Briefcase AI SDK

Governance infrastructure for AI decisions: enforce controls, record context, replay later.

Install

pip install briefcase-ai

Quick Example

import briefcase

briefcase.observe("console")          # send records to stderr (or "memory" / "runs.jsonl")

@briefcase.capture(decision_type="classify_text")
def classify(text: str) -> str:
    return text.upper()

classify("hello world")               # a decision record is emitted

capture works immediately — no briefcase.init() required. Call init() only when you use the native runtime features (storage backends, snapshots). Without a call to briefcase.observe(...) (or briefcase.setup(exporter=...)), @capture records decisions but has nowhere to send them.

Exporters. briefcase.observe(...) accepts "console", "memory" (records collected on exporter.records), a "*.jsonl" path, or any briefcase.exporters.BaseExporter instance — subclass BaseExporter to ship records to your own backend.

Evaluations and RL

Two bridges emit decision records from work that is not a single function call. Both use the exporter you already configured.

from briefcase.integrations.evals import EvalRun, from_inspect_log, replay

with EvalRun("gsm8k", model="claude-opus-5") as run:   # one eval.case per case,
    run.log_case("q1", inputs=q, outputs=a, passed=a == target)   # one eval.run
print(run.summary()["pass_rate"])

replay(from_inspect_log("logs/2026-08-12_gsm8k.eval"))  # inspect-ai .json/.eval

The parsers are stdlib only and never import the eval framework, so a log can be replayed on a machine that has neither installed. See examples/eval_runs/.

from briefcase.integrations.gym import GuardrailGymEnv, capture_episodes

env = GuardrailGymEnv(guardrail, tasks, injections)  # a guardrail as a gym.Env
env = capture_episodes(env)                          # rl.step / rl.episode records

Needs pip install briefcase-ai[gym]. See examples/rl_gym/.

capture_episodes exports on a background thread by default, since step capture is on the hot path. A script that exits right after close() can lose its records; pass capture_episodes(env, async_capture=False) in short runs. EvalRun already defaults to async_capture=False for that reason.

Logging

The library is silent by default (it installs only a NullHandler). Turn on visible logs explicitly:

import briefcase
briefcase.enable_logging("DEBUG")     # or set BRIEFCASE_LOG_LEVEL=DEBUG

Using with AI tools (Cursor, Claude Code, Codex, …)

This repo ships machine-readable usage guidance: llms.txt / llms-full.txt, an AGENTS.md, and copy-paste editor rules under docs/llm/. An MCP server is available via pip install briefcase-ai[mcp] then briefcase-mcp.

Extras

Most features are native- or pure-Python-backed and ship with the base package — their extras (replay, drift, sanitize, storage, routing, bitemporal, compliance, …) are convenience groupings that pull in no additional dependencies. The extras that install third-party packages are otel, lakefs, bitemporal-iceberg, bitemporal-glue, kdb, compliance-kms, gym, evals, mcp, opa, kafka, gcp-logging, integrity, the rag-* and vcs-* store adapters, and the framework auto-instrumentation extras.

Managed platform

Every library feature runs against infrastructure you own: storage backends (in-memory, SQLite, pyiceberg, Glue-authenticated Iceberg, kdb+), KMS-signed examiner bundles, guardrails, RBAC/ABAC/OPA policy evaluation, routing, replay, framework auto-instrumentation (briefcase.auto), lakeFS branching and lineage, vector-store and VCS adapters, event transports, and the controls layer, in both Python and TypeScript (@briefcase-ai/controls).

The commercial offering is the hosted platform, not gated code: a managed control plane, catalog provisioning and credential brokering, operational runbooks (DR failover, catalog migration, key rotation), certified retention and regulator attestations, licensed market-data ingest (Bloomberg BPIPE, Refinitiv, ICE), and SOC 2 / FedRAMP posture with SLA support.

Contact sales@briefcasebrain.com for the managed platform.

Telemetry

The SDK can report anonymous usage metrics (SDK version, OS, architecture, backend type) to help prioritize development. No personal data or decision content is ever collected. Telemetry is only transmitted when a collection endpoint is configured via BRIEFCASE_API_URL (it defaults to localhost, i.e. a no-op), and you can disable it entirely at any time:

export BRIEFCASE_TELEMETRY=0   # also accepts: false, no, off (case-insensitive)

Links

• Docs: briefcaseai.io
• GitHub: github.com/briefcasebrain/briefcase-ai-sdk
• Contributing: CONTRIBUTING.md