AI-Assisted Setup

Recommended Cursor · Claude Code · Copilot · Codex For: engineers

The fastest way to instrument decisions with Briefcase is to let the AI assistant you already code with do it for you. You don’t have to memorize the API — give your assistant the docs (or Briefcase’s own tools) and ask it to add capture, controls, and replay to a function.

Two ways, use either or both

Point your assistant at the docs so it answers from accurate, current Briefcase material. Add the MCP server so it can run Briefcase tools — redact PII, estimate cost, check drift — right inside your editor. They compose: docs for knowledge, MCP for actions.

Point your assistant at the docs

Every page on this site is also published as machine-readable text an assistant can ingest:

• https://briefcaseai.io/llms.txt — an index of the docs (titles + links), grouped by the five acts. Best as a map.
• https://briefcaseai.io/llms-full.txt — every page concatenated as plain text. Best for “read this, then help me.”

Both files are generated from the docs on every build, so they always match what’s published here.

1. Give your assistant the docs

   In Cursor, Claude Code, Copilot, or any assistant that can read a URL, paste the link (or add it to the project’s context / docs settings):

   Read https://briefcaseai.io/llms-full.txt — that's the Briefcase AI SDK.

2. Ask it to instrument a function

   Using Briefcase, add @capture and observe() to classify_ticket() in
   app/triage.py so every classification is recorded, then show me how to read
   the records back.

3. Review what it wrote

   The assistant should produce the canonical pattern — briefcase.observe(...) plus @briefcase.capture(...). Confirm it matches the Quickstart.

Note

llms.txt is an emerging convention for exposing docs to AI tools. Even assistants without a built-in “docs” setting can usually read a pasted URL.

Give your assistant Briefcase's tools (MCP)

The Briefcase MCP server exposes a small set of read-only tools your assistant can call directly: redact PII, estimate model cost, analyze drift, and look up usage. It also serves a briefcase://llms-full.txt resource so the assistant can read the guide in-editor.

1. Install the server

   pip install "briefcase-ai[mcp]"

   The [mcp] extra installs mcp>=1.2. The server runs over stdio with the briefcase-mcp command (or python -m briefcase.mcp).

2. Register it with your assistant

   Point your MCP-capable client at the briefcase-mcp command. The config file differs per tool, but the shape is the same:

   {
     "mcpServers": {
       "briefcase": {
         "command": "briefcase-mcp"
       }
     }
   }

3. Use the tools from your editor

   Ask the assistant things like:

   Use the briefcase tools to redact PII from this support ticket, then estimate
   the cost of classifying it with claude-haiku-4-5.

   The assistant calls sanitize_text and estimate_cost — the latter now supports a rate_card (e.g. "bedrock:batch") and returns a cache_cost, added in v3.2.1.

The server exposes sanitize_text, estimate_cost, analyze_drift, and how_to. See the MCP Server reference for each tool’s inputs and outputs.

Prefer to do it by hand?

The Quickstart walks the same path manually — record, persist, and replay a decision — in about 5 minutes.

Where this fits

Quickstart The manual path: record, persist, and replay a decision.

MCP Server Full reference for the tools and resource your assistant can call.

Why Briefcase What Briefcase governs, and the five-act lifecycle.

What's New in 3.3.0 oci-bai artifact graph launch, cost rate cards, and latest model pricing.