For teams already using AI
Connectors gave your AI access to documents. They did not give your organization a record of what it decided. A chat that reads your Confluence and drafts a PRD produces a document — not a decision.
pm.gold adds the missing half: the commit point.
Connect in one line
An admin issues a key from Agent access, then adds pm.gold to Claude Code, Claude, ChatGPT, Codex, or any MCP client. Clients that support OAuth can authorize with a click instead — the human approves on a consent screen stating exactly what the agent will and will not be able to do.
claude mcp add pmgold --transport http https://pm.gold/api/mcp --header "Authorization: Bearer pmg_…"Seven tools
find_decision_precedentReadsSearch past committed decisions on a topic, with how each one actually turned out.
read_approved_decisionReadsThe whole governed chain behind a decision: evidence cited, sections that honestly had none, precedent that informed it, versions, approver, whether what shipped matched intent, and the recorded outcome.
fact_check_artifactReadsGrade a document's claims against this workspace's own evidence — supported, contradicted, or unsupported, with the sources behind each verdict.
compare_intent_to_deliveryReadsCheck an approved PRD clause by clause against the implementation evidence in the workspace.
propose_prdProposesDraft a PRD grounded in the workspace's evidence and file it for human review. Sections with no evidence are flagged, never invented.
submit_for_approvalProposesSubmit text drafted elsewhere for review. It carries no citations and is marked agent-authored, so the reviewer sees that plainly.
capture_web_sourcesProposesHand pm.gold a URL your assistant found. It searches with its own tools; we fetch the page ourselves through a guarded crawler and file it as evidence for review — so research needs no search vendor of ours.
What stops an agent going wrong
A key is bound to the person who issued it, and principals are derived live from that person’s directory identity. An agent cannot retrieve a source its owner cannot see, and deprovisioning the human immediately narrows the agent. Same enforcement path as a browser session — only the credential differs.
There is no approve tool in the surface for an agent to call. This is structural rather than configured: the type describing what a tool may write has no value meaning “commit”, so a committing tool cannot be expressed, and the code that grants approval is not in scope for the MCP layer at all. An automated check asserts both on every push.
What an agent read, what it proposed, and who approved it are written to the same append-only trail as human actions — attributed to the human who authorized the key. Revoking a key stops the agent on its very next call and leaves the grant history intact. Keys also expire, and rotation issues a successor linked to the key it replaces, so the trail survives the swap instead of restarting at it. An expiry can only ever be shortened, never extended — enforced by the database rather than by policy, so “just give it another year” is not something an admin can do by accident.
Why a ledger, not another assistant
Generation is the easy part now. Your PMs already draft with Claude and ChatGPT, and the documents are good. But six months later nobody can say what evidence backed the call, who approved it, whether what shipped matched the intent, or how it turned out — because the reasoning lived in a chat window.
pm.gold does not compete with the AI your team already uses. It is where that work lands: agents propose, a named human commits, and the decision becomes queryable by every future agent and every future hire.
Six months, no charge, instrumented from day one — in exchange for honest feedback and permission to measure whether the record actually compounds.