shinthink/blitzstrike

What it is

Blitz Strike (shinthink/blitzstrike) is an MCP server exposing run_engagement as a single entry point that any MCP-capable agent (Claude Code, Cursor, Gemini, Hermes, OpenCode, Windsurf, Copilot, Cline) can call to drive a full pentest. The page adds concrete setup/operational detail the existing summary doesn't cover.

How it works

The LLM classifies the target automatically — a URL routes to the live testing pipeline, a filesystem path routes to the source-analysis pipeline — then the agent is guided through recon → analyze → verify → review → report by bundled doctrine (instructions, skills, per-step next_steps). Tool manuals are auto-attached: tool_lookup(name) pulls a tool's manual, and run_engagement attaches the manual tied to each matched escalation chain. Findings persist as append-only JSONL at ~/.blitzstrike/memory.jsonl, auto-capturing matched chains (deduped by chain id) so the knowledge base compounds across engagements; only entries marked verified=true are treated as authoritative.

Preconditions / Affected

Requires Bun runtime (TypeScript/Bun project). Distributed as a single compiled binary (bun build --compile) or via zero-install npx/bunx. blitzstrike install auto-detects installed agent CLIs and merges config into existing files (e.g., ~/.claude.json, ~/.cursor/mcp.json, project .mcp.json) without overwriting other MCP servers.

What to do

Install and register with detected agents:

npx -y blitzstrike doctor
npx -y blitzstrike install

Or run from source:

git clone https://github.com/shinthink/blitzstrike.git
cd blitzstrike
bun install
bun run src/index.ts serve --mcp

Then invoke via natural language ("audit ./src" or "audit https://example.com") after restarting the agent and calling run_engagement. Other useful commands: blitzstrike doctor (health check), blitzstrike install --dry-run (preview config), blitzstrike sync-data (fetch payload/template datasets on demand), blitzstrike update.

To extend the tool's knowledge base, edit chains.json directly rather than hardcoding escalation logic in source.

Caveats

The page states explicit hypothesis-guardrails the tool enforces, worth noting since they define what "verified" means here: a dangerous function in the same file as an unauth handler is not proof the handler calls it (use eagle_eye to confirm reachability); a writable file is not RCE unless the runtime loads it (CF-003); and because the default Apache FilesMatch regex is anchored with $, a filename like .php.jpg will not execute under default config. The page frames every finding as a hypothesis until strike_verify reflects a live marker and the chain's negative_control stays inert. Tool is stated as for educational/authorized security research only.