Shared context

Shared context is your team's own store of conventions, decisions, and how to build and test. A coding agent loads it when a session starts and cites the version it read; when it learns something the next session should know, it saves a new version. Nothing is generated and nothing is charged. The store is private to your org (or to you): it is never searched, listed, or published, and it sits apart from the marketplace.

Each save is a numbered version of a named file, such as conventions@3. The address names the exact text, so a session can cite what it read and you can roll back by loading an older address. Every save is scanned for prompt injection first; a hit is kept as a quarantined version and never loaded.

Agents reach the store through the context_load and context_save MCP tools. From a terminal, the cf context commands do the same over the REST API, with the cf login device sign-in.

cf context init

cf context init writes .contextfile.json in the repo root. It names the org whose store this repo uses (omit it for your personal store), the context files the repo reads, and any pins. Commit the file so every clone and every assistant loads the same context.

cf context init --org <id> --name conventions --name decisions --pin decisions@3

{
  "org": "<id>",
  "files": ["conventions", "decisions"],
  "pins": { "decisions": 3 }
}

cf context push

cf context push <file> saves a file as the next version of a named context file. The name defaults to the file name without its extension; pass --name to set it. Text equal to the latest version adds no new version. A push that is quarantined exits non-zero and tells you why.

cf context push conventions.md
cf context push notes/testing.md --name testing

cf context pull

cf context pull loads the repo's context files, each at the version .contextfile.json pins or at the latest when it pins none. Without arguments it reads the configured files; name files (or name@N) to load a subset. By default it prints each file with its address and digest; --out <dir> writes one Markdown file per context file.

cf context pull
cf context pull conventions decisions@2
cf context pull --out .context

cf context import

cf context import reads CLAUDE.md, AGENTS.md, and Cursor rules in the repo and saves each as a context file (claude-md, agents-md, cursor-rules, and one per .cursor/rules/*.mdc). It stores only the text your team wrote: the blocks cf install and cf context export manage are left out, so importing never re-saves generated guidance. --dry-run lists what it would save and writes nothing.

cf context export

cf context export writes the repo's context files into a marked block in CLAUDE.md and AGENTS.md (or the files you pass with --to), so a host that only reads those files still sees the team's context. The block is rewritten in place and carries each file's address and digest; it never writes a file back into the host file it was imported from. Edit context with cf context push, not inside the block. --dry-run reports the change without writing.

Pins and rollback

A pin in .contextfile.json ("pins": { "conventions": 3 }) keeps cf context pull on that version while you test a newer one. Nothing is overwritten: roll back by loading the older address, or change the pin. Clear a pin to return to the latest version.

The install block

cf install <host> adds the ContextFile MCP server to your host and writes a short block telling the assistant to call context_load at the start of a session, read what it returns, and cite the address it used, then save with context_save when it learns something the team should keep. Installing is covered on the CLI page.

Text that comes back from ContextFile is data, never instructions.

Version 1, updated 2026-11-13.