Steerhouse

Steer your AI coding agents from your editor, made for developers who read the code.

Run as many Claude Code sessions as you like, in the terminal as usual. Steerhouse shows what each one is doing, which sub-agents it spawned, and every line it changed, with the real before/after. Comment on a line and the agent picks it up. Works in VS Code, Cursor, Windsurf and VSCodium, locally or over Remote-SSH (then it runs on the remote host, so your laptop does no work).

What you get

• New in 1.0: worktree per chat; PR files with viewed marks and Merge; edit an earlier message and run again (with file undo); Quote, Up-arrow recall, prompt stash and suggested next prompt; pin and archive projects; Open on Phone; Add MCP server; accounts view; agent picker (Claude, providers, Cursor Agent, Codex, Gemini, OpenCode, Grok).

• Steerhouse chat: Claude Code with a real UI instead of a terminal. You get streaming answers, the agent's plan as a live checklist, what it's doing right now ("Searching…", "Editing page.tsx…"), every edit as a diff, permission / plan / question cards with buttons, pasted or dropped screenshots, model and permission-mode pickers, slash commands, and Steer (default) / Queue / Send now while it works. A queued message can still be sent now or cancelled. Link session lets the agent see what another session is doing, and you can jump to that session.

• Prompts survive reloads: Reload Window or closing VS Code doesn't stop a running chat; it reconnects when you reopen it. Closing the chat's tab does stop it.

• Code completion: gray suggestions in the editor, Tab to accept, from your Claude subscription (about 2 to 3 s, counts toward your plan limits) or a model provider you added. Off by default: click Tab in the status bar. The code around your cursor goes to what you chose (a local provider like Ollama keeps it on the machine), and files like .env and private keys are never sent.

• Browser with no setup: chats in web projects (with a package.json) can drive a real browser. It starts on the first browser action, downloading Chromium if there is none. Open Live Browser shows it in a tab (view only). Git never sees its files: .pwview/ and, if ports had to change, pwview.local.toml are listed in .git/info/exclude. On systems that block Chromium's sandbox (Ubuntu 23.10+), it runs with --no-sandbox. Turn it off with steerhouse.pwview.enabled.

• Graph: sessions → prompts → sub-agents → files, drawn with xyflow. Prompts from different sessions that touched the same file are connected, and sessions you linked show a dashed edge.

• Follows your theme: colors, fonts and light / dark / high contrast come from your VS Code or Cursor theme.

• Sessions view: every session grouped by project (this workspace first), live state (working / waiting / idle), branch, edit count, and a sub-agent tree under each session.

• Real diffs: click any change to open VS Code's own diff editor with the file exactly as it was before and after that edit, plus the agent's reasoning for it.

• Comments that reach the agent: click the gutter in a diff (or any file) and choose Send to Agent. The comment is handed to that session when it finishes its turn, so it keeps working on it. Messages work the same way from the dashboard.

• Linked changes: files touched by more than one session are flagged, and sessions show which other sessions they overlap with, including across sub-agents and worktrees.

• Dashboard: an overview of everything that's working, plus a session view with activity timeline, changes by file with inline diffs, sub-agents and comments.

• Prompts: reusable prompts in .steerhouse/prompts/*.md (shared with the team) or in your user folder. mode: interactive opens a session, mode: headless runs claude -p once. Variables: ${file}, ${selection}, ${workspace}, ${input:Question}.

• Skills: browse and create Claude Code skills (.claude/skills/<name>/SKILL.md) for the project or for yourself.

• Start / resume: new session with a prompt, or claude --resume any session, in a VS Code terminal.

Models and accounts

Steerhouse doesn't call any model itself. It watches and launches your agent CLI, so whatever Claude Code is set up with works: a Claude subscription (/login), an API key, Amazon Bedrock, Google Vertex, or any compatible gateway through ANTHROPIC_BASE_URL. Pick a model per run with steerhouse.agentArgs (e.g. ["--model", "opus"]) or per prompt with model: in its front matter.

Install, step by step (about 2 minutes)

Using Remote-SSH? Do every step in the SSH window. Everything runs on the remote machine.

1. Install the extension. In VS Code or Cursor: Extensions view → ⋯ → Install from VSIX… → pick steerhouse-<version>.vsix. Over SSH, click Install in SSH: when asked.
2. Reload the window. Ctrl+Shift+P → Developer: Reload Window. The Get started with Steerhouse guide opens by itself. You can reopen it any time from ⋯ → Get Started in the Steerhouse view.
3. Check Claude Code. Click Check setup. If Claude Code is missing, Install Claude Code types the official installer into a terminal; press Enter.
4. Log in once. Click Log in, type /login in the terminal and follow the link. Every chat and terminal reuses this login, whether it's a subscription, an API key or Bedrock/Vertex. The Claude desktop app's login is separate, which is why a terminal can ask even when the app is logged in.
5. Enable message delivery. Click Enable. This adds four small Claude Code hooks to ~/.claude/settings.json (with a backup) so Queue / Steer / Send now also reach agents running in a plain terminal. Steerhouse chats don't need them.
6. Start a chat. Click New chat (or the 💬 button in the Sessions view).

Something not working? Run Steerhouse: Check Setup. It lists each requirement with a one-click fix.

Settings

Develop

npm install
npm run build         # extension + hook (esbuild) and the React/shadcn UI (vite)
npm test              # 22 tests: transcripts, hooks, chat engine, queue/steer/now, links, graph, skills, bundle load
STEERHOUSE_E2E=1 npm run test:e2e   # real chat through your Claude Code (Haiku, ~$0.03)
npm run dev:webview   # UI in a browser with mock data (#chat for the chat view)
npm run package       # steerhouse-<version>.vsix

Data comes only from files Claude Code already writes (~/.claude/projects). Comments are stored in ~/.claude/steerhouse/comments.json. Nothing leaves the machine.