Skip to content

WEBMCP.md

Latest commit

 

History

History
64 lines (50 loc) · 3.43 KB

WEBMCP.md

File metadata and controls

64 lines (50 loc) · 3.43 KB
Raw

WebMCP surface

Proofpress exposes a same-origin owner page as both a human governance UI and a WebMCP tool host.

Use the URL for your own Proofpress deployment when exercising this surface.

What agents may do

Registered on the signed-in owner page via document.modelContext.registerTool (fallback: navigator.modelContext):

  • get_workspace_summary() — orient the agent to queue, state counts, and current knowledge
  • list_review_queue(state?, scope?, limit?) — enumerate the bounded work requiring attention
  • get_current_context(scope, task?) — read eligible governed claims
  • get_review_state(claim_id) — inspect checks, policy/LM recommendation, human-decision state
  • get_lineage(claim_id) — evidence, history, whether the ledger currently exposes the claim
  • prepare_review_response(claim_id, response) — prepare a bounded revision handoff for the connected agent MCP/CLI
  • run_deterministic_checks(claim_id) — execute non-authorizing integrity and prerequisite checks
  • open_review(claim_id, full?) — route the owner to the right decision surface
  • get_activity(limit?) — inspect semantic proposal, review, policy, and context-retrieval activity
  • get_review_policy() — read the active safe policy projection and provider configuration status
  • prepare_review_policy_change(...) — load a complete policy draft into Admin for explicit human review
  • get_agent_access() — inspect agent identities and credential lifecycle metadata without secrets
  • prepare_agent_credential_issue(principal_id, label) — fill an owner-reviewed credential request without issuing or exposing a key

What agents may not do

approve / admit is not registered. Human Approval stays on the owner decision bar.

How to exercise it

  1. Open your deployment URL in Chrome with WebMCP enabled, or ChatGPT's in-app browser.
  2. Sign in with the owner credential from the submission form.
  3. Ask the agent to call get_workspace_summary, then list_review_queue.
  4. Ask it to inspect one candidate with get_review_state and get_lineage.
  5. Let it run run_deterministic_checks, then open_review for the owner.
  6. Human: Request changes. Agent: prepare_review_response, then submits the revision through its agent MCP/CLI. Human: Approve.
  7. Call get_current_context again and confirm only the admitted claim is returned.

Agents that need to propose still use the hosted /v1/operations credential path (evidence.submit, claim.propose). That path also cannot admit.

Agent-native product rule

Every user-visible Proofpress capability must have an agent-addressable MCP or CLI path. Agent-addressable does not mean agent-authorized. Reads and bounded candidate writes may execute under an agent principal; policy, credential, retention, recovery, and admission mutations become reviewable change requests.

prepare_review_policy_change writes no server state and receives no provider secret. It opens Admin with a clearly labeled draft. The signed-in human owner must inspect it and select Save & activate; Human Approval remains available only in the owner UI.