Skip to main content
The server at https://mcp.sidecars.ai/mcp exposes 14 tools and one resource template, and no prompts. Each entry below links to its full contract: parameters, returned fields, errors and an example. Read tools change nothing. Write tools act as you in the workspace and are checkpointed and audited. Destructive write tools can replace existing content and require a version check.

Knowledge

Scope knowledge:read, granted on first connect.
  • search_knowledge (read): hybrid vector and full-text search over the workspace knowledge base, returning ranked excerpts.
  • ask_knowledge (read): a synthesized answer to a question, with cited sources.

Conversations

Scope conversations:read, by step-up consent on first use.
  • list_conversations (read): the rooms and pods you can reach, by pod name, slug or id and by room title — how a name becomes the conversation_id / pod_id the document tools take.

Documents

Reads need documents:read, writes need documents:write. Both arrive by step-up consent on first use.
  • list_documents (read): the documents and pages you can reach, filtered by conversation, pod or library section, with your role on each.
  • read_document (read): a document as markdown, windowed for long content, always with its version.
  • create_document (write): a new document in a conversation or in your library, optionally pre-filled, with idempotent retries.
  • edit_document (write, destructive): append, replace a quoted section, replace everything, or fill an empty document while people may be editing it live.
  • get_document_versions (read): the recent checkpoints and named versions, newest first.

Pages

Same scopes as documents.
  • create_page (write): a new HTML page artifact in a conversation or your library.
  • read_page (read): a page’s stored HTML and its version.
  • write_page (write, destructive): replace a page’s HTML under a required expected_version; a conflict returns a diff.

Export

Scope documents:read.
  • export_document (read): render a document, page or table to markdown, PDF, HTML, CSV or JSON and get a 15-minute download URL.

Comments

Reads need documents:read, writes need documents:write.
  • read_comments (read): the comment threads on a document or page, with anchors, replies and your role.
  • add_comment (write): start a thread, optionally anchored to a quoted passage, or reply in one.
  • resolve_comment (write, idempotent): resolve or reopen a thread.

Resources

  • sidecars://document/{id}: the documents you can reach, as text/markdown resources for hosts that mount them next to tools.

Which methods need a token

A presented-but-invalid token is refused on every method.

How to read a tool result

A successful call returns a text block for a person and a structuredContent object for the host. The two always agree.
A recoverable failure comes back the same way with isError: true and, when there is something to act on, an error code in structuredContent. See Troubleshooting. Every tool declares MCP annotations so a host can decide whether to ask before calling it:

Not exposed

These are not available through this server today, and no tool pretends otherwise:
  • Posting messages or reading a room’s messages (conversations:write is reserved for a future release; conversations:read lists rooms only).
  • Sharing, granting, moving or linking artifacts.
  • Restoring a version, naming a version, or deleting anything.
  • Editing tables (a table is filled in by the agent that created it).
  • Resource subscriptions (notifications/resources/updated).
  • Prompts (prompts/list is answered as unsupported).