OpenCockpit — The open Claude Code GUI for any LLM
One seat. One AI. Everything under control.
/ˈkɒkpɪt/ — like an aircraft cockpit
English · 中文 · Website · Blog
OpenCockpit is the open Claude Code GUI for any LLM — an IDE-like workbench for the whole dev loop. Run multi-project Claude sessions out of the box; pop open a tab for Codex, Grok Build, DeepSeek, GLM, Kimi, or local Ollama whenever you need, side by side, with no environment variables to edit. Built-in terminal, Chrome control, PostgreSQL / MySQL / Redis bubbles, code review, and slash modes — all local. And it’s web client–server: self-host it on a shared dev box and every teammate gets a seat.
https://github.com/user-attachments/assets/18f1a5dc-64f3-4ff6-b9fc-9cd08181fbb8
npm i -g @surething/cockpit && cockpit
Why Cockpit?
Anthropic ships Claude Code as a CLI. That’s the right call for power users — but the moment you have more than one project in flight, your terminal turns into mission control with no instruments.
Cockpit is the instrument panel. It does not replace Claude Code; it stands on top of the official Agent SDK and gives you the things a CLI can’t:
| Pain with raw Claude Code |
What Cockpit adds |
| Stuck on one model |
7 engines side by side in tabs: Claude (default), Codex, Grok Build, DeepSeek, GLM, Kimi, local Ollama — each its own session |
| One session at a time, terminal chaos at 3+ projects |
Multi-project tabs, parallel agent sessions, red-dot inbox, desktop notifications |
| Image attachments are awkward |
Drop / paste images straight into chat |
| “What was I debugging yesterday?” |
Cmd+K cross-project session browser, pinning, forking; /ss finds any past session from one sentence |
| Every subagent starts with a blank slate |
Bots: persistent, file-native subagents you tag with @name; each runs in its own inspectable session |
| Agent can’t reach your browser / DB |
Smart Bubbles: Chrome, PostgreSQL, MySQL, Redis — drivable by the agent |
| Reading an unfamiliar repo means a 90-min file-tree scavenger hunt |
Code Map chip view — caller / callee pins, click to walk the call graph |
| Reviewing AI output is friction |
Line-level comments on any file, send any comment back as AI context |
| Same “do X but don’t change code” prompt every day |
Built-in Skills /qa /spec /design /pair /fx /ex /go /cg /cc /cr + custom SKILL.md via the Skills sidebar |
| No automation hooks |
One-time / interval / cron-based scheduled tasks |
| “Cloud relay” trust concerns |
Fully local. No telemetry. API keys (DeepSeek / GLM / Kimi) stay in ~/.cockpit/<engine>/credentials.json on your laptop. |
How it compares
An honest snapshot as of July 2026 — each tool wins somewhere. Spotted an error? PRs welcome.
|
OpenCockpit |
Claude Code Desktop (official) |
Opcode |
| Positioning |
IDE-like workbench for the whole dev loop |
agent session companion |
session manager for Claude Code |
| Architecture |
✅ web client–server — deploy on a shared dev box; the whole team codes with AI in parallel, each on their own project / worktree |
single-user desktop app |
single-user desktop app |
| Open source |
✅ MIT |
❌ closed source |
✅ AGPL-3.0 |
| Engines |
✅ Claude + Codex / Grok Build / DeepSeek / GLM / Kimi / Ollama (BYOK) |
Claude only |
Claude only |
| Parallel multi-project sessions |
✅ |
✅ |
✅ |
| Agent-drivable browser & DB (Chrome / Postgres / MySQL / Redis) |
✅ Smart Bubbles |
❌ preview pane only |
❌ |
| Fully offline / air-gapped |
✅ via Ollama |
❌ |
❌ |
| Phone / tablet access |
✅ any LAN browser — code runs on your machine |
✅ via Claude Code on the web (cloud sandbox) |
❌ desktop only |
| Native desktop app |
❌ local web app (needs Node ≥ 22.19) |
✅ |
✅ Tauri |
| Newest Claude Code features on day one |
⏳ tracks Agent SDK releases with a lag |
✅ first party |
❌ |
| Automation triggers |
one-time / interval / cron |
✅ Routines: cron + API + GitHub events |
background agents |
| Session checkpoints / rewind |
pinning & forking only |
✅ |
✅ checkpoint timeline |
| Per-tool-call change snapshots (real disk diff for every Edit / Write / Bash / MCP call, 7-day local history) |
✅ |
❌ |
❌ per-prompt checkpoints only |
| Usage / cost analytics |
basic token counts |
n/a (plan-based) |
✅ full dashboard |
| Cost |
GUI free (MIT); AI billed by whichever engine you bring — can be $0 with local Ollama |
app free; needs a paid Claude plan or API billing |
GUI free (AGPL); needs a paid Claude plan or API billing |
| Actively maintained |
✅ |
✅ |
⚠️ last release Aug 2025 |
Pick Claude Code Desktop if you live inside the Anthropic ecosystem and want first-party polish. Pick Opcode if you want a native desktop feel with checkpoints and cost analytics. Pick OpenCockpit if you want more than a chat window — an open-source, IDE-like cockpit that covers the whole dev loop (read code, drive your browser and databases, review, automate) with any engine you bring, even from your phone.
Features
Engines — Claude by default, bring any LLM you want
- Claude (default) — full official Agent SDK; zero setup if
claude CLI is already configured
- Codex — reuses your
~/.codex config; same chat, same shell + bubbles
- Grok Build — ships a pinned Grok CLI and uses your subscription login and native coding tools through ACP. Select Grok from the new-chat menu; The model and reasoning menus use a fixed list (Grok 4.7, 4.6 and 4.5; Low through Extra High) and save choices per session. Commands and edits run with full access, like Claude/Codex. Session resume, streaming, tool results, and background-task continuations are supported. Independent tasks use isolated native context; prefix/single-turn forks and turn deletion synchronize both native and display history. Exact edits of turns already replaced by native compaction are refused to avoid retaining excluded content.
- DeepSeek — DeepSeek key + live model picker (
deepseek-v4-pro by default) and a balance check
- GLM (Zhipu / BigModel) — BigModel key + live model picker (
glm-5.2 by default) and a Coding Plan quota check; one key, two hosts — switch Region between open.bigmodel.cn and api.z.ai
- Kimi (Moonshot) — Kimi Code key + live model picker (
kimi-for-coding, k3, …) and a quota check; no kimi CLI needed
- Ollama — auto-starts the daemon; pick any pulled model from the chat header; fully offline
- Each engine lives in its own tab with its own session history; switch from the new-tab dropdown
- API keys stored locally in
~/.cockpit/<engine>/credentials.json; no cloud relay
Agent — AI chat that scales
- Default engine powered by the official Claude Agent SDK — zero extra setup
- Multi-project concurrent sessions with desktop notifications and red-dot badges
- Session pinning, forking, cross-project session browser (Cmd+K)
/ss — find any past session (any project, engine or date) from a one-line description; the reply links straight to it/dl — delegate a sub-task to a new session in another directory and/or on another engine, without waiting; ask later how it went. /dl --wait waits for the result and carries on in the same turn- Bots — tag a persistent, file-native subagent with
@name; it runs in a separate session with its own identity, context, commitments, and Skills
- Per-tool-call snapshots — every file-touching tool call is snapshotted (shadow git, fully local); review each reply’s changes as a git-history-style timeline, including what
Bash did
!command prefix to run shell from chat — output piped back as context- Image attachments, code references, token usage tracking
Explorer — Code & files
- 4-tab file browser: Directory · Recent · Git Changes · Git History
- Syntax highlighting (Shiki) with Vi mode editing
- Git blame, diff view, branch switching, worktree management
- LSP integration — go to definition, find references, hover info
- Code Map — every function as a chip with caller / callee pins; click to walk the call graph. Multi-language: TS/JS, Python, Go, Rust. No LSP, no project setup, works offline.
- CodeGraph — a code graph for the agent: same tree-sitter index exposed as 10 HTTP endpoints — 6 base (
search / callers / callees / impact / file / coedit) + 4 analytics (context / related / risk / affected, powered by PageRank · PPR · TF-IDF · Louvain, zero training) — so the agent queries coordinates instead of grepping text; trigger with /cg from chat. Coordinates only, never source — precise, cheap, and catches the conventional coupling no regex can.
- Fuzzy file search (Cmd+F), JSON viewer, Markdown preview
Console — Terminal & smart Bubbles
- Full xterm.js terminal with shell integration
- 🌐 Browser Bubble — control Chrome via accessibility tree (click, type, navigate, screenshot, network)
- 🐘 PostgreSQL Bubble — browse schema, run queries, export
- 🐬 MySQL Bubble — browse databases & tables, run queries
- 🔴 Redis Bubble — browse keys, inspect values, execute commands
- Drag-to-reorder, grid / maximized layout, per-tab env vars & shell aliases
- Line-level comments on any file
- Send any comment back to AI as context for an automated fix
Built-in Skills — guide how the agent works
/qa — Clarify requirements: produce a lean PRD with goals, scope, assumptions, and decisions; no solution design or code/spec — Specify: turn agreed requirements into behavior, contracts, and acceptance criteria; no implementation/design — Technical design: ground the solution in existing code, define responsibilities and tradeoffs, and plan implementation and verification; no implementation/pair — Programming partner: use the conversation and code context to explore, plan, design, code, review, verify, and explain as needed; give evidence-based advice when asked to take a look, and implement within the agreed scope when explicitly asked/fx — Diagnose: trace evidence to the root cause, then give blast radius, fix options, and verification steps; no code changes/ex — Deep analysis: answer first, then evidence, options, and open points; no clarification pauses or code changes/go — Implement: build agreed work in small verified slices, record decisions the spec did not cover, and report blockers and residuals/cg — Explore code: use CodeGraph symbols, call relations, and co-edit history to locate code, trace calls, and assess change impact/cc — Cockpit CLI: open and control browser bubbles, observe terminal output read-only, and query the codegraph; production instance by default/cr — Code review: independently review static correctness and state, timing, and concurrency; report a solution diagram, severity-ranked findings, and coverage; hard fans out by dynamic slice- Custom: drop any
SKILL.md and add it via the Skills sidebar — it auto-appears in the slash autocomplete menu (see Skills)
Scheduled tasks — Cron for AI
- One-time, interval, or cron scheduling
- Pause / resume, reorder, track results across projects
Skills — extensibility
- Drop in any
SKILL.md to teach the agent a new trick
- Invoke with
/skill-name from chat
- All managed from a single Skills sidebar
Bots — persistent subagents
- Create a Bot with
/bot, then register its ordinary local directory from the Bots panel or its BOT.md
- Tag it into any task with
@name; every turn runs in a separate, inspectable session and returns a link
- Memory is deliberate: normal turns are read-only, and files change only when you explicitly ask to remember, update, correct, or forget
- Attach an existing Skill with
@name attach /path/to/SKILL.md; review stale memory with @name review
- Bot directories stay local and portable Markdown — no hidden database, upload, or vendor lock-in
Read the Bots documentation.
Use cases
- Solo dev, multi-repo: “I have a refactor running in API, tests writing in Web, and a bug investigation in Pipeline — all at once, all visible.”
- A specialist that remembers: tag
@product for a roadmap decision today and again next month; it reads the same explicit, reviewable context without polluting the main chat.
- Day one in an unfamiliar repo: Open it in Code Map, click through caller/callee pins, walk the auth flow in five clicks instead of a 90-minute file-tree scavenger hunt.
- Team on a shared dev box: self-host one Cockpit where the code lives — every teammate opens a browser, takes a seat, and runs their own AI sessions in their own project / worktree.
- Reviewing AI-generated PRs: Switch the changed files into Code Map — changed functions are highlighted with their callers / callees still drawn around them, so blast-radius is one click away.
- Full-stack chore mode:
/fx in one tab on a backend bug, /ex in another to plan the frontend refactor, /go to land it — three slash modes, three different agent postures.
- Cheap second opinion: Same prompt in two tabs — Claude in one, DeepSeek v4-pro in the other — compare answers before you trust either.
- AI-driven QA: Browser Bubble + scheduled task = “every night at 2 AM, run this UI smoke flow and post a summary”.
- Privacy-sensitive code: runs on your laptop. Pair with an Ollama tab for fully air-gapped sessions. No telemetry, no relay.
Try online
No install, no AI chat (read-only sandbox, 5 min):
Prerequisites
- macOS, Linux, or Windows — Windows support is experimental (details)
- Node.js ≥ 22.19 — nodejs.org
- Claude Code — docs.anthropic.com/en/docs/claude-code (default engine)
- Git — for git features (blame, diff, worktree, etc.)
- Chrome (optional) — for Browser Bubble; install the bundled extension from
~/.cockpit/chrome-extension
Optional, per non-Claude engine
- Codex — log in once with
codex login to populate ~/.codex
- Grok Build — the official Grok Build CLI is installed with Cockpit. Run
grok login to sign in once. If it is not on PATH, use ~/.grok/bin/grok login (or npx --no-install grok login from this checkout). A separate global CLI installation is unnecessary. Cockpit launches its bundled binary with automatic updates disabled and uses the cached subscription login, including when XAI_API_KEY is set in the host environment. Display history is stored under $COCKPIT_HOME/grok/sessions (default ~/.cockpit/grok/sessions); native context and credentials stay in the CLI’s data directory.
- DeepSeek — get an API key at api-docs.deepseek.com; paste it into the engine picker
- GLM (Zhipu / BigModel) — get an API key (
<id>.<secret>) at bigmodel.cn/apikey/platform; paste it into the API Key field of the model picker in the chat header. The same key works on both hosts, so Region (open.bigmodel.cn / api.z.ai) is just routing — it defaults from your UI language and can be overridden in the picker.
- Kimi (Moonshot) — get a Kimi Code API key (
sk-kimi-…) at kimi.com/code/console; paste it into the API Key field of the model picker in the chat header. (A platform.moonshot.cn Open Platform key is a different product and will not work.) The kimi CLI is no longer used.
- Ollama — install ollama.com and
ollama pull <model>; Cockpit auto-starts the daemon
API keys are stored locally, each in its own file at ~/.cockpit/<engine>/credentials.json (e.g. ~/.cockpit/kimi/credentials.json, ~/.cockpit/glm/credentials.json) — never in settings.json. No cloud relay.
Install
npm install -g @surething/cockpit
cockpit # start cockpit → http://localhost:3457
cockpit . # open current dir as a project
cockpit ~/my-project # open specified dir
cockpit -h # help
Running in the background
cockpit runs in the foreground, so closing the terminal stops it. To keep it
running:
cockpit start # start in the background
cockpit status # running? which version, port, pid, uptime
cockpit restart # restart in place
cockpit stop
start accepts the same flags as cockpit (--port, --token). Output goes
to ~/.cockpit/logs/server.log.
restart asks the running server to replace itself, so the port and token it
was started with are preserved — you do not need to repeat them.
There is no launchd/systemd integration: after a reboot, run cockpit start
again (or add it to your login items).
Updating
cockpit update
…or click the version pill in the sidebar when an update is available.
If Cockpit is running, it updates in place: the server hands off to a
helper that installs while nothing is holding the install directory, then
relaunches with the same port and token. A failed install rolls back to the
previous version and still restarts, so an interrupted update never leaves you
without a working Cockpit.
If it is not running, cockpit update simply installs.
Once the new server is up, any browser tab left open from the old version shows
a “reload” prompt — its cached assets belong to the previous build.
Windows
Windows support is experimental — it is exercised by CI but has far less
real-world mileage than macOS and Linux. Cockpit runs natively; the Console
panel uses Git Bash when it is installed and falls back to PowerShell, the same
order Claude Code applies on Windows.
- Install Git for Windows for a bash
terminal. Without it the Console panel runs PowerShell, and the
cockpit.bash channel used by HTML apps is unavailable.
- The file browser, diff and chat panels behave as they do elsewhere.
Prefer a Linux toolchain? Run all of Cockpit inside WSL2, where it is an
ordinary Linux install:
wsl --install # once, from an admin PowerShell, then reboot
wsl # enter your Linux shell
npm install -g @surething/cockpit # install Node ≥ 22.19 inside WSL first
cockpit
Either way, open http://localhost:3457 in your Windows browser — WSL2
forwards localhost automatically.
Pick one side and stay there. Running Cockpit on Windows while your
project lives inside WSL (or the reverse) gives the agent one set of paths
and its terminal another: the same directory is C:\Users\me\proj on one
side and /mnt/c/Users/me/proj on the other, so paths passed between them
do not resolve. Keep Cockpit and your code in the same world.
The package ships a single command: cockpit. (The former cock alias has been removed — one name everywhere.)
From source
git clone https://github.com/Surething-io/cockpit.git
cd cockpit
npm install
npm run setup # build + npm link (registers `cockpit`)
CLI
cockpit browser <id> snapshot # capture accessibility tree
cockpit browser <id> click <uid> # click element
cockpit terminal list # list terminal bubbles
cockpit terminal <id> output # read terminal output (filter, cursor)
cockpit connection list --cwd . # list ALL bubbles (term + browser) with user-set titles
Read more
Development
npm run dev # dev server → http://localhost:3456
npm run build # production build
npm run setup # build + npm link
npm run lint # ESLint
Tech stack
Next.js 16 · React 19 · TypeScript · TailwindCSS · xterm.js · node-pty · Shiki · tree-sitter (WASM) · i18next · Claude Agent SDK · Vercel AI SDK
Contributing
Issues and PRs welcome. See CONTRIBUTING.md and the OpenCockpit docs.
Questions, ideas, or show-and-tell: join the community at r/opencockpit.
License
MIT © Surething
If Cockpit saved you 10 minutes today, a ⭐️ on GitHub is the cheapest thank-you we know.
评论