VITNAcontainment for rogue AI agents
Try the demoSign in
[ compatibility ]

Which agent setups we have run.

A setup is tested once we have run it with the VITNA guard and recorded the version and date. Until then it is documented, not yet tested: its config follows the client's own documentation. The config for each is on Connect your agent.

Tested

Run with the guard: a hold was held, approved and denied.

SetupStatusHolds
Claude CodeTested 2026-10-14, version 2.1.299. Proven:
  • A hold: a call to a tool the policy holds waited for a person.
  • Approve: the approved call ran.
  • Deny: the denied call did not run.
  • Expiry: a hold nobody answered was blocked.
  • The agent's own key refused: no key can decide a hold (the decide route answers 403).
Waits for the decision

VITNA Desktop as the guard's evaluator

Guard mode can send each tool call to VITNA Desktop on your own machine instead of the cloud: the app checks it before the server sees it, and you decide holds in the app. Records are signed by a key on that machine, not by VITNA.

Run on VITNA Desktop 0.8.0 (built from the v0.8.0 release tag, run with a test identifier and its own data folder), guard 0.5.6, 2026-10-13. Proven:

  • A hold: a call to a tool the policy holds waited for a person.
  • Approve: the approved call ran.
  • Deny: the denied call did not run.
  • Expiry: a hold nobody answered was blocked.
  • The look tool: the agent changed its emoji and colour in the aquarium, and nothing else.
  • Retry after approval: the same call, retried after an approval, ran once.
  • The agent's own key refused: no key can decide a hold (the decide route answers 403).

Documented

Config known from the client's docs, documented, not yet tested.

SetupStatusHolds
Claude Desktopdocumented, not yet testedWaits for the decision
Codex CLIdocumented, not yet testedWaits for the decision
Gemini CLIdocumented, not yet testedWaits for the decision
Cursordocumented, not yet testedHeld, not run; retry after approval
VS Code (Copilot)documented, not yet testedWaits for the decision
Clinedocumented, not yet testedWaits for the decision
Continuedocumented, not yet testedWaits for the decision
Zeddocumented, not yet testedWaits for the decision
Goosedocumented, not yet testedWaits for the decision
Windsurf / Devin Desktopdocumented, not yet testedHeld, not run; retry after approval
OpenClaw (VITNA plugin)documented, not yet testedWaits for the decision
OpenClaw (MCP servers)documented, not yet testedWaits for the decision
OpenAI Agents SDK (Python)documented, not yet testedWaits for the decision
OpenAI Agents SDK (JS)documented, not yet testedWaits for the decision
Google ADK (Python)documented, not yet testedWaits for the decision
LangChain MCP adaptersdocumented, not yet testedHeld, not run; retry after approval
CrewAIdocumented, not yet testedHeld, not run; retry after approval
AutoGendocumented, not yet testedWaits for the decision
Microsoft Agent Framework (Python)documented, not yet testedHeld, not run; retry after approval
LlamaIndexdocumented, not yet testedWaits for the decision
Mastradocumented, not yet testedWaits for the decision
Vercel AI SDKdocumented, not yet testedWaits for the decision

Not supported

Setups the guard cannot sit in front of, and why.

SetupStatusHolds
n8n (MCP Client Tool)Its MCP Client Tool connects over SSE and HTTP only, not stdio, so the guard cannot sit in front of it. A remote VITNA gateway is planned.–