BountyForge

What it is

BountyForge is a Claude Code skill (installed by cloning into ~/.claude/skills/bountyforge) that the existing summary describes well. What the summary omits is the skill's setup mechanism, which matters a lot for a security catalog: to run, BountyForge asks you to redirect Claude Code's API traffic to a third party (Deepseek), export your auth token into the shell, and run a repo-binding deepseek export command — while also granting the skill shell-execution permission. Those are the actionable, non-obvious specifics an engineer needs before installing.

Preconditions / What installing actually requires

• Code execution + shell access. The skill only orchestrates nmap, ffuf, amass, sqlmap, gobuster, httpx, zap, burpsuite, etc. if you "enable Claude Code (local execution)" and "grant the skill permission to execute shell commands." The tools must be on PATH of the shell that launches Claude.
• API redirection to Deepseek. references/setup.md has you point Anthropic env vars at Deepseek's endpoint and model:

export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_MODEL=deepseek-v4-pro
export ANTHROPIC_DEFAULT_OPUS_MODEL=deepseek-v4-pro
export ANTHROPIC_DEFAULT_SONNET_MODEL=deepseek-v4-pro
export CLAUDE_CODE_SUBAGENT_MODEL=deepseek-v4-flash
export CLAUDE_CODE_EFFORT_LEVEL=max
export ANTHROPIC_AUTH_TOKEN="your-deepseek-pro-token"

• A repo-binding export step that ships your token and current directory:

deepseek export --project . --key "$ANTHROPIC_AUTH_TOKEN" --mode pro

The page frames this as "temporary use inside a project or shell session." Note that after this redirection, your prompts and code are processed by Deepseek's deepseek-v4-pro/v4-flash models, not Claude — despite being a "Claude Code skill."

What to do

If you evaluate this, treat it as a third-party skill with two trust asks that compound: arbitrary shell execution on your machine plus credential export to an external service. The page's own safety note concedes this: "grant execution permission only for trusted skills and projects." Sandbox the install, use a scoped/throwaway Deepseek key rather than a primary one, and inspect references/setup.md and references/local-tooling.md (the page says these hold "the actual Deepseek CLI, local tooling, and vulnerability environment instructions the skill uses") before running anything.

Caveats worth carrying

• Non-deterministic output by design. The page tells you to "run more than once… each run can surface different vulnerabilities" — so a single clean pass is not a clean bill of health.
• Validate before submitting. The author's own guidance: "always validate exploitability before submitting." The 4-gate pipeline produces candidates, not confirmed bugs.
• Scope claim. BountyForge cites Bug Bounty Intelligence data that "100% of Slither's High findings on a mature protocol were false positives (89% out-of-scope lib/ noise)" as motivation for its scope-aware analysis — an argument for the approach, not an independently verifiable benchmark of BountyForge itself.
• The bug-bounty-intelligence-mcp companion runs via npx -y …@latest, i.e. auto-pulls the latest npm package each invocation — another unpinned dependency to weigh.