Most product code is now written with an AI agent in the loop, and that agent makes product assumptions constantly — usually leaving no trace. Sightspool exposes your workspace over the Model Context Protocol so the agent can ask what is already known before it guesses, and record what it assumed when it had to.
Connect it
On the Connections page, under Coding-agent access, issue an access key. It is shown once — it is a secret, stored only as a hash, and we cannot show it to you again. Rotate replaces it; revoke deletes it.
claude mcp add --transport http sightspool https://www.sightspool.com/api/mcp \
--header "Authorization: Bearer YOUR_SIGHTSPOOL_MCP_KEY"Any MCP client that speaks streamable HTTP works the same way; the Connections card also gives you the equivalent JSON configuration block.
What the agent can do
- check — what does the register already say about the code I'm touching? Entries are ranked by the paths they declare, then by relevance, and their state is reported honestly: still gathering is never dressed up as a verdict.
- triage — the workspace's proof-gated findings, ranked, with their citations.
- signals — first-party signals captured by the in-product SDK, filtered and paginated.
- diagnose — the provenance trail behind an entry: verdict history and what it depends on.
- declare — deposit a product assumption the agent just made, as a draft for you to review.
- flag — mark a committed entry that a change may have invalidated, sending it back to untested.
What it cannot do
Nothing reachable over MCP can act on your users or assert a conclusion. The four read tools have no write path at all. The two writing tools can do exactly two things — create a draft, or move a settled entry back to untested — which is the same de-escalation a commit already triggers. No tool can rule a verdict, commit an entry, or launch anything at a customer. This is a structural property of how the server is built and is covered by tests that fail if a world-mutating tool is ever added.