An engineer defines once what the agent may do. Teammates @ it in a Feishu group with one sentence; it runs on the local machine and posts the result back.
Each agent has its own role, the material it may access, and its engine. An engineer defines it once; whoever the team @s in Feishu shows up and does the work.
Edits campaign landing pages, tweaks styles, spins up preview links.
Reviews PRs, runs tests, suggests changes.
Answers "how does this feature work?" without touching code.
Checks monitors, pulls logs, pinpoints incidents first.
Summarises core metrics every Monday and posts to the ops group.
Product, ops and data each @ the agent they need in Feishu — no more queueing to interrupt an engineer.
Draft a PRD, digest user research, break down requirements — in one sentence.
Query monitors, pull logs, generate campaign reports, locate issues first.
Pull, clean, model, visualise — results land straight in the group.
Transparent and auditable end to end — code and credentials never leave the machine, no API key is injected, your existing subscription login is reused.
An engineer sets its role, boundaries, and which repos, material and workflows it may touch.
A teammate states the request in plain language — no terminal, no process to learn.
Claude Code / Codex runs on the engineer's machine with the real context.
Answers, preview links, progress and files are posted back to Feishu — previewable, reusable.
Code and credentials are never handed to a cloud service. Your existing Claude Code / Codex subscription is the auth; no extra tokens to buy.
All code is public, PRs welcome. Safe enough to let it touch local code.
Reuses your local subscription login — no injected API key, no extra tokens.
Code and credentials stay on the host; nothing passes through a third-party cloud.
Pick Claude Code or Codex per bot; more engines can be added smoothly.
Once the host is running, the team can @ it in Feishu. Current line: v0.3.x = Agent Workspace Runtime.
# install globally $ npm i -g larkway # check environment / subscription login $ larkway doctor $ larkway init # start the bridge; the team can now @ it in Feishu $ larkway start ✓ bridge online · Feishu connected
An open-source (MIT) thin bridge that turns your locally logged-in Claude Code or Codex into a Feishu (Lark) agent your whole team can @-mention. Feishu is the entry point, Larkway is the bridge, Claude Code / Codex is the engine.
No. Larkway reuses your local Claude Code / Codex subscription login. It never injects an API key and you do not buy extra tokens.
No. The agent runs on your own machine. Code and credentials stay local; only messages and cards travel between Larkway and Feishu.
No. Feishu pushes events over a persistent WebSocket; outbound access is enough.
Yes. An engineer defines a bot once (permissions, role, which repo); product, ops and data teammates @ it directly in Feishu.
Yes. @ the bot from the Feishu mobile app; the agent runs on your computer. See the post on dispatching a fleet of local agents from your phone.
Setups, rough edges, how to shape agents for your team — discussed and solved together. Mostly Chinese-speaking; English is welcome, or open an issue on GitHub.