MCP Server
Humanbound exposes an MCP (Model Context Protocol) server that lets AI coding assistants (Claude Code, Cursor, Windsurf, GitHub Copilot) query your security testing data directly. Ask natural-language questions about posture, findings, coverage, and experiments without leaving your IDE.
Setup
The MCP server is built into the Humanbound CLI and ships as an extra. Install it first:
zsh users
Quote the extras ('humanbound[mcp]') — unquoted, zsh treats [mcp] as a glob and fails with no matches found.
hb mcp speaks MCP over stdio and is meant to be spawned by an MCP client (Claude Code, Cursor, Windsurf, etc.) — don't run it manually in a terminal.
Claude Code
Register the server with the claude CLI instead of hand-editing JSON:
claude mcp add humanbound -- hb mcp # this project, just you (default)
claude mcp add -s user humanbound -- hb mcp # all your projects
claude mcp add -s project humanbound -- hb mcp # commit to repo (.mcp.json)
Scopes write to:
| Scope | Stored in | Visible to |
|---|---|---|
local (default) |
project .claude/settings.local.json |
only you, only this project |
user |
~/.claude.json |
only you, every project |
project |
.mcp.json (committed) |
anyone who clones the repo |
Verify with claude mcp list. Note: claude mcp only configures Claude Code — Claude Desktop reads a separate config.
Other clients
For IDEs/clients that take a JSON config directly:
If your client launches with a minimal PATH and can't find hb, replace "hb" with the absolute path from which hb.
Available Tools
| Tool | Description |
|---|---|
get_posture |
Current security posture score, grade, and trend data for the active project |
get_findings |
List findings with optional status/severity filters |
get_coverage |
Test coverage summary and untested categories |
get_experiments |
Recent experiments with status, results, and configuration |
get_logs |
Conversation logs and verdicts from a specific experiment |
get_campaigns |
Current ASCAM campaign status and phase information |
Example Queries
Once configured, you can ask your AI assistant questions like:
- "What's my current security posture score?"
- "Show me all critical findings that are still open"
- "What attack categories haven't been tested yet?"
- "Did the last experiment find any prompt injection vulnerabilities?"
- "What ASCAM phase is my project currently in?"
Requirements
Humanbound CLI must be installed and authenticated (hb login). The MCP server uses your existing CLI credentials and active project context.