Tool Guard and CrashMire
Tool Guard locks. CrashMire checks.
Tool Guard is free and open source: it puts your AI agent's rules in code, on its tools. CrashMire is paid: it finds the rules you are missing, then checks that the locks hold.
Tool Guard, free
Under the MIT licence, in PHP (a Symfony bundle), Python and Node: for Symfony AI, LangChain, the Vercel AI SDK, the OpenAI Agents SDK, MCP servers and any agent loop. It also sits in front of an MCP server you did not write (mail, drive, bank) and blocks a tool that changes or gives the model orders. Whatever the model was talked into, a refused call never runs.
- Permissions per tool: who may do what, on which order
- Limits: amount, frequency, number of calls
- Human approval above a threshold
- A journal of every refused or held action
Measured, not promised: on Tool Guard's demo shop, 19 of 25 attacks run without it, none with it (corpus replayed by its tests, published with the code).
CrashMire, paid
For any company with an AI agent, whatever its language or platform. We rebuild your agent with fictitious data, without access to your systems, and we manipulate it.
- Your agent facing manipulation: fake superior, hidden instruction, insisting customer, trapped tool
- What gave in, replayed step by step, on several models
- The rules to put in code for each weakness
- A new run with the fix, to check it holds
The two together
The crash test finds
Where your agent gives in, and under what pressure.
The lock blocks
You put the rule in code: with Tool Guard, in PHP, Python or Node, or in your own code.
The crash test checks
We replay the same situations: the rule holds, observed, not assumed.
Tool Guard already runs on our demo
On the public demo, "Léa after crash test" refunds through Tool Guard: up to €50 the refund goes through, above it Tool Guard holds it for Sophie, the manager, even when Léa was talked into paying. When Tool Guard is what blocked it, the conversation says so.
Why a test is not enough, and neither is a lock
An AI that knows it is being tested can behave during the test, like a car that detects the test bench. A lock trusts no one: it blocks in production, whether the AI is sincere or not. But a lock does not tell you which rules to set nor whether it was set right: the crash test shows that.