DurbarAI
Get notified

Limits

What DurbarAI cannot do yet

Reviewers trust a vendor that names its gaps first. This list covers what DurbarAI does not do today, in plain words. It follows the product's claims registry and changes when the product does.

Last reviewed 7 November 2026.

Release gate

  • It scores how an agent works through a task (tools, arguments, plan, completion, safety), not whether the final answer is right.
  • A model-graded check uses the same provider and model as the agent it grades unless you change it. Nothing enforces a different one.
  • Gate reports are sealed, but checking one offline needs Python and the DurbarAI source tree, and the signer shows as not authenticated unless you supply a trusted public key.

Audit log

  • The keys that sign the chain derive from one deployment secret held by the operator. The log exposes edits by anyone without those keys; it is not proof against the operator.
  • Run receipts are written best-effort. A run can finish without its receipt, and a missing receipt is not flagged on its own.

Stop button

  • If the store that holds the switch errors, the switch fails open and runs continue.
  • It does not cover document-intake submissions or the intake pipeline.

Approvals

  • Approvals gate tool calls. They do not review what an agent writes in a chat reply.

Company knowledge

  • Search matches words, not meaning. The embedding step is a placeholder, and retrieval by meaning is on the roadmap.

Deployment and channels

  • We do not publish container images; you build them from source.
  • There is no packaged chart for Kubernetes package managers, only kustomize manifests.
  • Agents cannot be embedded as an in-process code dependency; use the REST API.
  • Automatic TLS certificates, strict database TLS verification and a live cluster install have not been verified end to end.
  • Slack and the web widget have not been shown against a live customer workspace. Email relay only logs messages by default.

Identity and compliance

  • OIDC sign-on and SCIM provisioning are implemented but not tested against a real identity provider. OIDC is the one sign-on protocol supported.
  • There is no multi-factor sign-in.
  • Prashna Labs has no SOC 2 report and no penetration test. DurbarAI produces evidence your own audit can use; it does not make you compliant.

Agents

  • Agents do not rewrite their own prompts or tools. Changes come from people, and they reach users only through the release gate.