MCP: Archive docs

Over HTTP

/mcp serves the Model Context Protocol over streamable HTTP. Make a token for the assistant in Settings or with archive token create, and send it as a bearer token. The assistant then acts as you, within the token's spaces and roles, and its edits are attributed to the token.

claude mcp add --transport http archive https://archive.example.com/mcp \
  --header "Authorization: Bearer arc_…"

Over stdio

archive mcp serves MCP on standard input and output, acting as the first owner or the user in -as. Run it on the server, or inside the container:

claude mcp add archive -- docker exec -i archive archive mcp

Tools

Tool	Does
`describe`	Every type with its schema or fields, parent and description, and up to 3 sample records per user type
`sample`	A few records of one type
`query`	List objects with field filters; page with after; reveal for granted secrets
`get`	One object with its provenance and links
`upsert`	Create an object, or set some fields of one
`delete`	Delete an object
`propose`	Propose ontology ops
`preview`	Dry-run a proposed change
`apply`	Apply a proposed change
`revert`	Revert an applied change
`intake`	Read a folder or a stored file into a type
`files`	List files
`file`	One file's details
`file_text`	A file's extracted text, in pages of characters
`search`	Search records, files and raw items
`sources`	List sources
`sync`	Sync one source now

The server also sends instructions: describe first, propose only what the user asked for, preview before applying, and confirm before deleting. Validation errors come back as JSON with one entry per bad field.

Changing the ontology

An assistant can reshape your types live, through the same steps as every other surface: describe, propose, preview, then apply if you agree, and revert if you change your mind. Applying needs a token with the admin role.