Coming soon · Behavior and provenance for AI agent tools
What did this version actually do?
Coding agents install MCP servers, agent skills and editor extensions, and each one runs with an engineer's own access. Tarvite is building a review service that runs every version of such a tool in an isolated sandbox, records what it did, compares it with the version before, and checks that the published files match their source.
What a review covers
A company sends us the list of agent tools its engineers want to use. For each tool, pinned to the exact published file by its hash, the review reports:
Identity and provenance. Who publishes it; changes of maintainer or owner; install scripts; whether the published package matches its tagged source.
What it did when it ran. The files, network destinations, environment variables and credentials it touched, and the programs it started.
What it says against what it does. Tool descriptions checked for hidden characters and instruction-like text, and anything the tool does that it does not declare.
What changed since the previous version.
What free scanners and the marketplace say, and where we disagree.
One verdict per tool, with reasons
The client receives a written report, a machine-readable allow-list of the pinned versions, and the evidence behind each claim.
- Approve
- Approve with conditionsa scoped token, a configuration limit or a pinned version
- Reject
- Inconclusivethe runs did not give enough evidence, and we say what would settle it
Then we keep watching
The list becomes a watch list. When a tool on it ships a new version, we run that version, compare it with the one the client approved, and send a new verdict.
What a verdict means
A verdict records what we observed for one version, under the scenarios we ran. It is not a guarantee. Where we saw nothing wrong, we say that no disallowed behavior was observed under those scenarios. Where we could not tell, we say so.
How a review works
Pin it
We fetch the tool's published file, record its hash, and compare it with its tagged source and its maintainers' history.
Run it in a sealed sandbox
Each run gets a fresh virtual machine that holds only bait credentials and a fake internet, and is destroyed afterward. Every connection ends at a recorder, and we observe the tool from outside it.
Compare it and decide
We set what the tool did beside the version before and beside what free scanners report. A person reviews the evidence and writes the verdict.
We test only the published files, in our own sandbox. We never connect to a tool vendor's live service or to a client's systems, and hosted services are out of scope.
If we find a problem in someone else's tool
We tell its maintainers or the registry first, in private, under a coordinated-disclosure policy that we will publish on this site. A client's name never appears in a disclosure. We never sell early knowledge of an unfixed flaw.
Who it is for
Security and platform teams at companies rolling out coding agents, who decide which MCP servers, skills and extensions engineers may install.
Operators of agent-tool gateways and catalogs, who need to know what the tools they list actually do.
Security consultancies that assess agent tooling for their own clients.
Status
Coming soon: what exists and what does not
Tarvite is in development. This is where it stands.
Built or under way
- The review method and the design of the sandbox.
- Software that fetches and pins tools, checks their provenance and drives the sandbox, in development in a private repository.
- A dedicated sandbox server, kept apart from our other systems, now being installed.
Not yet
- No tool has run in the sandbox yet.
- We have published no findings.
- No client review has been delivered.
Next
- Test runs on our own planted test programs, to check the method, before any real tool.
- Reviews for a small first group of companies, then monitoring of new versions on their lists.
- Public pages for widely used tools, showing what each version did, once our disclosure policy is published.
Contact
If your teams are deciding which agent tools to allow, or you run a catalog of them, we would like to hear from you.