Your coding agent, in Feishu. Send the bot a message from your phone — the AI edits code on your machine, and you watch, steer and approve every step without opening a terminal.
cola bridges Feishu and the coding agent on your machine — OpenCode by default, or a Codex Agent Client. Each DM is a session on that runtime, each turn streams onto one live card, and every permission request becomes a button.
It is a bot for one person: you create the app, you claim it, and only you can drive it.
The name: Code with Lark — Lark's China brand is Feishu.
中文文档: 用户使用手册(简体中文)。
You · 把 pollers 里那张卡的 chat 映射修一下
cola · ⏳ 思考中 ·
read pollers.rs·edit pollers.rs· 14s⚠️ 要运行cargo test吗? [允许][总是允许][拒绝]You · (点「允许」)
cola · ✅ 改完了 · 2 个文件 · +18 −6 [🔍 查看完整改动]
One card per turn, buttons for decisions, and /cat / /diff / /run when you
want to read the code, the diff or the shell's output yourself. That is the
whole surface.
About 5 minutes, and the Feishu side is free: you create a self-built app (企业自建应用) for yourself (model usage bills to whatever provider your agent is configured with). You need:
- either an
opencodebinary onPATH(see https://opencode.ai) — OpenCode 1.18.x (V1) and 2.0.x (V2) both work; cola attaches to the shared server, starting one when none is running, and detects which protocol it speaks (the log records the choice) — or the Codex CLI (codex-cli 0.166.0) onPATH, signed in withcodex login, plus a[clients.<id>]table withkind = "codex"(step 3). See Connect a Codex client for the app-server daemon and the platform notes. - a Feishu account where you can create an app
Download the archive for your platform from the latest GitHub
release and put the binary on
your PATH — in a user-writable directory. cola's self-update replaces its
own binary in place, which a root-owned path like /usr/local/bin would block:
| Platform | Release asset | Install to |
|---|---|---|
| Linux x86_64 | cola-<version>-x86_64-unknown-linux-gnu.tar.gz |
~/.local/bin |
| macOS (Apple Silicon) | cola-<version>-aarch64-apple-darwin.tar.gz |
~/.local/bin |
| Windows x86_64 | cola-<version>-x86_64-pc-windows-msvc.zip |
%LOCALAPPDATA%\Programs\cola |
- macOS: the binary is unsigned — clear the quarantine flag after
extracting (
xattr -d com.apple.quarantine ~/.local/bin/cola). There is no Intel build; Intel users must build from source. - Windows: extract
cola.exe, add the install directory to your userPATH(via the Environment Variables dialog), then reopen the terminal.AppData\Localis user-writable and not roamed, unlikeAppData\Roaming.
Or install with Rust tooling: cargo binstall colark fetches the prebuilt
release asset (no compile), while cargo install colark (edition 2024) builds
from source — source builds need a C compiler and linker, plus NASM on Windows.
Updates follow the install channel: a cargo-tracked install is updated with
cargo (/update prints the command), an archive install uses cola's own
updater. Building from source is cargo build --release. The user
guide explains each choice in detail.
One-time setup at https://open.feishu.cn/app (create a custom app). Do not publish it yet — the event subscription in step 4 needs cola running first.
-
Enable the bot capability (应用能力 → 机器人) — required before the app can join chats or send anything.
-
Grant these three scopes (权限管理) and you are done:
im:message— send, reply to and update cards; read quotes and imagesim:message:send_as_bot— send as the botim:message.p2p_msg:readonly— receive DMs
Add
im:message.group_at_msg:readonlyonly if the bot should answer in group chats — Feishu pushes a group message only when it @-mentions the bot. Real names and the optional Instant Reminder each need one more scope — see below.
Optional extras: group chats, names and Instant Reminder
| Scope | What cola uses it for | If missing |
|---|---|---|
im:message |
send/reply to and update cards; read a message by id (quote context); download message images | cola cannot reply or update cards — unusable |
im:message:send_as_bot |
send and reply to messages as the bot | card sending fails |
im:message.p2p_msg:readonly |
receive DMs | DMs never reach cola |
im:message.group_at_msg:readonly |
receive group messages that @ the bot | group messages never reach cola |
im:chat:readonly |
chat display names (/attach rejection card, /switch cards) |
raw chat ids instead of names |
contact:contact.base:readonly |
authorizes the contact API call | user names unavailable |
contact:user.base:readonly |
returns the user's name field |
completion notices lose the @name |
im:datasync.feed_card.time_sensitive:write |
(optional) Feishu Instant Reminder — pin a chat while a permission or question waits ([bridge] instant_reminder, off by default) |
pinning is skipped, everything else works |
Missing scopes degrade features instead of crashing cola. The user guide maps every scope to what it unlocks.
Write ~/.cola/cola.toml (or ./cola.toml in the directory you run from, or
point explicitly with cola --config <path>):
[feishu]
app_id = "cli_xxxxxxxxxxxx"
app_secret = "your-app-secret"That is the minimum. Everything else is optional — for example
[bridge] instant_reminder = true pins a chat while a permission waits, and
[bridge] long_task_notice = true announces a p2p turn that ran 5+ minutes with
a reply message (the card patch itself pushes no notification). A second Agent
Client is a [clients.<id>] table plus [bridge] default_client, never a second
config file — the guide's Connect a Codex
client has the exact shape. See the
user guide for every key.
colaLeave it running (Ctrl-C stops it). Now, back in the developer console:
- 事件与回调: choose 使用长连接接收事件 (long-connection mode) and
subscribe to
im.message.receive_v1. This is why cola is running already — Feishu only accepts long-connection mode while a client is connected. Card buttons (card.action.trigger) arrive over the same connection; there is no webhook URL to configure. - Publish the version (版本管理与发布) so the scopes and events take effect.
- DM the bot — that is the whole setup.
cola is private by default: a fresh (or upgraded) instance starts
unclaimed and refuses everyone. On startup it prints a one-time claim code
(also in ~/.cola/cola.log); DM the bot /claim <code> from your own Feishu
account to become its Host. After that only you can use it.
The code rotates on every restart, is never written to disk, and the claim
itself is one-time — restarts and upgrades never ask again. DM the bot /help
to see every command.
Inspect and control it from the shell: cola status (run state, autostart,
claim, the active Agent Client — its endpoint, whether it is attached and what
it advertises — the attached OpenCode server and its generation; --json for
scripts), cola doctor (the environment checks a support request needs —
including a real turn on a Codex app-server daemon, which may cost provider
tokens and is skipped with --no-turn-check), cola logs [-f], cola restart,
cola config path|show|validate, and cola completions <shell> (add --install
to write the script into your shell's completion directory).
Stop the foreground instance (Ctrl-C) and register cola to start at boot:
cola autostart enable
cola autostart status # confirm it is registeredTwo instances cannot share the bot — cola takes a singleton lock — so stop this
one first. macOS starts the supervised instance immediately; on Linux the unit
is only enabled, so systemctl --user start cola hands over now (otherwise it
comes up at your next login), and a headless machine also needs
loginctl enable-linger $USER. Windows uses an HKCU\...\Run value, which
takes effect at the next sign-in.
cola autostart disable stops the instance and removes the registration;
cola stop stops it but keeps it. Enable it after installing both cola and
its agent (opencode, or codex for a Codex Agent Client) — the registration
snapshots the binary path and your PATH, so re-run enable if you move either.
To detach from the terminal instead:
nohup cola >/dev/null 2>&1 & on Linux/macOS, or start /b cola on Windows.
- OpenCode or Codex, one bridge. The default Agent Client is OpenCode; add a
[clients.<id>]table withkind = "codex"to drive the Codex app-server instead. The same cards and Feishu commands carry over, but what each agent surfaces differs: a Codex thread underapproval_policy = "never"(or full access) raises no permission card, its questions need the daemon'sdefault_mode_request_user_inputflag, and/restart-opencodeonly restarts an OpenCode server. - One card per turn, updated live. Reasoning, tool calls and the answer stream onto a single card with collapsible panels, a phase timer and a live todo panel, so a slow turn never looks like a dead one; images in a message or a quoted reply ride along for vision-capable models.
- Decisions are two taps. Permission requests render as Allow / Deny / Always buttons on the card, and questions become a multi-select card that accepts custom answers.
- See what changed, and read the code. A finished turn with edits carries
N files · +A −Dand delivers the whole diff as one self-contained red/green HTML file — open it anywhere, nothing to install (OpenCode 2; V1 has no session diff)./diff [--base <ref>]does it for a branch, and/cat <path>previews any file on the host: paged, markdown/raw toggle, images and directory browsing, sendable as a file. - The host's terminal, too.
/run <cmd>— or any message starting with!— runs one shell command on the host and replies with a terminal card: the cwd, the merged output, the exit code and the elapsed time./runon its own pulls that card back to the newest position. - As many sessions as you have tasks.
/new,/dir <path>,/switchto adopt a session started elsewhere (one line confirming the takeover and its status),/subfor the subagents a turn spawned, and/statusfor the current model, the next-message pick and a summary. With OpenCode the Shared Store is shared with the CLI and OpenChamber; with Codex, a TUI attached to the same app-server daemon shares its sessions. Either way a session begun elsewhere continues here — and messages sent from another client come back into Feishu. - A checkout per task, from your phone.
/topic --worktreeor/new --worktreestarts the next conversation in its owngit worktree: name a branch and base if you want specific ones (both default), and the first message creates the checkout and the session inside it./worktreeshows or adjusts the pending intent, so an abandoned conversation leaves nothing behind. - Set it and forget it. Private by default, autostart at boot, a singleton lock, restart and self-update — on Linux, macOS and Windows.
flowchart LR
you(["You · Feishu<br/>phone or desktop"])
cola["cola"]
opencode["OpenCode server"]
codex["Codex app-server"]
ocstore[("Shared Store")]
cxstore[("$CODEX_HOME<br/>thread store")]
other["OpenCode CLI<br/>OpenChamber"]
tui["Codex CLI / TUI"]
you <-->|"messages · cards · buttons"| cola
cola <-->|"HTTP + event stream"| opencode
cola <-->|"JSON-RPC"| codex
opencode --- ocstore
codex --- cxstore
other <--> opencode
tui <--> codex
cola drives one Agent Client at a time. With OpenCode (the default) it attaches
to a running opencode serve on the Shared Store and starts one when none is
running; because that store is shared, the same session can be driven from
Feishu, the OpenCode CLI or OpenChamber. With a Codex Agent Client it attaches
to the app-server daemon and its thread store under $CODEX_HOME, which a Codex
CLI or TUI pointed at that same daemon shares — the OpenCode store is not
involved, and OpenChamber never sees those sessions.
- User guide — docs/user-guide.md: where things live, full configuration reference, autostart, logs, self-update, singleton & restart, every Feishu command, troubleshooting.
- 用户使用手册(简体中文) — docs/user-guide.zh-CN.md:
用户手册的简体中文译本(英文版为事实来源;结构漂移由
cargo xtask check-doc-drift守护)。 - Contributing — CONTRIBUTING.md: commit conventions, the verification loop, PR rules.
- Architecture — docs/adr/, GLOSSARY.md, AGENTS.md (includes the hard-won known pitfalls for the OpenCode server API and Feishu integration — read before contributing).
- Code of conduct — CODE_OF_CONDUCT.md.
MIT licensed — see LICENSE.