Edge Connector

Edge Connector

Local AI agent security enforcement for Linux edge devices, with optional Fleet escalation over MQTT.

The Edge Connector is a C engine for Linux edge devices such as Raspberry Pi, Jetson and industrial single-board computers. It checks an agent's tool calls on the device with local rules and content inspection. When local rules cannot decide, it asks a DefenseClaw Fleet gateway over MQTT. It works with no network at all for the local checks.

Why a separate engine?

Use a connector for desktop coding agents such as Claude Code, Codex and Cursor. Those assume a reliable network, plenty of memory and a person at the keyboard. Use the Edge Connector when the agent runs on a small device, the network comes and goes, or a bad tool call causes a physical action.

What it does

  • Detects secrets, PII, credentials, exfiltration, prompt injection and dangerous commands in tool arguments and LLM responses.
  • Blocks requests to private ranges and cloud metadata addresses, and allows network destinations only from a public-IP allowlist.
  • Tracks multi-step capability patterns, for example fetch from the network then run a shell command.
  • Applies fixed rate limits: 60 tool calls, 30 network requests and 10 physical actuations per minute.
  • Keeps an HMAC-chained audit ring so changes to retained history are detectable.
  • Takes signed policy updates over MQTT with a rollback canary. Production builds accept HMAC updates only when built with DCLAW_ALLOW_HMAC_OTA=ON and given a separate HMAC key.

Evaluation path

The engine runs seven checks in order and stops at the first local block. A cache hit avoids a new Fleet request, but the earlier local checks still run.

cache miss
1. Validate inputname, hash, session
2. Rate limit3 token buckets
3. Scan contentsecrets, PII, injection
4. Check destinationSSRF and allowlist
5. Match sequencesrisky multi-step patterns
6. Verdict cacheby tool hash
7. Ask FleetMQTT 3.1.1

List view for small screens. Use the expand button to open the drawing.

  1. 1.1. Validate inputname, hash, session
    • 2. Rate limit
  2. 2.2. Rate limit3 token buckets
    • 3. Scan content
  3. 3.3. Scan contentsecrets, PII, injection
    • 4. Check destination
  4. 4.4. Check destinationSSRF and allowlist
    • 5. Match sequences
  5. 5.5. Match sequencesrisky multi-step patterns
    • 6. Verdict cache
  6. 6.6. Verdict cacheby tool hash
    • cache miss7. Ask Fleet
  7. 7.7. Ask FleetMQTT 3.1.1
The seven checks of one tool call, in order. Checks 1 to 5 are local and any of them can block. A Fleet verdict from check 6 or 7 decides what the local checks leave open.

The result is ALLOW, BLOCK, WARN or PENDING. The Fleet gateway decides unknown tool hashes by tool hash, and blocks them by default.

Build profiles

ProfileFeaturesUse
MINIMALNo MQTT, no content scanConstrained builds with no Fleet
STANDARDMQTT and content inspectionRaspberry Pi, Jetson, single-board computers
EDGEMQTT, content inspection, bloom filterEdge gateways

Measure size, memory and latency on your own target; TLS libraries and compiler flags change them.

See the install guide for the build and the Fleet connection steps.

Next steps