MCP

Precedent is one MCP server. Add it once and your agent can search the corpus while it designs, then cite each record it used by its id.

Install

Set PRECEDENT_KEY to the key from your account, then pick your client.

Run in a terminal
claude mcp add --transport http precedent https://precedents.dev/api/mcp \
  --header "Authorization: Bearer $PRECEDENT_KEY"

Tool reference

ToolReturnsExample call
precedent_searchSystems, decisions and failures that match a described problem, grouped by how their scale compares with yours: at your scale, at a different scale, or of unknown scale.precedent_search({ query: "job queue on postgres", workload: ["job-queue"], traffic: 1, team: 0 })
precedent_getUp to eight records or systems in full, each fact with its place in the source and an explicit note where the source is silent.precedent_get({ ids: ["notion-2021-postgres-sharding#d1", "healthchecks-2025-tarpit-outage"] })
precedent_facetsWhat the corpus covers: its size and version, the tier definitions, and the workloads, topics, tech tags and footprints in use.precedent_facets()

Worked example

Brief

Design the background job system for our Rails app. Under 100 requests a second. We are a team of 5.

Calls
precedent_search({ query: "background jobs rails", workload: ["job-queue"], traffic: 1, team: 0 })
precedent_get({ ids: ["tosbourn-2023-client-rails-jobs#d1", "37signals-2023-solid-queue#d1", "brandur-2015-postgres-job-queue#f1"] })
Cited in the design

Keys and limits

Every request needs a key, sent as a bearer token. A free key allows 1,000 tool calls a month; Pro will allow 10,000. Listing the tools and connecting do not count. Pricing →