Govern the action.Prove what followed.

The system of record for governed agents and model calls.

Install the SDK→Explore signed lineage

Apache-2.0 · Python + TypeScript · no forced proxy

Documented governed path

decision and evidence stay together

REQUEST

Agent intent

prompt · context

MODEL

Reason

supported route

OBSVR BOUNDARY

Decide first

allow · block · redact · review

ACTION

Execute

tool · handoff

EVIDENCE

Reconstruct

signed lineage

Before dispatch

Policy owns the supported execution boundary.

Through the run

Sources and parents remain attached to descendants.

Months later

Signed records reconstruct the recorded blast radius.

The stakes

The clock is already running

Dec 2, 2027

EU AI Act - Annex III high-risk rules

90+

organizations hit by prompt injection

34.8%

of data sent to AI is now sensitive

01

For teams shipping AI agents

Your agent can email, refund, query, and commit. Logs arrive too late.

Once agents can trigger real work, the team shipping them needs a policy decision before execution and a verifiable record afterward.

The action already ran

A log can describe the email, refund, export, or tool call after the side effect is real.

The data already left

Detecting sensitive content after a provider or tool receives it is not prevention.

The decision cannot be rebuilt

Ordinary logs rarely preserve the exact actor, model, policy version, reason, and event order together.

From 2 December 2027, Annex III high-risk systems enter the deferred rules window, including lifecycle logging obligations. Turn that on late and the gap is permanent.

See the article-by-article coverage

02

The gaps

Why existing tooling fails

Not an observability tool. Not a trace viewer.

LoggingControl

Logs record the violation after it happened. They can't block the call.

FilteringEnforcement

Output filters clean up after generation. They never stop what was sent.

MonitoringProof

Dashboards show what happened. They can't produce a record an auditor can verify.

DetectionCoverage

A detector only sees the traffic that reached it. It can't tell you about the call that bypassed it entirely.

obsvr = control + enforcement + proof + coverage attestation

03

What obsvr does

Agents require a security, governance and compliance layer

Set the rules for agent actions

Version rules, approvals, and workload boundaries before deployment.

Explore Secure

04

Measured, not marketed

Governance in microseconds

25µs

TypeScript sign + enqueue

24.7–25.1µs p50 across two complete 0.11.2 benchmark passes.

64µs

TypeScript full local policy

63.2–64.1µs p50 with rules, PII scan, quotas, hooks, signing, and enqueue.

525µs

Python full local policy

524.3–525.1µs p50 under the same local-policy benchmark shape.

100Kcalls

Per stress tier

Two complete passes verified retained signatures, chain links, accounting, and declared loss.

Reproducible: BENCHMARKS.md · raw evidence retained for both runs

05

Integrations

Supported agent and model routes, one decision format

Supported provider, framework, and protocol routes produce one decision format. The coverage grade tells you exactly where policy runs before dispatch and where a separate gate is needed.

Model providers

OpenAI
Anthropic
AWS Bedrock
Azure OpenAI
Google Vertex AI

Agent frameworks

LangChain
LangGraph
LlamaIndex
Vercel AI SDK
CrewAI
AutoGen

Connect via

Explicit SDK wrap
Node import hook
MCP client wrapper
Cloudflare Workers AI

Languages

Python
TypeScript
JavaScript

See the full integration guide→Coverage belongs to a route, not a framework logo.

06

FAQ

Security & compliance answers

8 questions

12

Contact

Start with one supported action path. See exactly where policy belongs.

Bring one provider, tool, agent, or MCP route. We will map its enforce-or-observe boundary, run the public SDK locally, and show which optional hosted evidence workflows are available in beta.

A real allow, block, redact, or review decision on your route

A signed local decision record with offline verification

The exact boundary: enforced, observed, or outside coverage

An optional hosted-beta ingest and evidence-export walkthrough

Govern the action.Prove what followed.

The clock is already running

Your agent can email, refund, query, and commit. Logs arrive too late.

The action already ran

The data already left

The decision cannot be rebuilt

Why existing tooling fails

Logging≠Control

Filtering≠Enforcement

Monitoring≠Proof

Detection≠Coverage

Agents require a security, governance and compliance layer

Set the rules for agent actions

Governance in microseconds

TypeScript sign + enqueue

TypeScript full local policy

Python full local policy

Per stress tier

Supported agent and model routes, one decision format

Model providers

Agent frameworks

Connect via

Languages

Security & compliance answers

Start with one supported action path. See exactly where policy belongs.

The clock is already running

Your agent can email, refund, query, and commit. Logs arrive too late.

The action already ran

The data already left

The decision cannot be rebuilt

Why existing tooling fails

Logging≠Control

Filtering≠Enforcement

Monitoring≠Proof

Detection≠Coverage

Agents require a security, governance and compliance layer

Set the rules for agent actions

Governance in microseconds

TypeScript sign + enqueue

TypeScript full local policy

Python full local policy

Per stress tier

Supported agent and model routes, one decision format

Model providers

Agent frameworks

Connect via

Languages

Security & compliance answers

Start with one supported action path. See exactly where policy belongs.

LoggingControl

FilteringEnforcement

MonitoringProof

DetectionCoverage

LoggingControl

FilteringEnforcement

MonitoringProof

DetectionCoverage