Clawscan Cli

ClawScan CLI

Overview

Use ClawScan to run one or more security scanners against agent skills, preserve raw scanner evidence, and optionally pass the evidence to an external judge harness. The public CLI is general-purpose; ClawHub-specific parity helpers belong outside cmd/clawscan. Keep secrets in environment variables, never CLI flags.

Command Surface

Run from the repo root during development:

bashCopy code
go run ./cmd/clawscan <target> --scanner <scanner-id> [flags]

Use the installed binary when available:

bashCopy code
clawscan <target> --scanner <scanner-id> [flags]

Choose the run mode:

Helpful metadata:

bashCopy code
clawscan --help
clawscan -h
clawscan --version
clawscan scanners
clawscan profiles
clawscan benchmark list

Unless --json is passed, ClawScan writes the full artifact to ./clawscan-results/artifact.json by default, preserves per-scanner JSON files in the same visible results bundle, and prints a concise key/value summary ending in full_results: ./clawscan-results/artifact.json. Use --output <path> to choose a different artifact path; explicit .json paths keep that artifact file and write scanner JSON beside it.

Targets, Profiles, And Config

If no target is passed with --scanner, --profile, or --config, ClawScan scans child skill directories under ./skills. If ./skills is missing or contains no children with SKILL.md, it fails with a target-discovery error. Plain clawscan without --scanner, --profile, or --config is invalid. Benchmark runs use clawscan benchmark <benchmark-id> and do not accept scan targets.

Built-in profiles:

Profiles are loaded from embedded built-ins. Project .clawscan.yml / .clawscan.yaml files are NOT loaded automatically: pass --config <path> to load a specific config, or --discover-config to load the nearest one found upward from the current directory. A project profile with the same name shadows the built-in whole profile. --config <path> without --profile runs every profile in that config and emits a clawscan-batch-v1 artifact. Discovery is off by default because project configs can define scanner commands that execute with the caller's environment and credentials.

Use clawscan profiles to inspect the built-in profile catalog. Use clawscan profiles -v to print it as pasteable YAML.

CLI flags override the selected profile for one run. Passing --scanner without --profile creates an ad hoc scanner-only run, so profile judges are not invoked accidentally.

Minimal config:

yamlCopy code
version: 1
sandbox:
  mode: docker
  env:
    - OPENAI_API_KEY
    - CODEX_API_KEY
profiles:
  review:
    scanners:
      - clawscan-static
    json: true
    judge:
      command: judge --out {{ output }}

sandbox.env is an allowlist of env var names to pass into the Docker runtime; store names there, not secret values. CLI equivalents are --sandbox, --sandbox-image, and repeatable --sandbox-env.

Scanners

Use clawscan scanners for the registry-backed scanner catalog and clawscan scanners <scanner-id> for one scanner's repository, description, env vars, and install guidance.

Accepted scanner IDs:

textCopy code
agentverus, aig, cisco, clawscan-static, skillspector, snyk, socket, virustotal

Credential rules:

Artifact env fields record only present or missing; they must never contain secret values. GenDigital/Gen Agent Trust Hub is not a built-in scanner because there is no local CLI for ClawScan to invoke.

Dependency setup:

bashCopy code
clawscan install aig cisco skillspector

clawscan install accepts one or more scanner IDs. It follows upstream scanner install docs where they publish an install command, including A.I.G's pip install aig-skill-scan, Cisco's uv pip install cisco-ai-skill-scanner, SkillSpector's uv tool install git+https://github.com/NVIDIA/skillspector.git, Socket's npm install -g socket, and AgentVerus' npm install --save-dev agentverus-scanner. Snyk is launcher-based, so ClawScan verifies uvx; built-in and simple API-backed scanners are skipped. The aig adapter runs aig-skill-scan --repo <target> --language en -o <result.sarif.json> and stores the SARIF 2.1.0 document as raw scanner evidence.

Starting in ClawScan v0.1.2, aig no longer uses the legacy A.I.G Docker/API service. Replace AIG_MODEL with DEFAULT_MODEL, AIG_MODEL_BASE_URL with DEFAULT_BASE_URL, and AIG_MODEL_API_KEY with LLM_API_KEY or OPENAI_API_KEY; AIG_BASE_URL and AIG_API_KEY are retired. The local scanner accepts directory targets only.

For normal runs, command-backed scanners and judges run in ghcr.io/openclaw/clawscan-runtime:latest. clawscan install is mainly for local development or --sandbox off environments.

Use --scanner-result when a test or fixture should supply stable scanner JSON:

bashCopy code
clawscan ./my-skill \
  --scanner skillspector \
  --scanner-result skillspector=./fixtures/skillspector.json \
  --json

The scanner must still be requested with --scanner or via the selected profile.

Benchmarks

Use clawscan benchmark list for the registry-backed benchmark catalog.

Supported benchmarks:

textCopy code
cuhk-zhuque/SkillTrustBench
clawhub-security-signals

Run SkillTrustBench with the canonical Hugging Face ID or the short alias SkillTrustBench:

bashCopy code
clawscan benchmark SkillTrustBench \
  --limit 10 \
  --scanner clawscan-static \
  --output /tmp/clawscan-benchmark.json

SkillTrustBench uses split benchmark. The first live run downloads and caches benchmark_full_v1.0.zip, then extracts only the requested case directories into temporary scan targets.

Use --ids <path-or-url> with SkillTrustBench to run a fixed subset from a plain text file with one ID per line or JSONL rows with an id field. The source is streamed and may contain at most 5,520 unique IDs, at most 256 bytes per trimmed ID, and at most 256 KiB (262,144 bytes) of retained ID text. These are selection limits, not a total source-size limit; JSONL sources may exceed 256 KiB. Individual lines must be smaller than the parser's 1 MiB buffer limit. --ids preserves source order, records idsSource, idsCount, and idsSha256 in the artifact, and is mutually exclusive with --limit and --offset.

ClawHub Security Signals splits: train, validation, test, eval_holdout. --limit 0 means run the full selected split. Use --offset with --limit for reproducible chunks.

For clawhub-security-signals, --output ./clawscan-benchmark.json also writes ./predictions.jsonl. Use --predictions-output <path> to choose another path. --predictions-output is only supported for clawhub-security-signals.

Benchmark artifacts use clawscan-benchmark-v1. Each case embeds the normal clawscan-run-v1 artifact, expected verdict metadata, and evaluation status. The summary includes case counts, scanner/judge statuses, and accuracy over scored cases. Inspect the JSON artifact for full evidence.

Judge Harness

--judge runs through the platform shell and must produce a JSON object on stdout or at {{ output }}.

Important placeholders:

Prompt files can reference requested scanner JSON:

mdCopy code
```json
{{ scanners.skillspector }}
```

If a prompt references an unrequested scanner, ClawScan should fail clearly. The built-in clawhub profile uses this same judge mechanism with embedded prompt and output-schema files.

Verification

Use static scanner smokes for local proof because they do not need secrets:

bashCopy code
go run ./cmd/clawscan ./README.md --scanner clawscan-static --output /tmp/clawscan-smoke.json
go run ./cmd/clawscan benchmark SkillTrustBench --limit 1 --scanner clawscan-static --output /tmp/clawscan-benchmark-smoke.json
go run ./cmd/clawscan --help
go test -count=1 ./...
go vet ./...

Common Mistakes

• Do not pass API keys as CLI flags.
• Do not run plain clawscan; choose --scanner, --profile, --config, or clawscan benchmark <benchmark-id>.
• Do not assume clawscan scans .; no target means discover ./skills.
• Do not expect a profile judge when passing explicit --scanner flags without --profile.
• Do not use benchmark flags such as --split, --limit, or --offset outside clawscan benchmark <benchmark-id>.
• Do not use --config without --profile for benchmark runs; all-profile config runs are target scans only.
• Do not assume host-installed scanner CLIs are used by default; command-backed scanners and judges use the Docker runtime unless --sandbox off is set.
• Do not add unsupported dataset names; built-ins are SkillTrustBench and clawhub-security-signals.
• Do not assume scanner failures are final policy verdicts; scanner output is raw evidence for comparison or judging.