feat(opencode): add MCP server + Agent Skill for opencode CLI - #29
namana843-bit wants to merge 1 commit into
Conversation
- mcp-server: Node MCP (stdio) exposing computer_screenshot/click/drag/type/key/scroll/move, bash_run (blocked rm -rf /, mkfs etc), browser_snapshot/navigate/tabs/click/type via CDP http://127.0.0.1:9222 - mirrors src-tauri/src/computer.rs, browser.rs, bash.rs, agent.rs; AI space 1280x800 - skills/computer-agent/SKILL.md: opencode-compatible skill (name: computer-agent) installed via npx skills add suitedaces/computer-agent -g -a opencode - opencode.json.example: mcp config for opencode - README: new Use with opencode CLI section - .gitignore: allow mcp-server/dist - verified: npx skills add ./skills -g -a opencode works, MCP builds and runs via node mcp-server/dist/index.js
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 9c6b4e426f
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| const img = await screenshot({ format: "png" }); | ||
| // img is Buffer, convert to base64 | ||
| const b64 = (img as Buffer).toString("base64"); |
There was a problem hiding this comment.
When computer_screenshot runs on any display that is not exactly 1280x800, this returns the native-size screenshot while the mouse tools still interpret inputs as 1280x800 AI coordinates. A coordinate read from the returned image is then scaled a second time by mapFromAI (for example, x=960 from a 1920-wide screenshot maps to x=1440), so clicks and drags land in the wrong place for the primary workflow of screenshot-then-act. The MCP path needs to resize like the Rust ComputerControl::take_screenshot implementation or expose/use native coordinates consistently.
Useful? React with 👍 / 👎.
| "optionalDependencies": { | ||
| "@nut/nut-js": "^4.0.0" | ||
| }, |
There was a problem hiding this comment.
Adding @nut/nut-js here without resolving it in mcp-server/package-lock.json makes clean installs fail before anything is installed: I checked npm --prefix mcp-server ci --omit=optional --ignore-scripts, and npm reports Missing: @nut/nut-js@ from lock file. Any CI or reproducible setup that uses npm ci cannot build the new MCP server until the lockfile is regenerated from this package.json.
Useful? React with 👍 / 👎.
| // Use CDP Page.navigate via websocket would be ideal; use http fallback via fetch to chrome's /json/new?url | ||
| if (url) { | ||
| // open new tab via /json/new | ||
| const r = await fetch(`${CDP_HOST}/json/new?${encodeURIComponent(url)}`); |
There was a problem hiding this comment.
When browser_navigate is called with a URL, this fetch defaults to GET, but Chrome's remote-debugging /json/new endpoint only accepts PUT for creating a tab. In a normal Chrome --remote-debugging-port=9222 session this returns 405, so the advertised URL navigation path never opens the page; pass method: "PUT" here or drive Page.navigate over the websocket.
Useful? React with 👍 / 👎.
1 participant
Summary
Makes computer-agent (taskhomie) runnable headless from opencode CLI via MCP server + Agent Skill - no Tauri window needed. Reuses the same Rust tool surface (computer.rs, �rowser.rs, �ash.rs, �gent.rs) over MCP stdio.
What's added
1. MCP server mcp-server/ (Node, @modelcontextprotocol/sdk)
ight_click/double_click (AI coords mapped to real screen)
m -rf /, mkfs, dd if= etc, truncated 8000 chars)
pm install && npm run build -> mcp-server/dist/index.js
2. Agent Skill skills/computer-agent/SKILL.md
ame: computer-agent compatibility: opencode (https://opencode.ai/docs/skills)
px skills add suitedaces/computer-agent -g -a opencode -y -> ~/.config/opencode/skills + ~/.agents/skills (opencode discovery)
px skills ls -g -a opencode shows computer-agent
3. Opencode config
json { "mcp": { "computer-agent": { "type": "local", "command": ["node", "mcp-server/dist/index.js"] } } }4. Docs
Verification
`�ash
skill
npx skills add ./skills/computer-agent -g -a opencode -y
npx skills ls -g -a opencode --json | grep computer-agent
MCP
cd mcp-server && npm install && npm run build
node dist/index.js # stdio -> "computer-agent MCP server running"
in opencode, add mcp to opencode.json and prompt:
"use computer-agent to take a screenshot and click at 600,400"
`
Tauri desktop flow unchanged (
pm run tauri dev still works).
Testing
pm --prefix mcp-server run build passes (TS strict, @nut/nut-js optional)
Closes headless/opencode gap - lets any opencode agent (local or remote) drive the desktop without the Tauri UI.