Stridemap Labs builds security infrastructure for the AI agent era. We help enterprises govern, monitor, and protect their LLM-powered agents — so they can deploy with confidence.
Why Stridemap Labs
We're building the security infrastructure that makes autonomous AI safe for enterprise deployment.
Enterprises are deploying LLM agents that read emails, query databases, call APIs, and execute code. They're powerful — and ungoverned.
Traditional security tools weren't built for AI agents. Prompt injection, data leakage, and unauthorized actions are the new attack surface.
SOC2, HIPAA, PCI-DSS, and ISO 27001 all require audit trails and access controls. AI agents need the same governance as human operators.
Stridemap Labs builds the security layer that sits between agents and their tools — intercepting, governing, and logging every action.
Products
Assistants now reach into Linear, Drive, Notion and everything else through the Model Context Protocol. MiniGate is the control point that sits in the middle of it.
A security gateway for the Model Context Protocol — the standard by which AI assistants reach outside tools. MiniGate sits between the assistant and the vendor's MCP server and mediates every message in both directions. Connecting an assistant to a SaaS tool is all-or-nothing today: the vendor decides which tools exist and the client hands every one of them to the model. MiniGate is the layer that was missing.
Runs as a local shim or a hosted gateway. No model-channel access, no TLS interception, no credential custody — OAuth is relayed, never minted. Powered by AgentGuard.
# Policy as code, reviewed like any other config
rules:
- id: MG-010
name: Block deletion across all servers
conditions:
tool_name: ".*\.(delete|remove|drop).*"
action: BLOCK
# 57 tools offered by the vendor
# 36 reach the modelThe governance library MiniGate runs on — policy evaluation, local PII detection, prompt-injection scanning and the cryptographic audit chain. It is also embeddable directly in a LangGraph or LangChain agent for teams who want the controls without a gateway in front.
Use Cases
Every one of these is a real failure mode of connecting an AI assistant to a work tool — not a hypothetical.
Linear's MCP server ships 57 tools, four of which delete things. You wanted read and comment. A policy file gets you 36 — no waiting on the vendor for finer scopes.
The name never changes, so every allowlist still says ALLOW. MiniGate hashes the whole definition at approval. Real case: Linear altered save_issue between two days in September. The diff exits non-zero, so it gates CI.
Anyone with your intake form URL can write into your database. A ticket saying “ignore previous instructions” is a code path an attacker controls without touching your infrastructure. Scanned on the way back, after policy already said the call was fine.
Names, emails and card numbers are redacted in flight; an SSN refuses the call outright, because redaction is the wrong answer when the data should not be in play. Scanning runs locally — nothing leaves the machine to be classified.
Read from the CRM: allowed. Post to the public tracker: allowed. In that order it is a breach. Every allowlist and DLP appliance evaluates one request at a time, so the information is never in front of them.
SDK callbacks need a line in the agent. You cannot add one to Claude Desktop, Cursor, or a vendor's agent. A proxy in the protocol path does not ask the code for permission.
MCP is bidirectional. A server can ask your client to run the model and hand back the result, or render a form to your user asking for an API key. The spec says it must not. That is an expectation, not a control.
“Show me what your agents did” is answerable with logs. “Prove nobody rewrote them” is not. Each event's hash covers the one before it, so any edit breaks the chain from that row on. Verification walks the whole thing in milliseconds.
MiniGate never sees your prompts and does not proxy the model API. Everything it governs is visible in the protocol itself — which tool, with what arguments, and what came back. No TLS interception, and no credentials held: OAuth is relayed, never minted.
How You Use It
No SDK to adopt and no agent code to change. MiniGate runs as a local shim beside your client, or as a hosted gateway your whole team points at.
A custom connector URL, or two lines in your MCP client config. Nothing in your agent changes — MiniGate speaks the same protocol the client already speaks.
https://gateway.example.com/mcp/linear
YAML, in your repo, reviewed in a pull request like any other config. Rules match on tool name, data pattern, direction and destination, in priority order.
tool_name: ".*\.(delete|remove|drop).*" action: BLOCK
One sync records what each tool looked like the day you reviewed it — the whole definition, not the name. When a vendor changes one afterwards, you find out.
minigate registry sync --server-name linear
Every decision lands in an append-only, hash-chained record. Verification walks the whole chain and tells you whether a single row was ever altered.
minigate verify
{ "is_valid": true, "total_events": 212 }Faults fail open, policy fails closed. An error inside MiniGate lets the call through — a security tool that breaks your agent gets uninstalled. A rule that says block, blocks.
How it works
Every message MiniGate mediates passes through these. They are AgentGuard, the library underneath — usable on their own if you would rather embed the controls than run a gateway.
Interceptor
LangGraph BaseCallbackHandler — captures every tool call with zero code changes.
Policy Engine
YAML rules with ALLOW / BLOCK / ALERT / BYPASS. Hot-reload, regex conditions, AND/OR logic.
DLP Scanner
Microsoft Presidio PII detection — 15+ entity types, 100% local, zero data egress.
Injection Detector
Two-stage: fast regex patterns (<5ms) + LLM classifier. OWASP LLM01-06 coverage.
Audit Chain
SHA-256 cryptographic chaining — tamper-evident, SOC2 and HIPAA audit-ready.
Early Access
Join the waitlist for early access to Stridemap Labs products. Help shape the future of AI agent security.