edge-ai-MCP

MAVSDK-only drone backend for the SAR stack. All flight I/O goes through MAVSDK (Action, MissionRaw, Telemetry). No raw MAVLink.

Used internally by gateway → frontend LLM via a drone-http compatible HTTP API on port 3001.

SAR stack (recommended)

cd /home/jetson/Code
./sar-stack.sh start proxy    # or: auto | serial

The drone tmux window runs scripts/stack_server.py:

Flow:

Frontend chat → gateway :3000/infer → LLM → POST :3001/v1/apply-tool → MAVSDK → FC

Hermes (external) → https://edge-ai.basilrari.com/mcp
                 → gateway (Bearer MCP_API_KEY) → 127.0.0.1:8765 → MAVSDK

Related MCP server: osracer-mcp

Prerequisites

cd MCP
pip install -r requirements.txt

Environment

Test from frontend chat

With sar-stack running, try in the dashboard:

• "Just hover in place for now"

• "Return to home immediately" (SITL / when safe)

Test scripts

python3 scripts/test_unit.py
python3 scripts/test_sitl.py --skip-flight

Mission upload (dashboard Mission Planner)

{
  "include_takeoff": true,
  "takeoff_alt_m": 15,
  "include_rtl": true,
  "waypoints": [
    {"lat_deg": 23.558, "lon_deg": 120.473, "alt_m": 15}
  ]
}

Posted to POST /v1/mission/upload (proxied by gateway as /drone/mission/upload).

Public MCP for Hermes (optional)

Set MCP_API_KEY in sar-stack.env and restart the stack. The gateway exposes:

• GET/POST https://edge-ai.basilrari.com/mcp — streamable HTTP (Hermes; requires auth)

• GET https://edge-ai.basilrari.com/mcp/sse — legacy SSE (requires auth)

• POST https://edge-ai.basilrari.com/mcp/messages/ — MCP messages (requires auth)

Auth: Authorization: Bearer <MCP_API_KEY> or header X-API-Key: <MCP_API_KEY>.

Cloudflare (your account)

If you already have a tunnel pointing edge-ai.basilrari.com → localhost:3000 (gateway), you do not need a new tunnel, DNS record, or ingress rule. MCP is served on the same hostname under /mcp/*.

Optional hardening in the Cloudflare dashboard:

1. Zero Trust → Access — not required; API key is enforced by the gateway.

2. Security → WAF — optional rate limit on path /mcp/* to reduce abuse.

3. Keep the tunnel config unchanged unless you moved the gateway port.

Hermes

1. Add MCP server URL: https://edge-ai.basilrari.com/mcp (not /mcp/sse)

2. Transport: SSE

3. Authentication: yes — paste the same MCP_API_KEY from sar-stack.env

Verify:

curl -sS -N -H "Authorization: Bearer $MCP_API_KEY" \
  https://edge-ai.basilrari.com/mcp/sse | head -5

You should see an event: endpoint line within a few seconds.

Standalone stdio MCP (optional, local dev only)

Not started by sar-stack. For Cursor/debug on the Jetson:

python3 server.py

Deploy

git remote -v   # git@github.com-edge-ai-mcp:basilrari/edge-ai-MCP.git
git push -u origin main