Watch Vekil at work
Every video was recorded in the real product and runs about a minute. All are subtitled, so they work with the sound off.
Recorded in a demo environment: models and remote servers are samples; the addresses, keys and records on screen are not real.
Core tours
Six short videos showing what Vekil does on every call, end to end.
How every number on the dashboard is computed from the evidence ledger, the fleet table and live events.
National ID, phone and email reach the model masked; the call is written to the ledger under the agent's name.
When policy asks for approval the agent waits; a misbehaving agent is cut off within a second.
Rules are built from blocks without code, compiled to Cedar and tested without touching live traffic.
A multi-step job: fetch, summarize, human approval and write back, with every step tracked live.
Legal deadlines, the processing record, impact assessments and erasure that keeps the chain intact.
Setup wizard
Connecting an application in four steps: model and provider key, agent, service key, first call.
A Claude model; only the address and key change in the app.
An application using the OpenAI SDK.
A Gemini model, with the same four steps.
With the deployment name and the resource endpoint.
Connector auth types
An agent's calls to external APIs: the credential is added from the vault, and secret echoes in responses are masked.
Even an open API is registered as a destination.
APIs that expect the key in one request header.
Several credential headers, such as an app key and an app secret.
A token in the Authorization header, the most common form.
Systems protected by a username and password.
APIs that expect the key in the URL.
The app carries its own credential; inspection still runs.
Tools from several MCP servers merged behind one endpoint.
Referer and Origin don't reach the remote server by default; an administrator turns them on where needed.