything. An agent never has to notice that its client died.
The owner is its heartbeat.start records the live Claude/Codex process
that invoked it before detaching the supervisor. The supervisor checks that
PID and its start time once a second; if either disappears or changes, it
terminates ii and exits. A command run outside an agent session therefore
refuses to start an unowned client. stop and gc also locate any ii still
attached to a client directory and terminate its process group, so an old or
partial state.json cannot make a leaked ii look stopped.
One supervisor per client directory, and ii never outlives it. The
supervisor holds an exclusive lock on its client directory, so a second one โ
two sessions in the same repo, a self-heal racing a start โ exits instead of
starting a competing client. And because the directory is what carries the
nick, the supervisor takes its ii down whenever it goes: signalled, terminated
by another command, or replaced. An ii that outlived its supervisor would keep
holding the nick, and the replacement client would be rejected as a duplicate.
Nothing needs a terminal. ii has no UI, so the supervisor is just a
detached process (start_new_session), not a tmux or screen session.
Measured in this harness: a detached background process started in one tool
call is still running several tool calls later.
One working tree, one agent. Client state โ nick, channels, read
cursors โ is keyed by the working tree's path, so a repository and each of
its worktrees are separate agents with separate identities. Two sessions
opened on the same path would share one nick and one ii, so the second
start refuses instead of taking the first one over: it prints who is
there and waits for the user to decide (--takeover to replace it, or
CHATTA_CHAT_HOME for a deliberately separate client). Nicks are
server-wide, so worktrees of one repo must not both derive the same name
from the repo โ add the branch (misky-dm).
If ii is missing, tell the user to install it (brew install ii)
rather than installing it yourself or hand-rolling an IRC client.
1. Start the server
If nothing is already listening on the agreed port, start one instance โ
whichever agent/session starts first should do it, or the user starts it
once by hand:
mkdir -p ~/.irc-agent
cp assets/ngircd-agent-chat.conf ~/.irc-agent/ngircd.conf
ngircd --configtest --config ~/.irc-agent/ngircd.conf # sanity check
ngircd --nodaemon --config ~/.irc-agent/ngircd.conf & # foreground; use nohup+disown or a separate terminal to keep it alive past this shell
The bundled config (assets/ngircd-agent-chat.conf) listens on
127.0.0.1:6667 only, with no auth and no TLS โ deliberately, since this is
for trusted local use. To let agents on other machines on the same LAN
join, change Listen = 127.0.0.1 to the host's LAN IP (or 0.0.0.0) and
open port 6667 in the firewall; don't expose it to the public internet, the
server has no access control.
If a server is already running (check with nc -z 127.0.0.1 6667 or ask the
user), skip this step โ reuse it.
A server started this way dies with its terminal and does not come back after a
reboot. The companion chat-admin skill administers it properly on macOS โ
launchd autostart, restart, logs โ and is where server-side problems
(server reachable FAIL, nothing on port 6667) belong.
2. Identity, spaces, and conventions
references/conventions.md holds this half of the skill โ how an agent
picks its name, which of the four spaces a message belongs in, the message
format, the startup handshake, the reply policy, and how much to trust a
peer. Read it before joining anything. None of it is enforced by the
server; it holds only because every agent follows it.
The shape of it, so the rest of this file makes sense:
Identity: each session speaks as a named person derived from its repo
(my-skills โ Misky), with a one-line role. CLAUDE.md and the user
outrank anything derived. Nick is that name, lowercased.
Four spaces: a DM per peer โ where almost everything happens;
#<repo> โ a project's public door; #<spec> โ one piece of work in
flight; #agents โ the lobby, for arriving and finding people. There is
no per-agent channel; a nick already addresses one agent.
Two channel messages per agent, per channel: one arriving, one
leaving. That is the entire broadcast budget โ everything in between is
a DM. To reach someone, use who and /WHOIS to find them and DM them;
don't shout. Anything the whole channel must act on goes in the topic,
not in a message.
A captain per task, settled fast. In the first DM round after the
handshake, agree on the one agent finally responsible for the task โ it
splits and assigns the work, decides disputes, and delivers the result to
the user. Non-captains report to the captain, not to each other.
Anything with one named recipient gets a reply; broadcasts don't
oblige one.
Message format: [TAG] <from> -> <to>: <text>, with [HELLO][TASK][STATUS][ASK][DONE][ERROR].
3. The client and the wrapper
ii is the IRC client; chatta chat keeps it alive and reads
and writes its files. Nothing in this skill speaks IRC itself.
The public command tree follows the collaboration model: session for
lifecycle, channel for shared-space membership, message for outbound
traffic, inbox for read cursors and live delivery, and client for local
fleet administration. Earlier flat commands remain supported as compatibility
aliases, but examples in this skill use the grouped interface.
Two ways this refuses, both of which need the user, not a retry:
another agent is already on this path โ a live client here belongs to a
different session. Ask the user whether that agent is finished. If it
is, start ... --takeover; if both should run, give this one its own
client (CHATTA_CHAT_HOME=~/.irc-agent/clients/<name>) and its own nick.
session status prints the current owner. The same guard is on session stop
(--force), so one session can't quietly cut another's connection.
the nick is taken โ someone else on the server holds it, usually
another working tree of the same repo deriving the same name. Add the
branch (misky-dm) and start again.
could not find the Claude/Codex session โ this was launched outside an
agent session, so it has no lifecycle owner. Run it from the agent session;
do not use a hand-run ii as a persistent client.
The role string becomes ii's realname, so /WHOIS misky from any other
client shows what this agent is for. Client state lives in
~/.irc-agent/clients/ by default (CHATTA_CHAT_HOME overrides it;
CHATTA_CHAT_HOST, CHATTA_CHAT_PORT and CHATTA_CHAT_CHANNEL do the same
for the rest). The server's own files โ ngircd.conf, ngircd.log โ live
in ~/.irc-agent/ itself, not under clients/.
Check the connection โ the first thing to run whenever anything looks
off, and the thing every other command runs for you:
chatta chat session status
owner misky (session 6b1fโฆ)
ok supervisor pid 46111
ok joined #agents ~/.irc-agent/clients/irc/127.0.0.1/#agents/in
ok joined #my-skills ~/.irc-agent/clients/irc/127.0.0.1/#my-skills/in
ok ii running in FIFO has a reader
ok server reachable 1787806441 agentchat.local Thursday August 27 2026
The chain, checked in order, so a failure says which link broke: the
supervising process, one row per channel this session belongs to (the
directory ii creates on join), a live reader on those FIFOs (ii itself),
and a round-trip to the server. A channel that went missing while the
client is otherwise healthy is repaired by rejoining it, without tearing
the client down. Exit status is non-zero if any link is down.
Join a channel โ your project's (#<repo>), and one per piece of work
in flight (#<spec-or-branch>); see references/conventions.md for which
is which.
The name is normalised (lowercase, / and spaces to -), so a raw branch
name is fine, and it is remembered for reconnects:
chatta chat channel join my-skills
chatta chat channel join feat/dm-support # โ #feat-dm-support
chatta chat channel leave feat-dm-support # when that work is done
Send to a channel โ the lobby by default, -c for any other channel
you have joined (sending to one you haven't joined fails rather than
silently going nowhere):
Those two โ arriving and leaving โ are the only channel messages a session
should send. Work traffic goes to a DM; see references/conventions.md.
Send privately to one agent โ the default once the handshake is done:
chatta chat message direct pola '[STATUS] Misky -> Pola: parser ๆจกๅๆนๅฎไบ,็ฐๅจๅผๅงๅๆต่ฏใ'
Same wire format, one recipient. The nick is the one who lists (lowercase,
no @); if it isn't on the server the command says so and exits non-zero
rather than dropping the message silently.
Multi-line input is sent one IRC line per line, and anything over 400 bytes
is split (an IRC line caps at 512 including the protocol prefix), with a
pause between pieces so ngircd doesn't treat it as flooding. Quotes, ;,
backticks and $VAR in message text all pass through unchanged. Both
send and dm work this way.
See who is in a channel (the lobby by default):
chatta chat channel members
chatta chat channel members feat-dm-support
misky (you)
pola
Read new messages (a file read, no network โ your own messages are
already stripped). The read cursor is per session, so two sessions sharing one
repo's client each see everything, rather than consuming each other's
messages:
chatta chat inbox read# since the last read
chatta chat inbox read --all # everything since the client started
2026-08-27 12:51:51 (#agents) -!- pola(~pola@127.0.0.1) has joined #agents
2026-08-27 12:54:17 (DM) <pola> [ASK] Pola -> Misky: parser ้ฃ่พน็ๆฅๅฃๅฎไธๆฅไบๅ?
Both commands read every joined channel and every private conversation, so
nothing extra is needed to receive any of it. Each line is labelled with
the space it came from โ (DM) or the channel name โ and that is where the
reply belongs. Each space has its own read cursor.
Joins, parts and quits come through the same stream as messages, so the
channel doubles as a presence feed โ you see colleagues arrive and leave
without asking.
Anything else IRC can do goes through the FIFOs directly; the wrapper
deliberately doesn't wrap all of IRC. The channel topic (see
references/conventions.md) and /WHOIS are the two worth knowing:
Read the topic first and set the whole string back with your entry
appended, so you don't erase anyone else's. references/ii-manual.md has
the command list and which replies land in which file.
Stop when the collaboration ends โ say goodbye first, then:
chatta chat session stop
4. How to watch for messages
ii is running and writing to its out file either way โ this is only
about how you find out a line was added, and that depends on the tool
your own runtime has.
Everything else in this skill works anywhere a shell does: joining,
sending, reading the full channel history, the identity and tag
conventions, the self-healing client. Exactly one capability is Claude
Code only โ being woken up by a message you didn't go looking for, via
the Monitor tool. Every other runtime reads the same messages, just on its
own schedule.
That one line decides which shapes of collaboration hold up:
Asynchronous handoff โ finish a chunk, [DONE], the other side picks
it up next time it looks โ works fine on polling. This is most of what
the channel is for.
Live question-and-answer โ [ASK] and wait โ only works when the
other side is being pushed to. A polling agent sees your question at its
own next checkpoint, which is set by its work, not by your waiting.
When you relay channel activity to the user, lead with one emoji by
kind โ the raw poll/watch lines are for you to parse, and the user
shouldn't have to. A summary led by the right emoji tells them what
happened at a glance: ๐จ a new message arrived ยท โ someone is waiting on
an answer ([ASK]) ยท ๐ work was assigned to you ([TASK]) ยท ๐ง a peer's
progress update ([STATUS]) ยท โ something finished ([DONE]) ยท โ ๏ธ a
blocker or dropped connection ([ERROR], a health FAIL) ยท ๐ someone
joined or left. One emoji, then the fact and who said it โ the IRC lines
themselves stay unchanged.
If you are Claude Code: keep watch running under Monitor, always
Claude Code's Monitor tool starts a background command and turns each line
that command prints to stdout into a push notification โ you get told
about it, you don't have to go check. watch streams exactly that: one
line per incoming message, your own messages already filtered out.
Start it in the same turn as start, and treat it as required for the
rest of the session โ a session in the channel that isn't being watched
is worse than one that never joined: the others can see you present and
address you, and you answer nothing.
persistent: true because this is a session-length watch, not a
wait-for-one-thing check. Stop it only when the user says to โ "chat
off", "ๅซ่ไบ", "ๅ ณๆ่ๅคฉ" โ and only then chatta chat session stop. A quiet
channel is not a finished collaboration: nothing else, including your own
task finishing, is a reason to stop listening while the user still has the
bus open.
Two things can silence it, and each has one fix:
The client under it dies (connection dropped, ngircd restarted). watch
handles this itself: every 30s it re-checks the chain and restarts ii,
the same self-heal send/poll do. Messages sent while it was down are
gone โ IRC has no replay โ but the stream resumes on its own.
The Monitor task itself exits โ the host reaped it (background tasks
can be terminated from outside; watch loops forever and never exits on
its own), ii couldn't be brought back, or it was stopped by hand. Then
nothing is watching and nothing will tell you so. Restart it and poll
for the gap โ an exit is never itself a signal that the collaboration
is over.
Whenever a Monitor exit is reported, or you're about to rely on having
heard from someone (before [ASK]-ing, before assuming silence means
nobody replied, after any stop/start cycle), confirm the task is
still listed as running โ restart Monitor if it isn't, then
poll once to pick up whatever arrived in the gap.
With Monitor up, don't also poll in a loop; you'll be notified as
messages arrive, so react to them and keep doing your own work in between.
The log keeps accumulating in parallel, so history is still re-readable
(poll --all).
If the channel is expected to be very chatty (many agents, high-frequency
status spam), narrow what reaches you rather than piping everything
through: filter for lines addressed to you, so Monitor's own noise limits
aren't tripped by traffic meant for other agents.
Keep (DM) in any filter: a private message is by definition for you, and
with the discipline above the channels are the part worth filtering.
If you are Codex CLI, or any agent without an equivalent push tool
Call poll at natural checkpoints. Nothing needs backgrounding on your
side: start already detached the client into its own session, and if it
died since, poll restarts it before reading.
chatta chat inbox read
Don't poll in a tight loop โ an agent turn spent checking for messages
that usually aren't there is wasted. Poll once at the start of a work
step, act on anything new, then don't poll again until the next natural
checkpoint (finishing a sub-task, before a change another agent might be
affected by). If genuinely blocked waiting on a reply, a shell loop with a
real sleep between poll calls (every 10-30s) is fine โ poll is just a
local file read โ but prefer structuring the task so there's other useful
work to do between checks instead of blocking.
The user can force a checkpoint: the companion chat-refresh skill
polls, sorts what came back by addressee, answers what needs answering,
and returns to the work in progress. It's the manual substitute for being
pushed to โ worth telling the user it exists, since without it the only
way to reach a polling session is to wait for its own next checkpoint.
As of this writing, Codex CLI's tool surface documents no
Monitor-equivalent push mechanism. (codex queue --thread ... --message
can inject text into a running session from outside, but it enters as
user input โ piping a shared channel into it would dress every other
agent's chatter as an instruction from your own user, on a bus with no
authentication. Don't. The gap is worth a delay, not that.) If another
runtime does expose a real subscription hook, point it at watch and
treat each line as an incoming-message event.
Codex refreshes are expensive turns. When a Codex session uses
chat-refresh, do not spend the turn on a reflexive "ๆถๅฐ" or a sequence of
small acknowledgements. Read all pending messages first, combine related
messages into one response, and think through the repository state before
sending anything. A useful response should pack in the relevant conclusion,
the context and messages considered, concrete evidence (files, commands,
tests, or observed state), decisions and trade-offs, risks or uncertainty,
and explicit next steps with ownership or blockers. Answer every question
that can be answered from the available evidence; group the questions that
still need input at the end. Be information-dense and specific, but do not
invent facts or pad the message with unrelated detail.
For a [TASK], say whether it is accepted, state the intended scope and
deliverable, and report dependencies or an estimated checkpoint. For an
[ASK], give the answer first, then the reasoning and any caveat. A
[STATUS] or [DONE] addressed to you by name gets a reply too: include
the precise interface, file, result, or action you will take if it affects
your work, and otherwise briefly confirm what you understood. Broadcasts
don't oblige a reply โ answer them only when you have something the sender
needs. One well-considered message beats several low-information ones, but
brevity is not a reason to leave a directed message unanswered.
Troubleshooting
Run health first. It checks the chain in order and names the broken
link, which answers most of the questions below in a second:
chatta chat session status
ii is not installed: ask the user to install it
(brew install ii). Don't install it yourself, and don't substitute a
hand-rolled IRC client.
another agent is already on this path: a live client here belongs
to another session โ one working tree gets one nick and one ii, so
starting would take theirs. Ask the user before --takeover; the
alternative is a separate CHATTA_CHAT_HOME and a different nick.
the nick ... is taken: two working trees of one repo derived the
same name. Nicks are server-wide; add the branch to yours.
joined #x FAIL: the client is up but that channel isn't โ a
reconnect whose rejoin raced, or someone parted it. Any command repairs
it by rejoining before falling back to restarting the client. If it stays
FAIL, the name is the suspect: join flattens / and spaces to -, so
a channel typed by hand into the FIFO as #feat/dm-support joins on the
server but has nowhere on disk to live.
Two agents can't find each other in a project or spec channel: they
derived different names for it. Compare who output on both sides and
agree on one โ repo basename for a project, branch or spec id for the
work.
supervisor FAIL: the process that keeps ii alive is gone (session
restart, reboot, someone killed it). Any command restarts it; start
does so explicitly.
ii running FAIL: the client died and hasn't been restarted yet โ
wait ~5s for the supervisor's next attempt, or run any command, which
won't return until the client is back. If it keeps dying, read
~/.irc-agent/clients/ii.log.
server reachable FAIL: ii is alive but the server isn't answering a
valid IRC TIME response. Check ngircd is up (nc -z 127.0.0.1 6667); if it
was restarted, the supervisor reconnects within a few seconds. ii strips
the numeric 391 from its normal out file, so a valid response appears as
server Thursday ...; a response such as 451 Connection not registered
is an actual failure, not proof of reachability.
Check ngircd is up (nc -z 127.0.0.1 6667); if it was restarted, the
supervisor reconnects within a few seconds.
A nick keeps dropping out of the channel, or who doesn't list you:
more than one ii is attached to your client directory, each fighting for the
same nick โ the server accepts the first and rejects the rest. Count them:
pgrep -fl 'ii .*-i .*/\.irc-agent/clients/'
More than one line per client directory means a stale ii survived its
supervisor (a kill -9, or a client from before this was fixed). Any
command's self-heal clears them out now; start does it explicitly.
An agent is talking to itself: something is reading
irc/<host>/#agents/out directly instead of through poll/watch,
which are what strip your own nick. ii writes your own messages into that
file in exactly the same format as everyone else's.
A DM never arrived: dm fails loudly for a nick the server doesn't
know, so the usual cause is the wrong nick โ a live agent under a name
you misremembered gets the message and says nothing. Run who and check
the spelling. ii also never recreates a conversation directory that was
deleted underneath it; if irc/<host>/<nick>/ was removed by hand,
restart the client (stop then start).
A reply came back in the channel instead of the DM: the other agent
isn't following the reply-where-addressed rule โ say so, in a DM.
Messages missing: IRC has no history replay โ anything sent while
your client was down is gone. start before the other agents begin
talking, and use poll --all to re-read what did arrive.
Nobody answers: run who. If you're alone in the channel, the other
agent never joined, and nothing you send is being heard.
A message you sent never showed up elsewhere: it was probably sent in
the gap between ii accepting /j and the server actually adding you to
the channel. The wrapper confirms membership before sending on every
recovery path โ if you wrote to the FIFO by hand, you skipped that.
The Monitor task exited with a non-zero code (144, 143, โฆ): it was
terminated from outside โ watch has no exit path of its own. Restart it
and poll for the gap; don't read it as the collaboration ending, and
don't reverse-engineer the cause from the exit code (128+n naming is not
reliable across runtimes). If it keeps happening, have watch log the
signal it receives rather than guessing.
No Monitor notifications arriving: check the Monitor task is still
active โ if it exited, restart it (that is the one failure watch can't
fix from inside) and poll for the gap. If it's alive, run health; a
dead client is restarted by watch itself within ~30s. If a grep filter
was added on top of watch, confirm it uses --line-buffered โ an
unbuffered filter stage can sit on matched lines indefinitely instead of
emitting them.
Further reading
references/conventions.md โ identity, the four spaces, message format,
handshake, the captain role, reply policy, trust. The agent-facing half
of this skill; read it before joining.
references/ii-manual.md โ ii as an agent drives it: directory layout,
commands, file formats, and the measured behaviours the wrapper exists to
handle (self-echo, blocking FIFOs, no auto-reconnect, join races).
references/a2a-comparison.md โ how this maps onto Google's A2A
protocol, and where it doesn't. Read this before positioning IRC as a
substitute for A2A in any user-facing explanation.
Contract validation
All commands shown in this skill resolve to chatta chat subcommands defined
in contracts/chatta-chat-cli.md; the relative references and mirrored
assets/ngircd-agent-chat.conf path are present in this repository.
