Custom AI Agent MCP

Install

Works in

• Claude Desktop
• Claude Code
• Codex
• Gemini CLI
• OpenCode
• GitHub Copilot
• SourceCraft
• Koda
• Coddy

Manual install

git clone https://github.com/six2dez/burp-ai-agent.git
cd burp-ai-agent
JAVA_HOME=/path/to/jdk-21 ./gradlew clean fullJar

What it does

Custom AI Agent (formerly Burp AI Agent) is an extension that builds AI into web application security testing in Burp Suite. You right-click a request in the proxy history, pick Analyze this request, and a chat opens with the analysis. The model comes from one of the providers: Burp's built-in AI, Anthropic, Ollama, LM Studio, NVIDIA NIM, Perplexity, any OpenAI-compatible API, or the installed gemini, claude, codex, opencode and copilot commands. The extension starts an MCP server on 127.0.0.1:9876. Through it Claude Desktop, Claude Code, Codex or another MCP client reads the proxy history and the site map, sends requests, creates Repeater and Intruder tabs, starts Burp scans, gets Collaborator addresses for out-of-band checks and records findings in Burp's issue list. The full build has 59 of them, the lighter build 8. The passive scanner runs as a PassiveScanCheck in Burp Professional and looks for signs of vulnerabilities from 62 classes. The active scanner puts what it finds into a queue of suggested targets, and requests to a target go out only after you confirm. For terminal agents the skills folder holds the burp-scan skill, which leads a test through its stages: scope and passive analysis, active confirmation, out-of-band checks and issue creation.

Who it is for. For pentesters, bug hunters and application security engineers who work in Burp Suite and want to analyze traffic and confirm findings together with a language model or their own terminal agent.

Example request

Connect to Burp MCP at localhost:9876 and check the proxy history for IDOR, only on hosts in the testing scope

MCP

Transport

sse

Authentication

not required

Limitations

This is a dual-use tool: apply it only to systems and data you have permission to test.

There are two builds: the full one with 59 MCP tools, five command-line providers, an external MCP client and an optional auto-queue, and a lighter one with 8 tools. Load only one: two copies give two tabs and two MCP servers on the same port. The version is 1.1.0, it is not yet in the Burp extension store, and the lighter build is waiting for PortSwigger's review.

You need Burp Suite Community or Professional 2026.2 or newer, and the passive scanner is available only in Professional.

API keys and the MCP token are encrypted with AES-256-GCM, but the master key sits in Burp's settings next to the ciphertext, and the TLS keystore password is stored as plain text. Versions 1.0.0 and earlier wrote the MCP token and full prompts to the audit, and the context menu bypassed the privacy mode, so after updating to 1.1.0 the author asks you to rotate the affected credentials.

MIT license.

How to disable. Turn off the MCP server in Settings, then unload the extension in Extensions, Installed: after unloading no threads, timers or MCP listener remain. Delete the skill from ~/.claude/skills/burp-scan. If the server was opened to the outside, rotate the token. The ~/.burp-ai-agent/ directory with settings and agent profiles stays on disk.

README in short

The description opens with notes on versions 1.1.0, 1.0.1 and 1.0.0, followed by a six-step quick start, a table of model providers, a section on the burp-scan skill, links to the documentation and privacy notes. The repository has a docs folder with guides on MCP hardening, interface safety and external MCP servers, the pentester, bughunter and auditor agent profiles, SECURITY.md and CHANGELOG.md, and tests.