Skip to main content
OpenClaw is an open-source local AI agent that connects to any LLM. Add a Spine tool config and OpenClaw will delegate research-heavy tasks to Spine automatically.

​
Install

Install OpenClaw if you don’t have it:
Drop a SKILL.md for Spine into OpenClaw’s user skills directory:
Verify OpenClaw discovered it:

Download skill file

spine.openclaw.md — drop into ~/.openclaw/skills/spine/
OpenClaw uses skills/, not tools/. The file must be named exactly SKILL.md inside a per-skill directory (~/.openclaw/skills/spine/SKILL.md). OpenClaw also accepts skills under ~/.openclaw/workspace/skills/<name>/SKILL.md if you want a workspace-scoped install.

​
Set your API key

Add SPINE_API_KEY to your environment or to OpenClaw’s .env file:
Create a key at platform.getspine.ai/keys.

​
What it does

The skill teaches the OpenClaw agent:
  1. When to delegate — multi-source research, competitive landscapes, reports, slide decks, spreadsheets, web apps
  2. The two-step pattern — POST /v1/run → poll GET /v1/run/{run_id} every 15–30s (runs average 10–20 min, up to 2 hr)
  3. How to pick a template — deep_research, report, slides, memo, excel, app, landing_page, or auto
  4. How to present results — final_output text + artifact download links

​
Headless invocation

For unattended use (CI, scheduled runs, scripts):
  • --local runs the embedded agent (no gateway required; needs a model-provider key such as OPENAI_API_KEY or ANTHROPIC_API_KEY in the shell).
  • --agent main (or --session-id …, or --to <E.164>) is required — otherwise openclaw agent exits with Pass --to <E.164>, --session-id, or --agent to choose a session.
  • OpenClaw loads skills lazily. If your prompt doesn’t clearly match the skill’s description, the agent may solve the task with its default tools instead. For deterministic routing, mention the capability by name (“using your Spine skill”) or prompt the agent to check its available skills first.

​
Example

For long-running production workflows, register a webhook instead of polling — see Webhooks. Polling loops inside the agent can exhaust turn budgets on runs longer than 5–10 minutes.

​
Templates

Pass the template that best fits the output: