Vaara MCP Server

The decorator drives the same engine you can call directly when you want the decision object in hand.

from vaara.pipeline import InterceptionPipeline

pipeline = InterceptionPipeline()
result = pipeline.intercept(
    agent_id="agent-007",
    tool_name="fs.write_file",
    parameters={"path": "/etc/service.yaml", "content": "..."},
    agent_confidence=0.8,
)
if result.allowed:
    pipeline.report_outcome(result.action_id, outcome_severity=0.0)
else:
    print(result.reason)

Every call gets a risk score and an allow / block / escalate decision against your policy, then the call, the decision, and the real outcome are written to the audit trail. report_outcome closes the loop: the scorer reweights based on which signals actually predicted the outcome. Every release carries SLSA v1 build provenance as a GitHub artifact attestation (SLSA Build Level 2), and the PyPI wheel is the same file, so gh attestation verify vaara-<version>-py3-none-any.whl --repo vaaraio/vaara checks either download. Optional ML classifier: pip install 'vaara[ml]'.

Writing a trail is the easy half. The half that matters is letting someone who does not trust you check it, with no key, no access, and none of your code. Every Vaara record is content-addressed and fail-closed on authenticity, and ships with public conformance vectors plus a standalone checker that imports no Vaara code, so an independent party reproduces every verdict offline.

vaara verify-bundle evidence-bundle.json

ok only when a signature is actually established, not merely present in a log. The same property drives the standards work behind the Vaara Receipt Internet-Draft: evidence that holds up for someone who runs none of your software. The full verifier set, the trust model for each verb, and where trust comes from in each case are in docs/verifying-evidence.md.

To check that claim yourself, without installing Vaara, run the standalone checker against the published vectors. Its only dependencies are cryptography and rfc8785:

git clone https://github.com/vaaraio/vaara
cd vaara
pip install cryptography rfc8785        # the checker's only dependencies
python tests/vectors/external_evidence_v0/_check_independent.py

It re-derives every verdict from the receipt bytes and the public key alone. The output shows the property the trail is built for: a receipt dropped from inside a declared boundary is a provable gap from the held set, with no issuer access and no external witness.

For the whole loop in one runnable file, produce a signed record, verify it yourself, then watch a single forged byte get caught, see examples/prove-it-yourself/. The logs-versus-evidence argument behind it is in docs/logs-vs-evidence.md.

vaara compliance report --db audit.db --format json against a real trail produces an article-level evidence record an auditor reads directly. Articles with no recorded events return evidence_insufficient, not a rubber stamp.

{
  "system_name": "Acme HR Assistant",
  "overall_status": "evidence_insufficient",
  "trail_integrity": {"size": 105, "chain_intact": true},
  "articles": [
    {"article": "Article 12(1)", "title": "Record-Keeping (Logging)",
     "status": "evidence_sufficient", "strength": "strong", "evidence_count": 105},
    {"article": "Article 15(1)", "title": "Accuracy, Robustness and Cybersecurity",
     "status": "evidence_insufficient", "strength": "absent", "evidence_count": 0}
  ]
}

Each verdict carries the threshold-versus-observed snapshot, the rationale, and the underlying records, so a reviewer traces status back to a concrete event. The same data renders as a Notified-Body PDF, a static HTML dashboard, or a Sigstore-signed handoff envelope. See docs/COMPLIANCE.md.

• Gate every tool call you route through it against your own policy: allow, block, or escalate. Coverage is whatever you wrap; each adapter documents the surface it sees, and the Claude Code plugin states its four explicitly.
• A tamper-evident trail an outside party verifies without trusting your stack, with the chain head anchorable to an external RFC 3161 / eIDAS timestamp so its existence is provable against a clock you do not control.
• Article-level EU AI Act evidence, honest about the gaps instead of papering over them.
• Governance of the model call itself, not only the tools around it: a hardware-rooted inference receipt that a second, different local model cross-checks. This is the sovereign inference harness, new in v1.0.
• Enforcement, not only a record (v1.1.0): a credential broker mints a signed, short-lived credential bound to the attestation digest and scoped to one tool, its argument commitment, and tenant, with typed capability scopes that bound what a call may do. A gateway in front of a protected tool refuses any call without a valid, attestation-bound grant, so a bypass stops being silent. Off by default.
• Gap-evident completeness (v1.4.0): each authorization receipt can carry a signed per-boundary sequence and running count, so a dropped receipt inside a declared boundary is a provable gap from the held receipts alone, with no issuer access and no external witness (vaara verify-contiguity). Off by default.
• Independent re-mint (v1.14.0): a second generator in each public vector set reproduces the signed carrier byte-exact from the declared canonicalization alone, with no Vaara import. A verifier that has never run Vaara software reproduces the same bytes.

Native adapters route the major Python agent frameworks through the same pipeline, each via the framework's own hook, emitting identical audit events:

To put Vaara in front of an MCP server, run it as a proxy. Every tools/call routes through the pipeline before reaching the upstream; allowed calls forward transparently, blocked calls return an MCP error.

vaara-mcp-proxy \
  --upstream npx --upstream-arg -y --upstream-arg @sap/mdk-mcp-server \
  --db ./mcp_audit.db --shadow

Start with --shadow: every call is classified, scored, and recorded, nothing is blocked. After a few days, vaara trail shadow-report --db ./mcp_audit.db shows what enforcement would have done; then drop the flag and enforce, starting from a ready-made perimeter for common MCP servers in examples/policies/mcp-starters/. Point your MCP client (Claude Code, Cursor, any host) at the proxy instead of the upstream. There is also an HTTP API (pip install 'vaara[server]', vaara serve) and a first-party TypeScript client on npm (@vaara/client) for non-Python agents. Framework details, the cloud and OSS guardrail adapters (Bedrock, Azure, GCP, NeMo, Guardrails AI, LLM Guard, Rebuff), and the multi-tenant proxy are in docs/adapters.md.

A policy is code, so it belongs in the pull request that changes it. The action validates the policy, runs its cases, and fails the build on a policy that does not parse, a failing case, or a trail whose chain or signature does not hold.

- uses: vaaraio/vaara@v2
  with:
    policy: policies/production.yaml
    cases: policies/production.cases.yaml

Point trail at a signed zip to verify one a job produced. Without a pubkey that checks the trail is internally intact; pass a key you obtained separately to bind the signer too, and the run says which of the two it did. Inputs, outputs and pinning are in docs/github-action.md.

This checks artifacts. Gating the agent is the runtime's job, at the moment of the tool call.

Each risk score blends five expert signals and keeps adapting as outcomes come back, and it carries a confidence interval with a coverage guarantee that holds regardless of the input distribution. On a held-out adversarial corpus the classifier reaches 85.6% recall (95% Wilson [83.4, 87.6]) at a 5.1% false-positive rate, and 0.0% FPR on benign calls under live injection pressure. The hot-path rule scorer adds 140 µs mean per call on commodity CPU; the ML classifier is opt-in (vaara[ml]) and off that path. make bench reproduces the classifier figures below against the bundles that ship in the repo; it needs pip install 'vaara[ml]' and downloads the embedding model on first run.

• 25,556-entry adversarial corpus (250 hand-curated + 25,306 LLM-generated). The v0.40 split assigns 24,619 of them leak-free by (category, generation batch) cell: 15,114 train, 2,921 val, 2,980 test, 3,604 holdout.
• Classifier v11 (254 hand-features + 384-dim MiniLM embeddings) at threshold 0.8800 on held-out TEST n=1,827: recall 85.6% [83.4, 87.6] at FPR 5.1% [3.8, 6.9]. Recall is up 0.9 points on v9 and the false-positive rate is up 1.0. That trade is deliberate and the rest of this section is why.
• Cross-model held-out recall 86.3% over n=5,600, against attacks written by Qwen2.5-72B, a model absent from TRAIN. v9 reads 67.4% on the same set. The previous published figure here was 66.8% [64.9, 68.7], measured on v8 against the v0.37 surface; v9 reproducing it within a point on a different attacker family and freshly generated data is the reason that number survived two releases unchanged. This is still the honest worst case, and the in-distribution number above is the easier denominator.
• On four attack categories that the corpus had barely covered until v0.40, measured on 5,600 entries generated after v11's operating point was fixed: recall 63.1% to 87.0% and false positives 6.6% to 2.5% against v9. Credential exfiltration alone moves 45.9% to 80.3% at 10.0% to 3.9% FPR. Full record in bench/V11-CANDIDATE.md, including the earlier candidate that was rejected and why.
• BIPIA-pressure FPR on benign tool calls 0.0% [0.0, 0.5]: none of 704 calls flagged across four agent backends (Claude Haiku 4.5, Llama-3.1-8B, Mistral-7B, Qwen-2.5-7B). v9 flags 2.7% of the same calls and v8 35.7%. scripts/eval_bipia_entries.py reproduces all three.
• Multi-attacker PAIR robustness, measured on v0.31 and not re-run on v11: 0/25 successes per attacker across Qwen2.5-32B, Qwen2.5-72B, Llama-3.3-70B on identical seeds, Wilson upper 13.3% (bench/vaara-bench-v0.31.md).
• Distribution-free conformal coverage on the score; MWU regret bound O(sqrt(T log N)).
• Chain of custody: corpus, split, training commit, and bundle SHAs locked and printed by every script.

Method and per-cell breakdown: docs/architecture.md and bench/.

• vaara.receipt/v1 is the canonical parent spec for the signed receipt format: hash-chained, canonicalized with JCS (RFC 8785), verifiable offline from a public key. The x402 settlement binding and an eIDAS qualified-timestamp profile are downstream profiles that pin to it rather than competing formats. Receipts can carry a self-hosted RFC 3161 timestamp that Vaara mints offline.
• SEP-2828 signed execution records, carried forward as the IETF Internet-Draft draft-sirkkavaara-vaara-receipt. A second independent implementation has reproduced the SEP-2828 conformance vectors from a clean checkout with no shared code. Published corpora with independent checkers cover the fallback binding path (tests/vectors/fallback_projection_v0/) and CrewAI governance decisions (tests/vectors/governance_decision_v0/).
• OVERT 1.0 (overt.is): Vaara is the Arbiter and emits Protocol Profile 1.0 Base Envelopes (canonical CBOR, Ed25519) alongside every record when attestation is on.
• Post-quantum: an optional parallel ML-DSA-65 / FIPS 204 signature over the same preimage, so a stripped post-quantum signature is a detectable downgrade rather than a silent loss.
• Root-agnostic evidence: the same Article 12 record is provable with or without a hardware TEE and re-expressible in IETF RATS EAR claims (AR4SI vector), whether rooted in a TPM 2.0 host, an AMD SEV-SNP confidential VM, or no TEE at all.
• Sovereign inference harness (v1.0): a local model behind a signing proxy that emits a hardware-rooted inference receipt a second local model cross-checks. Developed privately, published here under AGPL-3.0.

Details and the offline checkers for each: docs/standards.md.

The public surface is fixed: the signed envelope (vaara.receipt/v1), capability constraints, the credential grant and gateway, and the @vaara.govern entry point. No new primitives are planned. New behavior ships as profiles that pin to vaara.receipt/v1, not as new core types, and no new format bindings will be added (the last was v1.13.0). From here the work is hardening and subtraction within this surface, so anyone building on it has a stable target.