Docs

The loop

• CaptureYour agent writes down what it worked out — decisions, constraints, dead ends — as the work happens, not afterwards.
• LoadA later session opens the same context and starts from that understanding instead of rebuilding it.
• ForkWhen the work branches, fork it. The copy is independent, and it records where it came from.
• ContinueCheckpoints commit what changed, so the next session can see how the work moved and pick it up.

Connect an agent

Create an account, then generate a token from the Connect screen and drop it into your client's MCP configuration. The endpoint is https://api.contextfork.dev/mcp, over streamable HTTP.

claude mcp add contextfork \
  --url https://api.contextfork.dev/mcp \
  --header "Authorization: Bearer cf_sk_…"

One token per agent, so you can revoke a machine without touching the others. A token acts as you: an agent holding one can read, write, fork and delete your contexts, and can make a private context public.

What the tools do

A context is one coherent line of work. Its content lives in nodes — each a piece of markdown under a name you choose, with a declared kind so a decision is distinguishable from an open question. Committing a checkpoint freezes what changed, and the history survives later edits.

Your agent does not need this page to use them. The server sends its own instructions when a client connects, and every tool carries a description explaining when to reach for it — so the authoritative list is always the one your client just fetched, never a copy here that could fall behind.

What stays private

Contexts are private by default; publishing one is a deliberate act. Your email address, your private contexts, the contexts you star and the accounts you follow are never public. Published contexts can be read and forked by anyone — and once something has been public, treat it as public permanently.

The full detail is in the privacy policy.

Support

Questions, problems, or anything that looks wrong: hello@contextfork.com.