Skip to content

Repository files navigation

Runfile SDKs

Open-source SDKs and customer-facing tooling for Runfile — the tamper-evident audit trail for AI agents. This repo is where engineers integrate Runfile into their agent code, and where audit firms verify that the Verifier CLI does what we claim.

License: Apache-2.0. The git history of this repo is public forever — keep backend IP, sales context, and security discussion out of it (those live in the private runfile/platform monorepo).

Packages

Package Registry Language Wire sdk.name
runfile-ai PyPI (pip install runfile-ai, import runfile_ai) Python 3.11+ runfile-ai
@runfile-ai/sdk npm TypeScript / Node 20+ @runfile-ai/sdk
runfile-verifier GitHub Releases Go 1.22+

All three depend on the published schema contract — @runfile-ai/schemas (npm), runfile-ai-schemas (PyPI), and github.com/runfile-ai/schemas (Go) — as a versioned package, never a path link. A schema change is an explicit dependency bump here.

What the SDK does

Observes the customer's agent (via framework adapters or a manual API) and translates framework-native signals into Runfile runs and events:

  1. Capture events (LangGraph, OpenAI Agents, Claude Agent SDK, MCP, manual).
  2. Manage run lifecycle from observable signals (start / suspend / resume / end).
  3. Classify + redact PII client-side (the PII boundary).
  4. Envelope-encrypt the redacted payload locally under a per-(tenant, agent) data key fetched from POST /v1/data-keys (no AWS credentials client-side).
  5. Batch and ship to the Ingest API (POST /v1/batches) off the hot path.

The full design lives in the private design docs (sdk-design.md, sdk-ingest-api.md, event-schema.md).

The Claude Agent SDK, LangGraph, and OpenAI Agents Python adapters are exercised end to end by a credit-line decision agent — the same human-in-the-loop agent built on multiple runtimes, capturing a tamper-evident audit trail from the PyPI-published runfile-ai (via the runfile-ai[anthropic] / runfile-ai[langgraph] / runfile-ai[openai-agents] extras, never a path dependency). Capture is opt-in: a transparent pass-through unless RUNFILE_API_KEY is set.

Layout

sdks/
├── packages/
│   ├── python/         # runfile-ai (PyPI)
│   ├── typescript/     # @runfile-ai/sdk (npm)
│   └── verifier-cli/   # runfile-verifier (Go, GitHub Releases)
├── examples/           # end-to-end demo apps
└── docs/               # SDK docs (docs.runfile.ai)

Status

🐍 Python (runfile-ai) is live on PyPI — 0.3.0. The core (manual + context-manager capture API, encryption, hash-chained flusher, PII redaction) and the Claude Agent SDK, LangGraph, and OpenAI Agents framework adapters are shipped and exercised end to end by the credit-line decision agent example. The MCP adapter, the TypeScript core, and the Verifier CLI logic are still in progress. See docs/STATUS.md.

Development

pnpm install            # TypeScript workspace + tooling
pnpm -r build
pnpm -r test

Python and Go each have their own local setup — see their package READMEs.

About

No description, website, or topics provided.

Resources

Contributing

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages