| name | claw-advisor |
| description | Run a long-running or token-heavy agent task at ~10× lower cost using Claude's Advisor Tool — Haiku/Sonnet executes, Opus advises only at strategic checkpoints. Use when an OpenClaw cron, batch job, or background routine is too expensive on Opus alone but loses quality on a smaller model. |
claw-advisor — Skill for OpenClaw
This skill lets the agent delegate token-heavy or long-horizon tasks (mailbox triage, repo audits, daily briefings, log analysis) to a cheaper executor model with selective Opus advice.
When to use
- A cron task is repeating ≥1×/hour and consuming significant tokens.
- A task is mostly mechanical (collect data → format → flag highlights) but quality drops sharply on Haiku alone.
- The user has an Anthropic API key and accepts beta tooling (
advisor-tool-2026-03-01).
Do NOT use for direct chat replies, single-turn Q&A, or tasks the user explicitly asks Opus to handle.
How to invoke
ANTHROPIC_API_KEY="$ANTHROPIC_API_KEY" \
claw-advisor run /path/to/job.yaml > /tmp/result.md
The CLI logs cost telemetry to stderr; the result goes to stdout.
Job file format
name: <job-id>
executor: claude-haiku-4-5 | claude-sonnet-4-6 | claude-opus-4-6
advisor: claude-opus-4-7
max_advisor_uses: <int, default 5>
max_tokens: <int, default 4096>
system: |
<multi-line system prompt>
task: |
<multi-line user task; can use {{ placeholders }} populated by caller>
Cost expectations (per run)
| Job size | Pure Opus | Haiku + Advisor | Sonnet + Advisor |
|---|
| 5–10K tokens (briefing) | $0.30–0.60 | $0.03–0.08 | $0.08–0.15 |
| 30K tokens (repo sweep) | $1.50–3.00 | $0.15–0.30 | $0.40–0.80 |
| 100K tokens (log audit) | $5.00+ | $0.40–0.80 | $1.20–2.50 |
Always check the printed Total cost line after the first run to validate.
Failure modes
- Beta header rejected: API may have moved past
advisor-tool-2026-03-01. Update claw-advisor (npm update -g claw-advisor) and check release notes.
- Advisor returns
overloaded: the executor continues without advice. Quality may drop, but the job completes. Monitor advisor call success rate.
- Cost spike: check if executor is getting stuck in tool-call loops. Add
max_tokens cap to job file.
Reference
