What is DroneServer?
DroneServer, from Peter J. Burke's group, lets an MCP client fly a MAVLink aircraft, ArduPilot or PX4, simulated or real. The model calls named tools such as takeoff, go_to_location and kill_motors; the server turns each into MAVLink through MAVSDK and returns telemetry. It registers 98 tools, which its authors say cover 223 of the 238 client-side MAVSDK methods, with the other 15 documented as not applicable.
It is the software for the arXiv paper 2601.15486, which is under review. The repository is a fork of ion-g-ion/MAVLinkMCP, MIT-licensed, written in Python 3.11+, and released v2.0.2 on 4 September 2026, with its last commit on 20 September 2026. It has about 10 stars.
How to install DroneServer
git clone https://github.com/PeterJBurke/droneserver.git && cd droneserver && uv sync && uv run python -m droneserver.server --transport stdioIs DroneServer worth it?
Most hardware MCP servers trust the model. DroneServer is built on the opposite premise, and in our reading of the code the safety layer does what the README says. Every tool passes through a guard applied at registration. Each tool has a tier, and one missing from the tier table is treated as critical. Critical tools such as kill_motors, vehicle_power and autopilot_shell return a one-time token on the first call and run only on a second call quoting it; tokens are bound to the tool and its arguments and expire after 60 seconds. If a check itself throws, the call is refused.
The envelope is set in SAFETY_ environment variables, all on by default: a 120 m altitude ceiling, 20 m/s top speed, a 1,000 m radius around home, an optional polygon fence the server enforces itself, and per-client rate limits of 60 calls a minute and 6 critical calls. Out of the box no API keys are configured, so every client gets telemetry only, and over stdio, which carries no headers, you must deliberately grant control. The suite has 937 unit tests, 116 SITL integration tests against ArduCopter 4.5.7, and 29 adversarial cases covering prompt injection and forged tokens.
The catch: the edges. The code default for the HTTP host is 0.0.0.0; only the example .env file narrows it to 127.0.0.1, and the authors expect a private tailnet with no public ports. State preconditions such as must be airborne fail open when telemetry cannot be read, a deliberate choice so a glitch does not strand a flying aircraft. The operator can turn the whole layer off with one variable; the audit log then records that. Emergency land and return-to-launch skip confirmation by design. The README says the software has been tested in simulation, and real flight is the operator's responsibility.
Our take: as a reference design for putting a language model near something that can fall out of the sky, this is the repository to read. As something to fly, treat it as a SITL tool with a manual RC override in reach, and expect a small research project rather than a maintained product.
Where it earns its place
A university UAV lab evaluating language models as mission planners
Compare how different models handle the same missions under identical guardrails
Run the bundled benchmark harness against ArduPilot SITL in Docker; every call lands in the append-only audit log with the rule that allowed or refused it. Free apart from model tokens.
A drone start-up's safety engineer
Red-team an agent's control path before any real flight
Replay the 29 adversarial cases, then add site-specific ones such as out-of-fence waypoints and replayed tokens, with SAFETY_GEOFENCE_POLYGON set to the real test field.
A survey team with an ArduCopter on a test range
Let an operator task the aircraft in plain language while a pilot holds the RC override
Bind the server to a Tailscale address, give the operator's client a control-scoped key and keep critical tools behind the confirmation handshake. The README requires visual line of sight and a live manual override.
Before you commit
- Setup
- git clone, uv sync, copy .env.example to .env and point MAVLINK_ADDRESS at the aircraft or SITL, configure SAFETY_API_KEYS, then uv run python -m droneserver.server --transport stdio (HTTP/SSE is the default transport). docs/reproduce.md walks through SITL.
- Integrations
- ArduPilot and PX4 over MAVLink via MAVSDK, on TCP, UDP or serial; ArduCopter SITL in Docker; any MCP client, plus a scripted LLM-in-the-loop harness. Some MAVSDK calls, such as set_actuator_control, are accepted but inert on ArduPilot.
- Data and privacy
- Runs on your machine or a companion host. Telemetry goes to your model provider through the client. An append-only JSONL audit log records every call. API keys carry telemetry, control or admin scope; unauthenticated clients get telemetry only.
- Cost at scale
- Free under MIT. Costs are model tokens, the aircraft and a host on a private network.
- Limits
- Software emergency stop needs the network, server and model all working; docs/estop.md ranks it as the weakest of four ways to stop the aircraft. Unit-test line coverage is low by the authors' own account, because most tool bodies only run against a simulator.
What’s good
- Confirmation tokens bound to tool and arguments
- Server-side geofence independent of the autopilot
- Telemetry-only by default until API keys grant control
- 937 unit tests plus SITL and adversarial suites
What’s not
- Tested in simulation only, by its own account
- HTTP host defaults to 0.0.0.0 without the example .env
- Preconditions fail open when telemetry is unreadable
- About 10 stars; a research artefact under paper review
Best for
Researchers and drone developers who want to put a language model in charge of an ArduPilot or PX4 aircraft in simulation first, with refusals they can audit.
DroneServer alternatives
DroneServer: common questions
- Is DroneServer worth it?
- A drone server with the most carefully built guardrails we have read, made for researchers and simulation, and honest that refusing bad commands is not the same as making a drone safe. We rate it 3.7 out of 5.
- How much does DroneServer cost?
- Free and open source (MIT). You supply the aircraft or an ArduPilot or PX4 simulator.
- What are the downsides of DroneServer?
- Tested in simulation only, by its own account; hTTP host defaults to 0.0.0.0 without the example .env; preconditions fail open when telemetry is unreadable; about 10 stars; a research artefact under paper review.
- Who is DroneServer best for?
- Researchers and drone developers who want to put a language model in charge of an ArduPilot or PX4 aircraft in simulation first, with refusals they can audit.
- How much does DroneServer cost for a team or in production?
- Free under MIT. Costs are model tokens, the aircraft and a host on a private network.
- Is DroneServer safe for company data?
- Runs on your machine or a companion host. Telemetry goes to your model provider through the client. An append-only JSONL audit log records every call. API keys carry telemetry, control or admin scope; unauthenticated clients get telemetry only.
- What are the best DroneServer alternatives?
- The closest alternatives we have reviewed are ROS MCP Server and Bagel.
Answers checked Nov 16, 2026.