VITNAContain your agents.
Prove that you did.
AI agents take real actions now. They send, they buy, they deploy, they delete. Sometimes one does something nobody asked for, and by the time you find out, it is done. VITNA checks a risky action before it runs and returns allow, hold or block. In guard mode it enforces that answer for the MCP servers it wraps; elsewhere your code does. Then it signs a record of the decision that anyone can verify without trusting us.
Containment covers the MCP servers you put behind the guard, and nothing outside them. Every session record says which ones those were.
VITNA produces Ed25519-signed evidence records that anyone can verify offline with a published public key and an open-source verifier, with no need to trust VITNA's servers. Its detection is heuristic, and those limits are documented publicly. In guard mode it refuses to forward a tool call it has not allowed, for the MCP servers put behind the guard and no others; everywhere else your system enforces the decision. Either way, it is independently verifiable proof that an AI agent's actions were checked, and what was decided.
How the three categories of agent safety tooling compare, and where each falls short: AI Agent Safety and Compliance Tools: A 2026 Comparison.
Four agents that went rogue in production.
PocketOS
A coding agent hit a credential mismatch in staging, went looking for a token, and found one in an unrelated file that had been created to manage custom domains. The token could also delete volumes. One API call deleted the production database and the volume-level backups stored with it. Railway later restored the data.
“It took 9 seconds.”
Replit
During a code freeze, Replit's agent deleted SaaStr founder Jason Lemkin's production database, then told him a rollback was impossible.
“It turns out Replit was wrong, and the rollback did work.”
Gemini CLI
Asked to move a user's files into a new folder, the agent's mkdir failed silently, it read its own move commands as successful, and the files were lost.
“I trusted these messages without verification.”
An OpenClaw inbox
Summer Yue, a Meta AI security researcher, told her agent to confirm before acting on her inbox. It started deleting instead, and did not stop for the commands she sent from her phone.
“I had to RUN to my Mac mini like I was defusing a bomb.”
None of these needed a model to be malicious. Each needed an agent to act on a wrong guess, with nothing outside it to catch the action. The seven questions that follow from them: Seven questions to ask before your agent goes rogue.
Enforcing is not the hard part. Proving it is.
Gateways, identity providers and policy engines enforce, and then you trust their dashboard that enforcement happened. VITNA signs a record of each decision with Ed25519, and anyone can verify it offline with a published key, without trusting us, our servers or yours.
Outside guard mode, VITNA is not in your execution path: your code asks it and your code runs the action. In guard mode it is, for the MCP servers it wraps and no others: a call reaches the server only if VITNA allows it. When the MCP server cannot reach VITNA, it tells your agent to stop rather than proceed, because a check that did not run must never read as a pass.
Because the record is produced by the check itself, it exists whether or not anyone remembered to collect it. That is what record-keeping under Article 12 of the EU AI Act depends on, and it is what a customer, an insurer or an auditor can check after the fact.
Two stacks, one missing answer.
The buyer's actual problem
AI agents now produce regulated data flows. A retrieval-augmented chatbot touches a Quebec resident's record. A coding assistant logs an Aadhaar. The buyer's question isn't "can I see what my agent did?" or "am I compliant?". It is both at once: "what just happened, and was it legal?"
Why this doesn't split cleanly along existing tool lines
Observability tools (Datadog, Sentry, Splunk) capture what happened. Compliance tools (Drata, Vanta) document policies + run audits. Neither answers "is this agent action legal in real time?" The buyer has to wire two stacks together and write their own glue, usually after a regulator asks.
Why VITNA is built differently
VITNA captures every event AND tags it with the regulatory frame (DPDP / GDPR / PIPL / etc.) AT INGEST. The DPIA, ROPA, breach classifier, and cross-border registries all run against the live event stream. The dashboard and the compliance reports are the same surface, drawn from the same database.
Side-by-side.
Numbers you can verify.
Verify any number: /api/compliance/global-status, /api/.well-known/openapi, and /api/health/security are public + machine-readable.
Four things that make us different.
Built for the people running the agents
VITNA is built for whoever puts an AI agent in front of real systems. A solo founder can start free in minutes, with no card and no audit kickoff. A security or platform team can start with a fixed-scope, four-week pilot. Either way, your own people decide the held calls.
Honest pre-audit readiness
We don't claim SOC 2 / ISO 27001 / HIPAA attestations we don't have. We DO publish detailed pre-audit readiness scorecards mapping every control to evidence. Procurement teams get the truthful answer they can verify.
Indigenous-owned
COSTRINITY INC. is an Indigenous-owned company, run from Treaty 4 territory (Regina, Saskatchewan).
Local-first
VITNA Desktop keeps events, decisions and holds on the operator's machine, and its guard sends nothing to our servers. On the API, storage_mode=local checks each event and returns it without storing it.
VITNA for agents, called by agents
The @costrinity/vitna-compliance-mcp package exposes 23 VITNA compliance tools as MCP. Your AI agent can ask, mid-task: "Is this transfer DPDP §16 OK? Is this incident reportable? Is this AI use case high-risk under the AI Act?" and get the answer before it acts.
Try VITNA in three steps.
Read the surface
Hit /docs or read the API contract at /api/.well-known/openapi, which documents the public subset.
Wire up MCP
Add @costrinity/vitna-compliance-mcp to your agent. Your LLM can now call VITNA.
Want a person to walk your actual systems through the whole Act? Fixed-scope EU AI Act readiness assessment, $7,500 US — book a 30-minute call.