patchbot

Bring-your-own-scanner, bring-your-own-feed vulnerability scanning, with Claude Code / Codex opening the fix PR.

inventory ──► feeds (match pkg@ver) ──┐
                                      ├──► findings ──► report (table/json/sarif)
scanners (trivy/grype/osv-scanner/sarif/any cmd) ──┘        └──► fix (agent → branch → PR)

• Inventory: built-in npm/pnpm/yarn/pip/go/cargo manifest+lockfile parsers, plus any CycloneDX SBOM (syft, cdxgen, or your own).
• Feeds: OSV.dev by default. Bring your own threat feed as an OSV-schema JSON file or URL — no custom format to write.
• Scanners: bring your own scanner. Built-in wrappers for trivy, grype, osv-scanner, or point command at any tool that emits trivy/grype/osv-scanner/SARIF JSON.
• Fix: claude or codex CLI bumps the vulnerable package, re-scans to confirm the advisory is gone, and opens a PR per package.

Install

pip install -e .

(Not yet published to PyPI — install from a local clone or a git URL for now.)

Quickstart

patchbot scan                          # table report, uses OSV by default
patchbot scan --format sarif -o out.sarif
patchbot fix --agent claude --dry-run  # print the fix prompt, make no changes
patchbot fix --agent claude --pr       # open PRs for real (needs `claude` + `gh` on PATH)

Config (patchbot.toml)

[inventory]
paths = ["."]
# sbom = "sbom.cdx.json"

[feeds.osv]
enabled = true

[feeds.internal]
type = "url"          # or "file" with `path = "..."`
url = "https://intel.example.com/osv-feed.json"

[scanners.trivy]
enabled = true

[scanners.mine]
type = "command"
cmd = "./scan.sh"
format = "sarif"       # or trivy | grype | osv-scanner

[report]
fail_on = "high"        # critical | high | medium | low | none
ignore = ["GHSA-xxxx-xxxx-xxxx"]

[fix]
agent = "claude"        # or "codex"
max_prs = 5
test_cmd = "npm test"

CLI flags override the config file.

Bring your own scanner

Any command that emits trivy, grype, osv-scanner, or SARIF JSON on stdout works:

[scanners.custom]
type = "command"
cmd = "my-scanner --json"
format = "sarif"

cmd runs via the shell using your config, so only point it at commands you trust.

Bring your own threat feed

Any URL or file serving OSV schema JSON (a bare array of vuln records, or {"vulns": [...]}):

[feeds.mine]
type = "file"
path = "./our-advisories.json"

Writing a plugin

Register a scanner, feed, or agent from your own pip package via entry points — no fork of patchbot required:

# your_package/pyproject.toml
[project.entry-points."patchbot.scanners"]
mytool = "your_package.scanner:run"          # def run(cwd: str, config: dict) -> list[Finding]

[project.entry-points."patchbot.feeds"]
myfeed = "your_package.feed"                 # module with match(packages, config) -> list[Finding]

[project.entry-points."patchbot.agents"]
myagent = "your_package.agent"               # module with run(prompt: str, cwd: str) -> int

patchbot plugins lists everything currently registered.

GitHub Actions

- uses: ./  # or a published action ref once this is published
  with:
    fail-on: high
    fix: false

See action.yml for all inputs, including fix: true to open auto-fix PRs (requires ANTHROPIC_API_KEY or OPENAI_API_KEY, and permissions: { contents: write, pull-requests: write }).