Open-Source Guard Proxy for MCP
Security Middleware
for AI Agents
Latch sits between your agents and its MCP tools. Every tool call is intercepted, classified, and enforced against your policies.
Self-hosted · MIT Licensed
Approval Required
Agent paused — awaiting human review
Guard Proxy for MCP
Latch wraps your MCP servers. Safe operations flow instantly — shell commands, external sends, and destructive actions get caught.
Human in the Loop
Your agent wants to search for passwords or nuke a database? Latch pauses it, notifies you on Telegram, and waits for your call.
Self-Hosted. MIT Licensed.
Run on your own infrastructure with Docker. No cloud dependency. Your data never leaves your control. Read every line of code.
LLM-Evaluated Policies
Write rules
in plain English.
No regex. No YAML. Write policy conditions in natural language and an LLM evaluates each tool call against them in real-time. Or skip the LLM entirely — create rule-based policies by tool name, action class, or domain.
Pending Approval
Claude Desktop requested a shell command
Approval workflow
One tap to approve.
One tap to block.
Risky tool calls pause and notify you via dashboard or Telegram. Single-use approval tokens prevent replay attacks. Time-limited leases auto-expire.
Full audit trail
Every tool call.
Every decision.
Full history of every request, what Latch decided, and why. Sensitive arguments are automatically redacted. See what your agent tried — even when it was blocked.
How it works
Three tiers of defense
Every tool call your agent makes passes through Latch. Each one is classified by action type, matched against your policies, and routed accordingly.
Pass Through
Reads and internal writes flow automatically. Zero latency for safe operations your policies explicitly allow.
Require Approval
Shell commands, external sends, and risky actions pause until you approve — via dashboard or Telegram with one tap.
Block Entirely
Payments, destructive operations, password searches — blocked before they ever execute. Most-specific rule wins.
Under the hood
One proxy, full coverage
Policy Engine
Rules by tool name, action class, domain, and recipient. Most-specific rule wins.
LLM Policies
Write conditions in plain English. An LLM evaluates each tool call against them in real-time.
Approval Workflow
Risky actions pause for your approval. Single-use tokens and time-limited leases prevent replay.
Telegram Alerts
Get notified when your agent needs permission. Approve or deny with one tap from your phone.
Audit Log
Full history of every tool call, decision, and outcome. Sensitive arguments auto-redacted.
Action Classification
Every tool call classified — read, write, execute, send — so your policies make sense.
Tool Discovery
Auto-discovers tools from upstream MCP servers. Write per-tool policies without config.
Upstream Management
Import from Claude Desktop or Cursor configs. Supports stdio and HTTP transports.
Self-Hosted
Run on your own infrastructure with Docker. No cloud dependency. Your data stays yours.
Open Source
MIT licensed. Read the code, contribute, or fork it. Security through openness.
Common questions
FAQ
Latch is an open-source guard proxy that sits between your AI agent and its MCP tools. Every tool call is intercepted, classified by action type, and enforced against your security policies — before it ever reaches the upstream server.