Agent reliability, provable · developer preview

The reliability
layer for
AI agents.

CloakPipe certifies every agent release against your evals, enforces that decision on each model and tool call, and keeps signed evidence anyone can verify offline.

Running agents in production? Become a design partner →

1identity for the full agent release
4Evaluate · Certify · Enforce · Prove
MITinspectable enforcement core
Agent ReleaseCANDIDATE
support-agent@184sha256:8e7a3fb…
code8fd29ac
promptsupport-v31
modelprovider/model@snapshot
toolssupport-tools@17
retrievalsupport-index@892
policysupport-prod@11
Certification gateBLOCKED tool accuracy 0.972 · required 0.980
Privacy
PASS
Critical cases
1 FAIL
Evidence
SIGNED
CloakPipe Cloud · early access

Hosted registry, certification, runtime enforcement and signed audit packs.

The cloud is opening in waves. The open-source CLI works today; start with the quick start.

Join the waitlist →
01 · Evaluate
Measure
Import JUnit, Braintrust, and Langfuse results against the exact release hash.
02 · Certify
Decide
A deterministic policy decision, signed and bound to the manifest hash.
03 · Enforce
Control
Uncertified or revoked releases lose production authority at runtime.
04 · Prove
Verify
Signed ledger and certifications that verify offline, without the console.
The production problem

Your API can be healthy
while your agent is wrong.

Agent behavior changes when a prompt, model snapshot, tool schema, MCP server, retrieval index, permission, or policy changes. Conventional monitoring sees successful HTTP calls. CloakPipe tracks the complete behavior-affecting release and connects pre-production assurance to runtime authority.

Configuration drift

The production agent no longer matches the configuration that passed evaluation.

Silent regression

The model still responds, but tool choice, schema adherence, task completion, latency, or cost degrades.

Unbounded authority

A prompt injection or changed tool grants the model more authority than its certified release received.

Unverifiable logs

A mutable dashboard cannot independently prove which configuration ran or why an action was allowed.

The control loop

From change
to verified decision.

Explore the architecture →

CloakPipe does not replace your evaluation platform, gateway, telemetry, or CI system. It binds their evidence to one Agent Release, applies a deterministic decision, enforces it, and records the result. Every step below is in the open-source core or the cloud today.

01 · EVALUATE
Release assurance
Test the exact agent configuration, not an isolated prompt or model.
  • cloakpipe eval import reads JUnit XML, Braintrust experiments, and Langfuse dataset runs
  • Fail closed: an unscored case is an error, never a pass
  • release diff maps each changed component to the suites it requires
02 · CERTIFY
Signed certification
Turn results into a deterministic, reproducible release decision.
  • Rules for critical failures, pass-rate regression against a baseline, coverage, and metrics
  • Ed25519-signed DSSE certification bound to the manifest hash
  • Offline verify-cert, revocation statements, and a GitHub Action
03 · ENFORCE
Runtime control
Only certified releases receive production authority.
  • Cedar sees the release; revoked releases are blocked and uncertified ones stay pseudonymized
  • Production promotion requires a valid certification or a recorded break-glass reason
  • Sentinels revoke on live error, block, or deny rates; the MCP gate allows only declared tools
04 · PROVE
Portable evidence
Verify the basis and outcome of a decision outside the dashboard.
  • Hash-chained, Ed25519-signed ledger records bound to the release hash
  • PII types, policy decisions, and tool hops; never raw values
  • Offline verifier for ledger bundles and certifications
Open core

Inspect enforcement.
Verify proof.

Browse the repository →

The components customers must trust stay open: the Rust proxy, privacy transformations, policy runtime, release identity, evidence formats, ledger, verifier, SDKs, and adapters. CloakPipe Cloud provides managed orchestration and fleet operations.

Open-source core
Private managed cloud
Boundary
Proxy, privacy, Cedar runtime, MCP gate
Scoped keys and per-environment release pointers
Enforcement is inspectable; operations are managed.
Release schema, hashing, diff, CLI
Release registry, baselines, Releases page
Identity is portable; collaboration is hosted.
Ledger, anchoring, certifications, audit packs, verifiers
Production sentinels; per-release audit pack download
Evidence verifies without trusting CloakPipe Cloud.
EvaluationRun format and importers
Evaluation upload from CI and the dashboard
Existing tools integrate; orchestration funds the service.
Product status

Built foundation.
Focused roadmap.

We label product state explicitly. The developer preview now runs the full loop: release identity, imported evaluations, signed certification, runtime enforcement, and offline verification. make demo runs all of it locally. External anchoring and release audit packs are on main.

Available now

Rust privacy proxy, streaming rehydration, PII pseudonymization, vault, signed ledger, and offline verifier. Agent Release manifests and registry, eval importers, signed certification with revocation, runtime release enforcement, scoped keys, production sentinels, the MCP tool gate, and the dashboard. External anchoring to RFC 3161 timestamp authorities and Sigstore Rekor, signed per-release audit packs verified offline, evaluation upload from CI and the dashboard, and an India DPDP policy pack.

Building next

Importing Langfuse experiments, per-tenant Cedar policies from the dashboard, Rekor v2 and incremental multi-batch sealing, and certificate revocation checks. An RBI conduct pack is planned.

Design-partner phase

Drift detection, approval workflows, canary hooks, incident timelines, SSO, and regional deployment.

Not claimed

CloakPipe is not currently SOC 2 certified and does not yet offer a HIPAA BAA, production SAML or SCIM, managed KMS, automatic key rotation, or autonomous rollback.

Certify the release.
Control the runtime.
Verify the evidence.

We are working with design partners operating real AI agents in production. Bring one agent, one critical workflow, and the changes you need to control.

Become a design partner →