An If-Match precondition stands between a coding agent and a teammate's lost edit

The mechanism is two fields. A read returns the block's stable ID and a fingerprint for the document version it observed, and the tutorial's read-only prompt asks the model to report both ^[16][12]. The write sends that fingerprint back as an If-Match style precondition, and the server applies the change only while the selected representation still matches the condition the caller supplied ^[16][17]. If someone saved in between, the write does not land, and the instruction to the model is to stop and show the conflict ^[15]. Without the precondition the same sequence completes quietly: the agent reads version A, a teammate creates version B, and the agent writes an edit computed from A over the top of B ^[18].

The addressing requirement is what makes the read-only phase checkable at all. A paragraph should keep the same address when it moves ^[5], so you can ask for title, heading path, stable block ID and the exact evidence relied on, then open the document and look ^[12]. A confident answer with no address is one you get to re-derive by hand.

Count the requirements the author sets out for an editing contract and the weighting is clear. Four of the seven are about identity, version or write granularity: same address for a moved paragraph, a read that names its version, a stale write that fails visibly, and a small change that does not replace the whole document ^[5][19]. One is about search returning evidence rather than only a generated answer ^[5]. The author's finding is that a reliable boundary matters more than added context ^[20], and the list bears that out.

Doco is the author's own product, and the piece says to read it as a first-party tutorial rather than an independent comparison ^[2]. Two things would have to be true for the recipe to transfer to whatever store you already run. Your reads have to answer which block and which version. Your write path has to reject on mismatch rather than merge or last-write-wins. That is the lost-update protection the tutorial points at in RFC 9110 ^[17], a property of the protocol rather than any one vendor.

The connection details are ordinary and worth getting right anyway. Anthropic recommends HTTP for remote services and stdio for local processes, and the tutorial gives one command for each ^[6][7][8]. Keep the token out of committed configuration: Claude Code expands environment variables in project-scoped `.mcp.json`, so a shared config can reference a local secret without it landing in Git ^[9]. Run `claude mcp list`, then `/mcp` inside the session, and inspect the tools before asking the model to call them ^[10]. The Claude Code documentation warns that servers fetching external content can expose clients to prompt injection, and that knowledge-base text is data rather than trusted instruction ^[11]. That warning is the reason the write token is created with the minimum scope needed for one test on one disposable document ^[14].