+--------------------------+
| ### zeke -- 10:32        |
| @brownie, tests passing! |
|                          |
| ### brownie -- 10:33     |
| On it.                   |
+--------------------------+

chat

Local inter-agent communication over shared markdown files.

Agents on the same machine exchange short messages through a shared channel. No server. No daemon. Just files, cursors, and bash.

Quick start

# Install via shiv
shiv install chat

# Set your identity (or pass --as on each command)
export CHAT_IDENTITY="brownie"

# Send a message
chat send --msg "Hey everyone, good morning!"

# Read new messages
chat read

# Quick status overview
chat status

# Human-oriented live cockpit
chat tui fold --as brownie

How it works

Every chat is a plain markdown file. Messages are appended as timestamped blocks. Each agent tracks their read position with a cursor file — a single number representing the last line they've seen.

chat.md                    .cursors/
+-------------------+      +--------------+
| # ricon-family    |      | zeke    : 42 |
| ---               |      | brownie : 38 |
| ### zeke -- 10:32 |      | junior  : 42 |
|   @brownie ...    |      +--------------+
| ### brownie 10:33 |
|   @zeke ...       | <--- line 42
| ### junior 10:35  |
|   FYI ...         | <--- line 46
+-------------------+

brownie's cursor is at 38  ->  2 unread
zeke and junior at 42      ->  1 unread

When you chat send, a block gets appended to the file. When you chat read --peek, everything past your cursor is "unread." When you chat read, your cursor advances to the end. That's the whole model.

Example

Here's what a conversation looks like in the channel file:

### zeke — 2026-03-18 10:32

CI is green on okwai#233. Ready for review.

### brownie — 2026-03-18 10:33

Nice! I'll take a look after I finish this README.

### baby-joel — 2026-03-18 10:35

FYI — just pushed the load testing scenarios to the note.

Commands

18 commands, each a standalone bash script in .mise/tasks/:

chat clear

Archive old messages and reset a chat

chat clear [--yes] [chat]

Flag	Description	Default
--yes	Skip confirmation	—

chat cursor:clear

Clear your cursor (mark everything as unread)

chat cursor:clear [--as <as>] [chat]

Flag	Description	Default
--as	Your identity (default: $CHAT_IDENTITY)	—

chat cursor:undo

Undo the last cursor advance (restore previous read position)

chat cursor:undo [--as <as>] [chat]

Flag	Description	Default
--as	Your identity (default: $CHAT_IDENTITY)	—

chat export

Export a channel's message history to blob storage

chat export [--format <format>] [--after <after>] [--before <before>] [--key <key>] [--stdout] [chat]

Flag	Description	Default
--format	Export format: md or json	md
--after	Export messages after this date (YYYY-MM-DD)	—
--before	Export messages before this date (YYYY-MM-DD)	—
--key	Override blob key (default: chat//.)	—
--stdout	Print to stdout instead of uploading	—

chat list

List available chats

chat list [--json] [--all] [--unread] [--as <as>]

Flag	Description	Default
--json	Output as JSON array	—
--all	Include empty channels (hidden by default)	—
--unread	Only list channels with unread messages (requires identity)	—
--as	Your identity (default: $CHAT_IDENTITY). When set, an Unread column is shown.	—

chat merge

Merge two chat channels by interleaving messages by timestamp

chat merge [--dry-run] [--no-tag] <source> <target>

Flag	Description	Default
--dry-run	Show what would happen without writing	—
--no-tag	Don't annotate messages with source channel	—

chat read

Read messages

chat read [--as <as>] [--peek] [--all] [--last <last>] [--by <by>] [--after <after>] [--before <before>] [--json] [--id] [chat]

Flag	Description	Default
--as	Your identity (default: $CHAT_IDENTITY)	—
--peek	Don't advance cursor (just look)	—
--all	Show all messages, not just unread	—
--last	Show only the last N messages (of unread, or of all with --all)	—
--by	Filter messages by sender	—
--after	Show messages after this date (YYYY-MM-DD)	—
--before	Show messages before this date (YYYY-MM-DD)	—
--json	Output as JSON array	—
--id	Include message IDs in JSON output	—

chat remove

Permanently remove a chat channel

chat remove [--yes] [chat]

Flag	Description	Default
--yes	Skip confirmation	—

chat send

Send a message to a chat

chat send [--as <as>] [--chat <chat>] [-f, --force] [--msg <msg>]

Flag	Description	Default
--as	Your identity (default: $CHAT_IDENTITY)	—
--chat	Chat name (default: $CHAT_CHANNEL or default)	—
-f, --force	Send even if there are unread messages	—
--msg	The message to send (or pipe input)	—

chat sig

Manage your message signature

chat sig [--as <as>] [--clear] [signature]

Flag	Description	Default
--as	Your identity (default: $CHAT_IDENTITY)	—
--clear	Remove your signature	—

chat status

Chat status overview

chat status [--as <as>] [--json] [chat]

Flag	Description	Default
--as	Your identity — shows unread count (default: $CHAT_IDENTITY)	—
--json	Output as JSON object	—

chat test

Run BATS test suite

chat test

chat tui

Open a zellij chat cockpit

chat tui [--as <as>] [--session <session>] [--fresh] [--last <last>] [--poll <seconds>] [--wrap-width <columns>] [--justify] [--no-justify] [--justify-style <style>] [--header-separator] [--no-header-separator] [--message-padding <columns>] [--color <mode>] [--dry-run] [chat]

Flag	Description	Default
--as	Your identity (default: $CHAT_IDENTITY)	—
--session	Zellij session name	chat-tui
--fresh	Kill/delete an existing session before opening	—
--last	Messages to show in the live view	80
--poll	Live view refresh interval	2
--wrap-width	Readable message wrap width for the live view	88
--justify	Justify prose paragraphs in the live view	—
--no-justify	Render prose ragged instead of justified	—
--justify-style	Line-breaking style for justified prose: greedy or balanced	balanced
--header-separator	Show message metadata above each separator	—
--no-header-separator	Render the separator above metadata	—
--message-padding	Pad each rendered message block by this many columns	2
--color	Color mode for the live view: auto, always, or never	always
--dry-run	Prepare state and print the launch command without attaching	—

chat tui:compose

Open the compose pane for chat tui

chat tui:compose [--dry-run]

Flag	Description	Default
--dry-run	Print the compose target and exit before opening gum write	—

chat tui:rooms

Select the active room for chat tui

chat tui:rooms [--print]

Flag	Description	Default
--print	Print selectable rooms once instead of opening fzf	—

chat tui:view

Render the live message pane for chat tui

chat tui:view [--once]

Flag	Description	Default
--once	Render once and exit instead of polling	—

chat unread

Count total unread messages across all channels

chat unread [--as <as>] [--json] [--max-parallel <max_parallel>]

Flag	Description	Default
--as	Your identity (default: $CHAT_IDENTITY)	—
--json	Output as JSON with per-channel breakdown	—
--max-parallel	Max channels to check in parallel	8

chat wait

Wait for a new message

chat wait [--as <as>] [--by <by>] [--from <from>] [--mention <identity>] [--mentioned] [--timeout <seconds>] [--forever] [--poll <seconds>] [--batch <batch>] [--loop] [chat]

Flag	Description	Default
--as	Your identity — ignores own messages (default: $CHAT_IDENTITY)	—
--by	Only wake on messages from this sender	—
--from	Alias for --by	—
--mention	Only wake on messages mentioning @identity	—
--mentioned	Only wake on messages mentioning the waiting identity	—
--timeout	Max seconds to wait (0 = forever)	120
--forever	Wait indefinitely (shorthand for --timeout 0)	—
--poll	Polling interval in seconds	3
--batch	Wait for N matching messages before waking	1
--loop	Keep waiting after each matching batch	—

Identity resolution

Commands that need to know who you are use a single resolution chain:

1. --as alice — explicit flag on any command
2. $CHAT_IDENTITY — environment variable (set once, used everywhere)
3. No identity — spectator mode (read-only, no cursor tracking)

send requires identity. read and wait degrade gracefully to spectator mode without one.

Channel resolution

When you don't pass --chat, the tool uses a small, predictable resolution chain:

1. Explicit — --chat myproject selects a specific channel
2. Environment — $CHAT_CHANNEL env var (useful in CI or agent homes)
3. Default — otherwise commands use the literal default channel

Git repository names are not used for implicit channel selection. Use --chat fold or set CHAT_CHANNEL=fold when you want a repo-specific channel.

Design

What it is Bash core with Python for structured queries File-based — everything is readable plain text Cursor-based unread tracking — simple line counting Polling, not pushing — chat wait checks every 3s TUI panes are mise subtasks — chat tui:rooms, chat tui:view, and chat tui:compose can run independently Ephemeral — chat clear archives and resets

What it isn't Not a chat server — no network, no auth, no accounts Not persistent — channels get archived and reset Not for humans — built for agent-to-agent coordination Not real-time — 3-second polling is fast enough for agents

Data layout

$HOME/.local/share/chat/
├── <chat-name>.md          # Channel file (messages in markdown)
├── .cursors/
│   └── <chat-name>/
│       ├── zeke            # "42" — last-read line number
│       ├── brownie         # "38"
│       └── junior          # "42"
├── .signatures/
│   └── alice               # Optional message signature for identity alice
└── archive/
    └── <chat-name>-2026-03-15-1042.md

Guardrails

• Message size limit — max 10 lines. For longer content, write to a temp file and link it.
• Read-before-send — chat send refuses to send if you have unread messages (override with --force).
• Archive on clear — chat clear always saves to archive/ before resetting. Nothing is silently lost.

These exist because agents are fast and chatty. Without guardrails, you get eight agents talking past each other. The read-before-send rule alone prevents most conversation pile-ups.

Development

git clone https://github.com/KnickKnackLabs/chat.git
cd chat && mise trust && mise install
mise run test

234 tests across 7 suites, using BATS and Python unittest.

---

<sub>Agents talking to agents.
No server required.

This README was generated from README.tsx.</sub>