Skip to content

README.md

Latest commit

 

History

History
243 lines (171 loc) · 6.67 KB

README.md

File metadata and controls

243 lines (171 loc) · 6.67 KB

Local LLM Assistant for Burp Suite

Use your own LLM (local or remote) to generate payloads and fire them via Burp. No Burp AI credits.

Highlights

  • Seed from any request (context-menu) and mutate it automatically
  • Generate → Parse → Send in one click (URL/BODY/JSON/COOKIE params)
  • Robust JSON/YAML parsing with auto-repair for common quoting issues
  • Timing: prints generation finish and send completion with ms totals
  • Encoding variants: optionally send URL/Base64/HTML-encoded payloads
  • Repeater integration: also queue each mutation in Repeater (optional)
  • MCP bridge: expose an HTTP bridge + stdio MCP server for tools like Claude Desktop or VS Code Copilot Chat to read the seed and call “send_payloads”

Only use on systems you’re authorized to test.

What’s new

v2

  • Better defaults (fast local models)
  • Stream OFF by default; YAML-only ON
  • Added Clear output button
  • Timing markers: [gen] and [fire] Done … (fire=…, total=…)
  • Encoding toggle: send normal + encoded variants (URL/Base64/HTML)
  • More robust YAML/JSON block extraction and quoting fixes

v3

  • MCP bridge inside the extension (/v1/seed, /v1/send)

  • Small Node stdio MCP server exposing:

    • get_seed – returns the selected Burp seed
    • send_payloads – posts payloads to the bridge

  • Works with Claude Desktop and VS Code Copilot Chat (MCP)

Requirements

  • Burp Suite (Pro/Community) with Montoya API
  • Java 17+
  • For local models: Ollama or any OpenAI-compatible endpoint
  • For MCP: Node 18+
  • (Optional) Claude Desktop ≥ version with MCP; VS Code with Copilot Chat (MCP)

Build

# from repo root
mvn -q -DskipTests clean package
# Output:
# target/local-llm-assistant-<version>.jar  (shaded/fat JAR)

Load in Burp: Extensions → Add → Select JAR.

Default settings (extension UI)

  • Base URL: http://127.0.0.1:11434/api/chat
  • Model: gemma3:4b
  • Temperature: 0.15
  • Max tokens: 256

Toggles:

  • Stream: OFF
  • Payloads only (YAML): ON
  • Strip <think>: OFF
  • Debug lines: OFF
  • Send via Burp: ON
  • Also add to Repeater: OFF
  • Command mode: ON (prompt box may be empty; uses seed + controls)
  • Encode variants: OFF by default (pick URL/Base64/HTML to enable)

Usage (Local LLM tab)

  1. In Proxy/Repeater, right-click a request → Local LLM → Use this request as seed. The tab shows: Seed: <METHOD PATH @ http(s)://host:port>

  2. Either:

    • Leave prompt empty and keep Command mode ON → choose Family (NoSQL/SQL/…), Where (URL/BODY/JSON/COOKIE), # (count), then Send. or
    • Write your own prompt. The model can return either JSON or YAML (the extension handles both).

  3. Watch output and Logger → Sources: Extensions for sent traffic.

Example minimal free-form prompt (when not using Command mode):

“Generate 5 low-noise NoSQL detection probes for likely URL params in the seed. Return a single JSON object: {"payloads":[...]} with fields {name,param,payload,type}.”

Payload format (either is fine)

JSON (preferred)

{
  "payloads": [
    { "name":"boolean_true", "param":"q", "payload":"' OR '1'='1", "type":"URL" },
    { "name":"regex", "param":"q", "payload":"/.*/", "type":"URL" }
  ]
}

YAML

payloads:
  - { name: boolean_true, param: q, payload: "' OR '1'='1", type: URL }
  - { name: regex, param: q, payload: "/.*/", type: URL }

The extension auto-repairs common single-quote/escape mistakes.

Encoding variants

Enable Encode variants and choose one:

  • URL → URLEncoder UTF-8
  • Base64 → standard base64
  • HTML → HTML entity escaping

For each payload the extension sends normal and encoded variants, e.g.:

- CategorySQL_Regex (category.URL) => 200 3556B 1487ms
- CategorySQL_Regex [enc=URL] (category.URL) => 200 3556B 1668ms

MCP mode (Claude Desktop / VS Code)

1) Start the HTTP bridge (in the extension)

In the tab, tick Expose MCP bridge.

  • Port (default): 7071
  • Optional Token: set any string; clients must send Authorization: Bearer <token>

Bridge endpoints:

  • GET /v1/seed → {"method","url","headers","params":{url,body,json,cookie}}
  • POST /v1/send → {"encode":{"enable":bool,"kind":"URL|Base64|HTML"},"payloads":[…]} Returns: {"count", "results":[{"name","status","bytes","ms"}], "timing":{"fire_ms"}}

Requires JDK with the jdk.httpserver module (present on standard JDKs 17+). If Burp’s runtime lacks it, run Burp on a JDK that includes it.

2) Run the stdio MCP server (Node)

src/mcp-burp/server.js (ESM) exposes tools over stdio:

  • get_seed → fetches /v1/seed
  • send_payloads → POSTs to /v1/send

Install & run:

cd src/mcp-burp
npm i
node server.js

Set env if you customized bridge:

BURP_BASE=http://127.0.0.1:7071 BURP_TOKEN=yourtoken node server.js

3) Hook into a client

Claude Desktop Add to claude_desktop_config.json:

{
  "mcpServers": {
    "burp": {
      "command": "node",
      "args": ["/absolute/path/to/src/mcp-burp/server.js"],
      "env": {
        "BURP_BASE": "http://127.0.0.1:7071",
        "BURP_TOKEN": "your-bridge-token-or-empty"
      }
    }
  }
}

VS Code Copilot Chat (MCP) Create .vscode/mcp.json:

{
  "servers": {
    "burp": {
      "type": "stdio",
      "command": "node",
      "args": ["/absolute/path/to/src/mcp-burp/server.js"],
      "env": {
        "BURP_BASE": "http://127.0.0.1:7071",
        "BURP_TOKEN": "your-bridge-token-or-empty"
      }
    }
  }
}

Example MCP prompt (minimal)

“Use only burp tools. Call get_seed, pick a likely URL param, generate 5 benign NoSQL probes, then call send_payloads with {"encode":{"enable":true,"kind":"URL"}}. Respond only with tool calls.”

Troubleshooting

  • No seed found: Right-click inside request editor → “Use this request as seed”.
  • Bridge 401: Token mismatch. Set the same token in the extension and MCP env.
  • Nothing in HTTP history: Extension requests appear in Logger → Sources: Extensions.
  • Model too slow: Use gemma3:4b or similar, temp=0.15, max_tokens=256, Stream OFF.

Options Explained:

  • Stream: stream output from the model (nice when it works; turn off for easier debug).
  • Payloads only (YAML): enforce a single fenced YAML block with payloads: [...].
  • Strip : remove chain-of-thought noise if your model emits blocks.
  • Send via Burp: actually fire the generated payloads (through Burp’s HTTP stack).
  • Also add to Repeater: each mutation is queued in Repeater with caption LLM/.