drive-openhub
Use to connect to and drive a OpenHub instance from an external agent (Claude Code / Cursor / any MCP client). Delegate goals to OpenHub's autonomous agent, drive its browser / computer / knowledge base / files, and manage the platform — over MCP or REST, with a per-companion access token (you run AS the bound companion). Use this whenever the user asks to control, automate, or hand work off to "OpenHub", "their OpenHub", "the desktop companion", or a running OpenHub server.
- License 许可
- Apache-2.0
- 最近更新
- 2026-07-28
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- drive-openhub
- description
- Use to connect to and drive a OpenHub instance from an external agent (Claude Code / Cursor / any MCP client). Delegate goals to OpenHub's autonomous agent, drive its browser / computer / knowledge base / files, and manage the platform — over MCP or REST, with a per-companion access token (you run AS the bound companion). Use this whenever the user asks to control, automate, or hand work off to "OpenHub", "their OpenHub", "the desktop companion", or a running OpenHub server.
03 SKILL.md 正文
Drive OpenHub (external companion)
OpenHub exposes its full platform capability set — an autonomous agent, browser automation, computer control, knowledge bases, files, terminals, and platform management — to external callers through one MCP endpoint and an equivalent REST API, authenticated by a per-companion access token. Each token is bound to one specific companion: calling with it runs you as that companion, inheriting its profile model / persona / knowledge bases. Connecting makes you an "external companion": you drive the platform exactly as its built-in desktop companion does.
1. Connect (MCP, recommended)
Configure OpenHub as a Streamable-HTTP MCP server:
{
"mcpServers": {
"openhub": {
"type": "streamable-http",
"url": "http://<host>:25808/mcp-agent",
"headers": { "Authorization": "Bearer <token>" }
}
}
}
<host>is the machine running OpenHub (127.0.0.1locally, or its LAN/public address with WebUI remote access enabled)./mcp-agentadvertises a tight, curated tool set for getting work done (agent delegation, browser, computer, knowledge, files). Use/mcpinstead for the full platform-control surface (~140 tools incl. channels, companions, cron, providers, …).<token>is a per-companion access token — it binds you to one companion. Get it from the OpenHub operator: in the desktop WebUI/remote panel, or by minting one (see "Minting a token" below). The companion you bind to must have a configured model foropenhub_agent_run(and other model-backed caps) to work — otherwise minting returns awarningand those calls fail until a model is set in Model Management.
REST equivalent (for scripts): POST http://<host>:25808/v1/tools/<name> with the
same Bearer token; GET /v1/tools?profile=agent lists tools; GET /v1/openapi.json?profile=agent is a machine-readable contract.
Minting a token (operator, local-trust only)
The mint/query/revoke endpoints are local-trust gated (reachable only from
the desktop client / loopback, not from a remote browser). {id} is the
companion id you want the token bound to:
# Mint — returns the plaintext token ONCE, bound to that companion.
curl -X POST http://127.0.0.1:<port>/api/webui/companions/<id>/access-token
# => { "token": "<64-hex token>", "companion_id": "<id>" }
# If that companion has no usable model yet, the body also carries
# "warning": "…" (the token is still minted, but openhub_agent_run etc. will fail
# until you configure a model).
# Query whether one is configured (does NOT return the token):
curl http://127.0.0.1:<port>/api/webui/companions/<id>/access-token
# => { "configured": true }
# Revoke:
curl -X DELETE http://127.0.0.1:<port>/api/webui/companions/<id>/access-token
For a headless server, seed a token at startup via the
OPENHUB_COMPANION_TOKEN env var — it binds to the default companion (only
if no token is configured yet; it won't overwrite an existing one):
OPENHUB_COMPANION_TOKEN="$(openssl rand -hex 32)" openhub-core # or openhub-web
2. Delegate a goal (the headline move)
To hand a whole task to OpenHub's own autonomous agent, call openhub_agent_run:
{ "goal": "Research X and write a summary to notes.md", "timeout_secs": 600 }
It spins up a fresh autonomous OpenHub agent (full tools, using the bound
companion's profile model — which must be configured), runs the goal to
completion, and returns { "status": "completed", "text": "<final answer>", "conversation_id": <id> }.
For long tasks it returns { "status": "running", "conversation_id": <id> } —
poll openhub_agent_result with that conversation_id until completed.
3. Or drive capabilities directly
Use individual tools when you want fine control: openhub_browser_* (navigate /
observe / act), the computer tools, openhub_knowledge_* (search / read / write
knowledge bases), openhub_fs_* (read / write / browse files), openhub_create_terminal,
and the conversation tools. GET /v1/tools (or MCP tools/list) is the live,
authoritative catalog with JSON Schemas.
4. Confirmations & limits
- Destructive actions (deletes, etc.) return
{ "needs_confirmation": true, "restate": "..." }. Restate the exact action to the user, get agreement, then re-call the same tool with"confirm": true. - Sensitive actions (secrets, factory reset) are denied on this surface.
- Trust model: holding a per-companion access token grants full, RCE-equivalent control of that OpenHub instance (as the bound companion). Treat the token as a high-value secret; only connect to instances you are authorized to drive. Revoking a token affects only its companion; other companions' tokens are unaffected.
5. Failure handling
401→ missing / invalid / revoked access token.- REST
409(orneeds_confirmationin the body) → re-call withconfirm: true. - REST
422/ a{ "error": ... }body → the tool rejected the arguments; check the schema from/v1/toolsand retry. - Connection refused → OpenHub isn't running or the URL/port is wrong.
04 同仓库更多 Skills
example-skill
Fixture skill used by startup_materialize integration tests.
xiaohongshu-recruiter
用于准备在小红书上发布高质量的 AI 相关岗位招聘帖子。包含自动生成极客风格的招聘封面图和详情图,并提供确认门控的发布脚本。当用户需要发布招聘信息、寻找 Agent 设计师或其他 AI 领域人才时使用。
x-recruiter
用于在 X (x.com) 发布招聘帖子。包含文案规范、图片生成提示和自动化发布脚本。发布 AI 相关岗位或设计类岗位时优先使用。
weixin-file-send
Use when the user wants a local file or image sent back, such as "send me the file" or "发给我".
ui-ux-pro-max
Professional UI/UX design research skill with a bundled searchable database for product patterns, visual styles, color palettes, typography, landing-page structure, dashboard charts, UX practices, and stack guidance. Use when designing, building, reviewing, or improving interfaces.
story-roleplay
Parse and apply character cards and world info files in multiple formats (PNG, WebP, JSON), fully compatible with SillyTavern format. Supports automatic parsing, keyword triggering, and dynamic updates.