Write-up
Wiring BurpMCP-Ultra to a Local Model: Driving Burp Suite with LM Studio / Bionic
Setting up a local, offline LLM with BurpMCP-Ultra and LM Studio or Bionic to keep authorized LLM red team testing on your own machine, and where MCP tool calling gives way to generating prompts by hand.
Wiring BurpMCP-Ultra to a Local Model: Driving Burp Suite with LM Studio / Bionic
Most of my Burp Suite work is wired up to a hosted assistant, and for the majority of engagements that is exactly what you want. But recently I helped a friend who was working on an LLM chatbot and had been required to run jailbreak test cases against it, covering the prompt injection and jailbreak categories from the OWASP GenAI / LLM Top 10. Testing an LLM is a different shape of problem from testing a web app. A good part of that work is generating adversarial and jailbreak prompts, and hosted assistants will, reasonably, decline to help with parts of that while you are in the middle of an engagement. That is expected behaviour, not a complaint, but it does stall the loop when the work is authorized and you just need to keep moving.
The pragmatic answer is to run a model locally. Everything stays on the machine, there is no third party in the request path, and the workflow does not pause on a refusal you have to work around. I had wanted an excuse to set this up for a while, so I documented every step. This is that setup.
A note before the steps. Point this at systems you are authorized to test, and nothing else. Running the model locally does not change the rules of engagement. It only changes where the compute lives.
What You Need
- Burp Suite with the BurpMCP-Ultra extension installed
- LM Studio or Bionic installed. I will skip their installation here, since both are a standard download and run
- A machine with enough headroom to hold a midsize model in memory
Choosing a Model
Once LM Studio or Bionic is installed, open Settings > Explore and search for the model you want to run. I started with a 27B model:
https://huggingface.co/HauhauCS/Qwen3.6-27B-Uncensored-HauhauCS-AggressiveOn paper my hardware should have handled it, an M5 Pro with 48 GB, and with plain prompting it ran fine. The problem showed up only when I drove it through MCP. Tool calling against Burp made it noticeably slow. Worth calling out, because it reads like a config mistake when it is really just the model size under an agentic workload.
Dropping to a 9B model fixed the responsiveness with no meaningful loss for this task:
https://huggingface.co/HauhauCS/Qwen3.5-9B-Uncensored-HauhauCS-AggressiveThe figure below shows the model downloading:
Configuring BurpMCP-Ultra
In Burp, open BurpMCP-Ultra > Server. Two values matter here, the MCP SSE endpoint and the Authorization token. Note both down, because you will paste them into LM Studio next.
Connecting LM Studio or Bionic
Back in LM Studio or Bionic, go to Settings > MCP and enter the details you just copied, the SSE endpoint and the Authorization token.
Verifying the Connection
Once the details are saved, the MCP connection registers. You should see BurpMCP show up as connected:
Driving Burp Through the Model
With the connection live, you can hand Burp tasks straight to the model, so it drives the extension rather than you clicking through it. As a quick sanity check, the figure below shows it retrieving a JWT through the MCP link:
When MCP Slows You Down: Generating Manually
With the model connected, you can ask it to craft the test payloads for you directly. After spending some time driving Burp through MCP, though, I settled on running the local model on its own, outside MCP, and copying the output across by hand. For payload generation the performance was simply better that way, and the manual step costs almost nothing.
For this I moved to a larger model:
https://huggingface.co/HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-AggressiveThe image below shows the local model generating the jailbreak test prompts for the engagement, without the refusals that interrupt a hosted assistant partway through a task.
Closing Thoughts
None of this is exotic. It is plumbing, but it is the kind of plumbing that quietly widens what your testing rig can do. A local model wired into Burp gives you an assistant that stays on your machine, keeps engagement data in house, and does not drop out of the loop during authorized work.
Two caveats worth keeping in mind. First, model size matters far more under MCP than under plain prompting, so start smaller than you think you need and scale up only if the quality demands it. Second, the convenience cuts both ways. A model that will drive Burp for you is only ever pointed at scope you are cleared to test. Keep it there.