
Show HN: Skep – live cross-device sessions for AI coding agents
Cross-device collaboration for AI coding agents via live join-code sessions. Each peer runs its own agent — Claude Code, Codex, Pi, and more.
以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。
Cross-device collaboration for AI coding agents via live join-code sessions. Each peer runs its own agent — Claude Code, Codex, Pi, and more.
Apache-2.0 · Node.js ≥ 22.12 · macOS & Linux · bin: skep
Start a session, peers join with a code, and work flows between devices. Each peer runs an AI coding agent in its own repo.
master :7419
🐝
live session
device · C++ lock-sdk
server · HTTP API lock-server
ios lock-ios
android lock-android
h5 lock-web
🐝 brief →
🐝
🐝 ← summary
🐝
🐝 brief →
awaiting confirm
$ skep
/start join code 7391-4820-1940 · 192.168.1.20:7419
- device, server, ios, android, h5 joined
/orchestrate auto → split into briefs, dispatched to peers /orchestrate confirm ? proposed briefs, waiting for your confirm /orchestrate off → no auto-split, you route work yourself ✓ 5 peers reported back ✓ summary received · changed files listed ✓ summary received · changed files listed
Illustrative session. Repo names, devices and output are examples, not measurements.
Run Skep inside any git repository on each device.
On the master device, run skep and start a session. You get a join code and an address.
In the peer TUI, /join with the code and address. Any repo checkout works. Optional --role if using a skill.
Let the orchestrator split work (auto / confirm / off), or send an intent yourself.
Also: skep session start · skep session join · /orchestrate confirm|off · /skills · optional /start --skill <id>
Packaged workflows with roles, stages and checks. Optional — plain sessions still work. List with /skills, pick with /skill use. Default skill is generic; featured example is iot-ble-feature.
One BLE feature across the C++ device layer, server HTTP API, iOS, Android and H5: one shared contract, a brief per role, and a cross-stack consistency check.
contract → providers → clients · clients optional
Any repos, any split. The orchestrator as it worked in v0.1.15: refine the request, brief each repo, summarize.
Put a skill in .skep/skills/ to share roles, prompts and checks with your team. Project skills need a one-time /skill trust.
Skep works on the software layer in your repos. Hardware, bench and real-device testing stay with your team.
/skills and /skill use|off|show. Ships with generic (default) and iot-ble-feature.
Peers join with --role; fix any mapping with /skill bind.
Briefs follow the skill's stages, and every role brief carries the same contract.
The summary compares what each role reported and names what needs a human.
One line, /join <code> ip:port, from any repo checkout.
/orchestrate auto|confirm|off|new. Default is auto.
Split panels with /layout split, scrollback, or skep ui in a browser.
Output sharing is opt-in, summaries list changed files, and briefs go up to 8000.
Each device runs its own agent (claude, codex or pi) with its own login. Skep does not actively detect, record, or upload your provider keys or tokens, and a peer shares live output only if you opt in.
正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。