Skip to main content

Open source · Apache-2.0

Rehearse every action.
Ship with evidence.

Stateful rehearsal and outcome verification for AI agents. Run your agent's tool calls against simulated systems, inject the failures that matter, verify the final state — and block a release when a new model, prompt or policy regresses.

Get StartedProduct tourView on GitHub
Verixa console overview — pass rate, verdict trend and recent rehearsals

The Verixa console, captured from a running server with demo data.

The answer sounded right. The refund happened twice.

An agent can give a convincing final message and still change the wrong object, retry a payment that already committed, ignore a permission boundary, or follow instructions hidden in a support ticket. Grading the last sentence misses all of it.

Verixa grades actions and outcomes. Every tool call runs against a stateful simulated world with real policy, real idempotency and real balances. Faults land exactly where you put them — before or after the commit — and assertions check what the world looks like when the agent is done.

Questions a transcript can't answer

Did the refund happen once?Stateful ledger, balance and idempotency keys
Did a timeout hide a successful commit?Faults injected before or after the simulated commit
Did approval cover the actual request?Exact order and amount matching
Did ticket text change infrastructure?Tool policy and final deployment state
Did the new version regress?Same-revision comparison and a release gate

Stateful simulated systems

Orders, refunds, approvals, support tickets and deployments with real balances, idempotency keys and per-world state — never a production gateway, ticket system or cluster.

Precise fault injection

Timeouts, responses lost after commit, expired credentials, rate limits and malformed results — at an exact tool and invocation, before or after the simulated commit.

Policy that actually binds

Allowed tools, refund limits and replica caps are enforced by the engine even when the agent asks for something else. Denials are recorded as violations.

Outcome assertions

Check the final world, not the transcript: refund count and amount, ticket status, replica count, call budget, violations. Missing paths fail.

Release gate

Compare two runs of the same scenario revision. Any failing candidate or regressed assertion blocks the gate — in the console or with exit code 3 in CI.

Evidence and replay

Every run records arguments, results, denials, state diffs and a hash chain. Export evidence, export the action trace, and replay it against fresh state.

Bring any model

Any OpenAI-compatible tool-calling endpoint proposes calls; Verixa executes them against the simulated world. Network calls are opt-in, token usage is recorded.

A real console

Netra-style sign-in, overview with pass-rate and verdict trend, an evidence drawer with an action timeline, a fault studio and version comparison.

Deploy anywhere

Zero runtime dependencies. One script deploys over SSH as a hardened container, a K3s Helm release or a systemd service, with HTTPS and a session login.

A real product, not a mockup

Captured from a running server with the bundled demo data. See the full tour →

Evidence drawer — outcome checks and action timelineFault studio — step builder with live previewVersion comparison — release blocked

Running in one command

Python 3.11+, no runtime dependencies, no frontend build and no API key for the demo. The demo seeds eleven real simulator runs — ten reference runs and one known regression.

Gate CI on exit codes: 0 passed, 1 error, 3 failed outcome or regression.

Read the quickstart →
git clone https://github.com/zyvorai/verixa && cd verixa
VERIXA_TOKEN=Admin@321 python3 -m verixa serve --demo
# open http://127.0.0.1:8788 — sign in as admin / Admin@321

python3 -m verixa run all --junit artifacts/reference.xml
python3 -m verixa run ticket-injection --agent regression  # exits 3

Open, and honest about its limits

Apache-2.0. CI on every push runs the Python suite on 3.11–3.13, a console-to-API harness, real Chromium tests and a container build. Evidence is labelled simulated-tool-state: Verixa proves what your agent did against simulated systems, not that it will behave identically in production. The event hash chain detects edits; it is not a signature.

Read the security model →
CI statusApache-2.0Python 3.11+Zero runtime dependencies

Bring your own scenarios

New simulated tools, fault types, agent adapters and scenario packs are all welcome. Start with an issue or a pull request.

Star on GitHubContributing guide