Agent skill installation
Install safely into Codex, Claude Code, OpenClaw or a workspace and run model workflows with your own key.
Updated
Before you start
- Python 3.10+, a host supporting local skills, and your own LINK42 API key.
Choose a host and install
Open the Skill page
Select a host on /skill and copy its checksum-verified command. Domestic and international installation use the same package.
Check the install location
Codex: ~/.agents/skills/link42-api; Claude Code: ~/.claude/skills/link42-api; OpenClaw: ~/.openclaw/skills/link42-api; workspace: .agents/skills/link42-api. Global configuration is untouched; updates keep a backup.
Set a private environment variable
Set LINK42_API_KEY in the agent's secure environment and reload skills. Never put the key in prompts, URLs, command arguments or commits.
Discover, quote, execute
The skill joins your current /v1/models scope with public options. Ambiguous names preserve versions: V4.1 and V4 are not interchangeable. Select supported resolution, ratio, duration and input mode before calling.
CLI generation requires --execute. An estimate is not a settlement ceiling; use key budgets for hard limits. Save task and idempotency IDs, and resume timed-out tasks instead of creating paid duplicates.
python3 scripts/link42_agent.py discover
python3 scripts/link42_agent.py estimate --json quote.json
python3 scripts/link42_agent.py image-create --json image.json --execute --max-estimate-usd 0.10
python3 scripts/link42_agent.py image-wait --id TASK_IDScope and recovery
Beyond individual operations, doctor checks scope, plan resolves models and validates known options before quoting, and workflow-plan totals estimates for 1–8 explicit stages. Plans neither submit tasks nor fabricate future assets. Installed workflows.md and request-examples.md cover conversation/tool-result continuation and reference workflows.
| Capability | Operations | Boundary |
|---|---|---|
| Text / Vision | chat, responses, messages, count-tokens, gemini, embeddings | Model support determines tools, reasoning and input types. |
| Image | image, image-edit, image-create/get/wait | Reference edits use JSON, not multipart. |
| Video | video-create/get/list/wait/cancel | Roles/options come from the model; cancellation requires intent. |
| Assets / Usage | asset-upload/list/get/delete/complete, usage-report/logs/csv, score | Owned resources only; usage access must be enabled. |
- Audio, Realtime, Batch and hosted agents are unsupported. No admin authority is granted; marketplaces are not yet published. Failures retain sanitized codes/request/task IDs, never credentials.