Changelog

What we shipped

Releases, launches, and milestones for the Zansn assurance toolkit. The open-source packages are the developer entry point to independent agent assurance.

  1. Release

    Framework adapter: Model Context Protocol (MCP)

    A fourth drop-in adapter that turns an MCP client session into a portable Zansn evidence pack, filling MCP's missing standard audit-logging layer.

    • @zansn/mcp: createZansnCollector() records tool-call requests, results, errors, and duration as they happen; finalize() returns an evidence pack.
    • Works with any MCP client (not tied to a specific SDK): recordToolCall() for request/result pairs, or beginToolCall() when they're handled separately.
    • Reuses the local zansn assurance engine (policy, privacy, cost, latency, readiness) with no new logic, and ships as a peer-dependency-only package.
    @zansn/mcp
  2. Milestone

    Signed, independently verifiable evidence standard

    Every Zansn evidence pack can now be signed into a tamper-evident record that any third party can verify offline, no Zansn account or network call required.

    • zansn 0.2.0 adds generateSigningKeyPair(), signEvidencePack(), and verifyEvidencePack(): Ed25519 over a canonicalised, SHA-256-hashed pack (zansn.signed-evidence-pack.v1).
    • Ships a standalone zansn-verify CLI: npx zansn-verify pack.json re-canonicalises, re-hashes, and checks the signature entirely offline.
    • The public key is embedded in the signed record, so it is self-verifying: an auditor, insurer, or regulator can check integrity without depending on Zansn Labs being reachable.
    • All four framework adapters (@zansn/ai-sdk, @zansn/langchain, @zansn/openai-agents, @zansn/mcp) produce packs that can be signed with this standard.
    npm: zansn
  3. Release

    Framework adapters: LangChain and OpenAI Agents

    Two more drop-in adapters that turn an agent run in a popular framework into a portable Zansn evidence pack.

    • @zansn/langchain: a LangChain BaseCallbackHandler that records tool calls, usage, and spans; finalize() returns an evidence pack.
    • @zansn/openai-agents: fromRunResult() turns a completed Agents SDK run into an evidence pack in one call.
    • Both reuse the local zansn assurance engine (policy, privacy, cost, latency, readiness) with no new logic, and ship as peer-dependency-only packages.
    @zansn/langchain@zansn/openai-agents
  4. Release

    Vercel AI SDK adapter

    The first framework adapter: capture a Vercel AI SDK run as a Zansn evidence pack with one hook.

    • @zansn/ai-sdk: createZansnCollector().onStepFinish captures the AI SDK per-step trail; finalize() seals the evidence pack.
    • Maps AI SDK step, tool-call, tool-result, and usage shapes to Zansn types; usage field names line up exactly.
    • Includes a runnable example and a pull-request-gate GitHub Action recipe.
    @zansn/ai-sdk
  5. Launch

    zansn: the open-source developer starter kit

    The local-first entry point to Zansn assurance is live on npm. Run an agent workflow, get a reviewable evidence pack.

    • Local evaluation runner with task, scorer, trace, finding, policy, and readiness types.
    • Tool allow/block policy checks, configurable sensitive-data pattern checks, and cost and latency checks.
    • JSON and Markdown evidence-pack exporters, pack validation, and OpenTelemetry-compatible attribute mapping.
    • Runs locally with no account; produces portable evidence that later ingests into hosted Zansn.
    npm: zansn

Follow along at npm.