Precedentdocs

Agents over MCP

Any MCP client can search the record, read a decision with its rationale, and propose drafts. It sees exactly what the signed-in person sees.

Connect a client

  1. Sign in and copy the MCP endpoint from Settings: it is your deployment URL plus /mcp.
  2. Add it as a remote MCP server in Claude, ChatGPT, Cursor, or anything that speaks MCP over streamable HTTP.
  3. The client registers itself through OAuth 2.1 with PKCE, and a consent screen in the app asks you to approve it with read or read-and-draft scope.
  4. Revoke a client any time under Settings, Agents. Every tool call it made is in the audit log, and in the signed audit export.
An agent acts as the person who approved it. Restricted nodes, workspace roles, and visibility rules apply to every tool call, so an agent can never read a decision its person cannot.

The tools

Read tools: search_decisions (hybrid keyword and semantic), get_decision, get_node_context, trace_impact, list_reviews_due, list_nodes, list_templates.

Write tools: propose_decision, update_draft, link_decisions, propose_supersession, add_evidence, mark_condition_met. Every write creates or amends a draft that waits in a person’s inbox. Agents propose; people confirm in the web app. There is no tool that confirms, supersedes, or archives, on purpose.

The decision-librarian skill

Tools alone do not tell an agent when to use them. The decision-librarian skill does: search before acting on a consequential choice, cite conflicts by ref instead of burying them, load a node’s context before planning work in it, draft only after the person agrees. Your deployment serves it at:

Download the skill

Or install it straight into a Claude Code project:

mkdir -p .claude/skills/decision-librarian
curl -s https://<your-deployment>/skill/decision-librarian \
  -o .claude/skills/decision-librarian/SKILL.md

For agents without a skill system, paste the same file into the system prompt. The MCP server also carries the same rules as its instructions, which many clients surface to the model automatically.

What this gives you

An agent that checks precedent behaves differently: asked to “move search to Elasticsearch”, it finds the decision that chose the current engine, quotes why, and asks whether to supersede it rather than silently planning the switch. The skill makes this reliable for stated choices. To also check the work an agent produces, not just the choices it is told about, see Validating agent work.