Verdict · Apache 2.0 · public beta

Monitor agents in production with the evidence to act.

Verdict records the work your agent actually performs and connects supported provider calls to that work. Its guided UI takes you from an issue or usage change to the affected run, a reviewed check, and ongoing monitoring.

Get started ↓View the current UISource and docs
Agent evidence, not isolated charts

See the execution around the model call.

Application evidence

Runs, turns, tools, outcomes

Use the framework-neutral Python SDK to record turns, tool actions, tests, retries, feedback, and explicit business outcomes. Parent-run links show relationships the application records.

Model evidence

Genuine provider calls

Supported Anthropic, OpenAI, and Google SDK entry points capture request and response metadata, latency, errors, and provider-reported token usage where supplied.

Existing evidence

Import what you have

Import supported LLM telemetry, bounded text-only voice records, or local Claude Code and Codex history. Source gaps remain visible; unrelated records are not linked by guesswork.

The product flow

A clear next step from the first screen.

Home shows what was captured and what needs attention. Suggested setup builds a source-aware plan. Investigation follows each finding to the recorded event.

See the screens →
Synthetic agent run investigation with recorded tool and test failures
Current beta UI · synthetic agent run
Measure, evaluate, monitor

Different evidence supports different decisions.

Key-free checks

Recorded agent issues

Deterministic analysis identifies captured tool and test failures and links to affected runs. Schedule recurring checks with a separate worker and verify its actual cycle status.

Optional evaluation

Judge a selected scope

Preview bounded evidence for a day, session, run, turn, event, or supported call. A provider-backed judge is optional; compare candidate behavior on fixed reviewed cases.

Prospective monitoring

Model-call drift

Preview reference and current cohorts of eligible genuine model calls. Activate a reviewed policy only when its sample and measurement gates make sense. Statistical drift alerts apply to model calls.

Customer-managed single tenant

Keep the evidence in your environment.

Start with SQLite. For a shared deployment, use a dedicated PostgreSQL store, dashboard, and supervised monitoring worker. Producers can connect directly or write local spool files uploaded by the shipper to an authenticated collector.

Deployment responsibilities

  • Run the dashboard behind your TLS and access control
  • Choose capture fields, retention, and provider keys
  • Verify the worker's real cycle status and alert destination
  • Test your provider paths and data volume in your environment
Customer-managed operations guide →
Get started

Choose the evidence you already have.

Local agent history

Try Verdict on your own Claude Code or Codex sessions with no provider key and no source checkout.

python3.12 -m venv .venv
.venv/bin/python -m pip install --pre 'cognifity-verdict[dashboard]'
.venv/bin/verdict

Choose the detected local history in the setup screen, preview, and approve capture.

Production agent or existing telemetry

Instrument a Python agent with verdict.init() and agent_run(), or import a supported telemetry source. Use the full onboarding guide for provider extras and deployment topology.

python -m pip install --pre 'cognifity-verdict[anthropic,openai,google,dashboard,postgres]' cognifity-verdict-eval
Read the install and instrumentation guide →

The currently released packages are beta pre-releases, so these commands use --pre without hardcoding a version. Deterministic capture and investigation need no judge key.

Evaluate Verdict on one real workflow.

Agree on the question, capture boundary, and success criteria before starting a POC.

Discuss a POCRead the docs