Build OpenAI-compatible LLM agents on Runware that can reason, call your functions, and answer with grounded tool results.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install llm-agent
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install llm-agent using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is LLM Agent?
Openclaw Skills for LLM Agent turns Runware-hosted models into chat assistants and tool-using agents that can reason across multiple steps, call your APIs, and incorporate the returned JSON into the final answer. It uses the standard OpenAI chat-completions and tool-calling pattern, so existing OpenAI client code and agent frameworks can connect to Runware with minimal changes.
This skill is designed for developers who need a model that can interact with their own systems without direct system access. You define the tools, constrain the schema, choose a live AIR model, and let the agent loop until it stops calling tools. The result is a reliable Openclaw Skills workflow for database lookups, search, orchestration, and other multi-step tasks.
LLM Agent Use Cases
- Building a chat assistant that can call internal APIs, search services, or database endpoints.
- Letting a model gather context before acting, or act immediately based on the system prompt policy.
- Multi-step reasoning tasks that require reading tool results and making another decision.
- High-volume or latency-sensitive agents that should route to faster models like Claude Sonnet 4.6.
- Long-context, tool-rich workflows that benefit from Gemini 3.1 Pro, Claude Opus 4.8, or Qwen3.5-397B.
- Migrating an existing OpenAI-compatible agent to Runware-hosted models without rewriting the loop.
How LLM Agent Works
- Gather the job, the tools, the system prompt, and any constraints such as latency, cost, or context-window needs.
- Point an OpenAI-compatible client at
https://api.runware.ai/v1, choose a live AIR model or slug, and send the conversation plus thetoolsarray. - Let the model return either text or
tool_calls. When tool calls appear, parsefunction.argumentsas a JSON string and execute the corresponding functions. - Append the assistant message and one
toolmessage per call, making sure eachtool_call_idmatches the callid. If multiple tool calls arrive in one turn, process them all and continue the loop. - Repeat until
finish_reasonisstop, then return the final answer. Keep the loop bounded with an iteration cap so the agent cannot run forever. - Route by sub-task when needed: use stronger reasoning models for complex planning and faster tiers for cheap or high-volume calls.
LLM Agent Setup
- Export your Runware key and OpenAI-compatible endpoint.
export RUNWARE_API_KEY="your-runware-api-key"
export OPENAI_BASE_URL="https://api.runware.ai/v1"
export OPENAI_MODEL="google:[email protected]"
- Verify the model is live and inspect its schema before production use.
runware-models
runware-run
- Wire any OpenAI-compatible SDK to
https://api.runware.ai/v1and setmodelto a live AIR or slug. - Define each function tool with a strict JSON Schema: use
required,enum, and narrow argument shapes. - Keep the system prompt and tool list stable across turns to improve prompt caching.
- Implement a bounded tool loop that parses
function.arguments, executes the function, and appends a matchingtoolmessage. - Return tool failures as JSON objects like
{ error: ... }so the model can recover instead of crashing the loop.
LLM Agent Data Schema & Taxonomy
Openclaw Skills for LLM Agent is request-scoped; it does not require generated project files by default. The runtime data model is a chat-completions envelope with tools and tool results.
| Entity | Shape | Purpose |
|---|---|---|
model |
AIR / slug + live status | Selects the Runware-hosted LLM and confirms it is runnable. |
messages |
Array of system, user, assistant, and tool messages |
Carries the conversation and preserves every tool round-trip. |
tools |
Array of { type: function, function: { name, description, parameters } } |
Exposes tightly scoped tool schemas to the model. |
tool_calls |
Array of calls with id, type, function.name, and JSON-string function.arguments |
Represents every requested function invocation. |
tool message |
{ role: tool, tool_call_id, content } |
Returns compact JSON or { error: ... } back to the model. |
| Control metadata | tool_choice, temperature, max_completion_tokens, iteration cap |
Governs autonomy, determinism, and loop safety. |
| Usage metadata | finish_reason, usage.prompt_tokens_details.cached_tokens |
Tracks loop termination and prompt-cache reuse. |
Recommended taxonomy: prompt, tools, loop_state, tool_results, model_selection, and usage. Persist transcripts as JSONL or request/response logs only if you need audits or replay.
LLM Agent Advanced Features
- OpenAI-compatible drop-in support for existing OpenAI agent code and frameworks.
- Parallel tool calls in a single assistant response, with independent
tool_call_idmatching. tool_choicerouting forauto,none,required, or a named function when the action is predetermined.- Prompt-cache-friendly execution by keeping the system prompt and tool definitions as a stable prefix.
- Model routing by sub-task, letting you pick stronger reasoning or lower latency without changing the loop.
- Error-aware tool execution that returns
{ error: ... }instead of throwing, so the agent can recover. - Bounded autonomy with explicit iteration caps to prevent open-ended loops.
- Live model validation via AIR,
runware-models, andrunware-runbefore shipping.
SKILL.md
Loading
Related Openclaw Skills
Logos and Vectors
Openclaw Skills creates real SVG logos or converts raster images into editable vector art that stays crisp at any size.
Photoreal Stills
An advanced AI generation skill designed to defeat synthetic rendering artifacts by prompting for optical behaviors, real-world camera properties, and intentional imperfections.
Reframe Video
Openclaw Skills reframe videos to new aspect ratios by extending the scene instead of cropping the subject.
Runware Model Selector
Openclaw Skills helps agents pick the right live Runware model by matching the task outcome to verified capabilities, status, and price.
Image to 3D Asset
Openclaw Skills converts photos or prompts into production-ready GLB 3D assets with controllable topology, materials, and polygon budgets.
2D Game Assets
Openclaw Skills for 2D game assets generates cohesive sprite sets, icons, stickers, and transparent cutouts that all share one visual style.
Featured*
