Scan an MCP server before your AI connects.
MCP servers and APIs can place instructions directly in an AI agent's context. Toolproof checks the model-visible text for prompt injection and unsafe tool design, then gives you a signed letter grade and the evidence behind it.
Scan an MCP server free- 1
Paste the endpoint
Use the MCP server or API URL you are considering.
- 2
Review the evidence
See exactly which descriptions, prompts or declarations triggered a finding.
- 3
Decide before connecting
Use the grade, signed passport and your own risk policy to decide what happens next.
Prompt injection
Instruction overrides, exfiltration requests and covert-action language in tool text.
Hidden text
Zero-width, bidirectional and other invisible Unicode characters that can change what a model reads.
Unsafe tool design
Credential-shaped parameters, destructive defaults, missing authentication signals and plaintext transport.
Proof you can keep
An ed25519-signed passport that can be verified offline and shared with your security review.
What does an MCP security scanner check?+
Toolproof inspects the public text and declared surface of an MCP server or API for agent-hijack patterns: hidden Unicode, instruction overrides, exfiltration language, embedded credentials, risky parameter defaults, insecure transport and missing authentication signals.
Does scanning a server execute its tools?+
No. Toolproof evaluates the endpoint and the model-visible descriptions it exposes. A scan is static analysis, not a sandbox or a guarantee about every future deployment.
Can I verify a grade independently?+
Yes. Each verified result carries an ed25519-signed passport over canonical JSON. The public key and an offline verification example are available in the API docs.
Make the check automatic.
Add one rule to your agent's instructions so it verifies unfamiliar tools before connecting and reports the grade to the user.
Set up the agent rule →