KavachKnow every registered tool
Auto-discover tools from FastMCP, LangChain and custom servers. Track schemas, auth scope and which agents can call them.
The Model Context Protocol is the connective tissue of the agentic era — and it hands agents direct access to files, databases and clouds. Kavach puts a guard in front of every tool your agents can reach.
A single LLM tool call can delete a bucket, dump a database or steal a secret. Static config scans never see it — Kavach intercepts at runtime.
Auto-discover tools from FastMCP, LangChain and custom servers. Track schemas, auth scope and which agents can call them.
sensitive_tools patterns — exact or wildcard — scope protection to aws.*, filesystem.* and beyond. Strict mode raises, advisory mode logs.
Every blocked call is recorded with rule, severity and masked payload — safe to ship to Datadog, Kafka, ELK or CloudWatch.
MCP standardizes how models reach your files, databases and clouds. That same reach is the attack surface. Kavach wraps the protocol's tool-call boundary with the same detection engine used by cloud security teams.
aws.*, filesystem.delete| MCP primitive | Risk | Kavach control |
|---|---|---|
| tools/list → registered tools | Shadow / over-privileged tools | Live inventory + ownership mapping |
| tools/call → heavy S3/Db/FS actions | Deletion, destruction, exfiltration | tool-abuse & cloud-destructive rules |
| tool arguments (prompts) | Prompt injection, secrets, PII | DetectionEngine.scan() over arguments |
| resources/read | Path traversal, data leakage | file-traversal + data-exfiltration rules |
| logging & traces | Secrets in audit logs | Masking + compliance-ready logger |
Middleware installs in minutes. The platform adds live inventory, attack-path scoring and orchestration.