Briefcase ships as briefcase-ai — a Python SDK over a Rust core. Install the base package, then add extras only when a concrete need appears.
Python 3.9+ Precompiled wheels No Rust toolchain needed
pip install briefcase-aiuv add briefcase-aiThe base install carries the whole core loop — capture a decision, store it, replay it, verify it — with no extras. The exports below are grouped by the act of the journey they belong to.
Recording
capture · observe · setup · DecisionSnapshot · Snapshot · Input · Output · ModelParameters · ExecutionContext · HardwareMetadata
Plus the stock exporters under briefcase.exporters and enable_logging / get_logger. Record each classify_ticket call and emit it to the console, a file, or memory.
Store & query
init · init_with_config · is_initialized · BriefcaseConfig · SnapshotQuery
Start the native runtime and query stored decisions. Cost types live in briefcase.cost — also base, no extra.
Replay & verify
Wired through the runtime above
Persist a decision, then re-run it and confirm its record is intact. The replay engine itself is gated behind the replay extra below.
Extras gate the import surface of optional submodules. Most install nothing and simply mark intent — only otel, lakefs, bitemporal-iceberg, and mcp pull in third-party dependencies. Install only what your deployment needs.
| Group | Extra | What it adds | Recommended for |
|---|---|---|---|
| Core | storage | SqliteBackend, BufferedBackend | Persisting classify_ticket decisions to a SQLite file |
| Governance | replay | ReplayEngine | Re-running stored decisions to confirm a change reproduces them |
| Governance | drift | DriftCalculator, DriftMetrics | Measuring how consistent repeated decisions are |
| Governance | validate | PromptValidationEngine | Checking prompt references before a call runs |
| Governance | guardrails | GuardrailEnv framework | Allowing or denying an action before it executes |
| Governance | routing | AgentRouter, PolicyRegistry | Versioned, policy-based routing of decisions |
| Governance | compliance | ExaminerBundle | Building a tamper-evident, verifiable bundle for a decision |
| Governance | bitemporal | BitemporalRecord, in-memory store | Reconstructing any past state of recorded facts |
| Governance | bitemporal-iceberg | Iceberg-backed store (installs pyiceberg, pyarrow) | A scalable bitemporal store |
| Capture | sanitize | Sanitizer (redaction) | Stripping sensitive spans from inputs/outputs |
| Store & Query | external | ExternalDataTracker | Snapshotting external data a decision read |
| Store & Query | rag | VersionedEmbeddingPipeline | Versioning an embedding index for reproducible RAG |
| Store & Query | lakefs | lakeFS VersionedClient (installs lakefs) | Reading versioned files with their commit SHA |
| Store & Query | vcs | VCS client base protocol | Implementing a custom versioned data source |
| Operate | otel | OpenTelemetry helpers (installs opentelemetry) | Correlating decisions with existing traces |
| Operate | correlation | Multi-agent workflow tracing | Correlating decisions across agents in one workflow |
| Operate | events | BriefcaseEvent emitter | Emitting events on low confidence or drift |
| Operate | mcp | briefcase-mcp server (installs mcp) | Exposing SDK tools to MCP clients |
| All | all | Everything above | Evaluation or local development |
| Dev | dev | Test and lint tooling | Contributing to Briefcase |
Install the base package — it covers recording and inspection, which is all the Quickstart needs to begin.
pip install briefcase-aiAdd storage and replay for the Quickstart’s persist-and-replay steps.
pip install "briefcase-ai[replay,storage]"Add further extras later, only when a need appears — for example otel for tracing or guardrails to gate actions.
pip install "briefcase-ai[all]"Installation is step zero of the journey. Next, capture and inspect a decision, then learn the data model behind it.