Simple work goes through the direct job tools — submit, keep the job id, poll status and result. Anything that chains steps, like analyse a video then rewrite it then import the result and return a link, goes to the agent run instead, where Docsie opens a hidden chat session and orchestrates internally while you track a single run id. Using the job tools for a chained workflow means writing the orchestration yourself; using the agent for a one-shot job means giving up the deterministic path.
The hosted MCP surface for Docsie's documentation platform. It covers two things at once: reading approved knowledge — help centres, SOPs, release notes, support articles — and running Docsie's video-to-documentation pipeline, from analysis through generation, export and publishing. Which tools appear depends on the organisation, workspace and permission packs granted.
- Workspace content searched and one article's markdown fetched, plus the workspaces available to your connection — `docsie_search`, `docsie_get_article`, `list_workspaces`
- A video turned into documentation as a job you submit and poll: start from a video URL or a Docsie file id, check status, fetch the result — `video_to_docs_submit`, `video_to_docs_status`, `video_to_docs_result`
- Documentation generated or regenerated from a finished analysis, and exported to DOCX, PDF or PPTX — `video_to_docs_generate`, `video_to_docs_export`, `video_to_docs_export_generated`
- Jobs listed, cancelled, and priced before you commit to them — `video_to_docs_list`, `video_to_docs_cancel`, `video_to_docs_estimate`
- Two to four videos compared, or two to ten combined, as a native job — `video_compare_submit`
- Narration: the available voice providers discovered and audio generated — `voice_options`, `voice_speech`
- A Docsie Chat-style agent run for workflows that chain several steps — analyse, rewrite, fill a DOCX template, run a policy check, import and publish — tracked by a run id — `docsie_agent_start`, `docsie_agent_status`, `docsie_agent_result`
- AI credits topped up through a returned Stripe Checkout link — `ai_credits_buy`
A Docsie account. The hosted endpoint is `https://app.docsie.io/mcp` over Streamable HTTP, authenticated with OAuth2 Authorization Code plus PKCE; you then choose an organisation and workspace and grant permission packs before any tools are exposed. The packs are the real gate — `workspace_read`, `workspace_write`, `video_docs`, `video_presentation`, `imports`, `publishing`, `billing` and `agent_runs` — and granting the smallest set that covers your workflow is the documented advice. Clients that cannot speak remote HTTP can bridge through `npx -y mcp-remote https://app.docsie.io/mcp`, or run the local wrappers `npx -y docsie-mcp` / `uvx docsie-mcp` with a service token in `DOCSIE_MCP_TOKEN`. Service tokens are scoped to one organisation, one workspace and explicit packs, which is the path for a shared agent that should not require every downstream user to log in.
One command — npx -y docsie-mcp