Skip to main content

​
Agent overview

The Agent is a coding assistant exposed via an OpenAI-compatible API. It supports tools (e.g. read file, write file, list directory, run terminal, search and replace). The API returns tool calls; your client (e.g. IDE or CLI) executes them and sends the results back. Execution is client-side—the platform does not run code on your machine.

​
Capabilities

  • Chat — Multi-turn conversation with a supported model.
  • Tools — The model can request tool use. Supported tools typically include:
    • read_file — Read file contents (path parameter).
    • write_file — Write contents to a file.
    • list_directory — List directory contents.
    • run_terminal — Run a shell command.
    • search_replace — Search and replace in a file.
Tool names and parameters follow the schema you send in the request; the response includes tool_calls with the function name and arguments (e.g. JSON string).

​
How it works

  1. You send a request with messages and optionally a tools array (OpenAI-style tool definitions).
  2. The model may respond with text and/or tool_calls.
  3. Your client executes each tool call (e.g. read a file, run a command).
  4. You append the tool results as new messages and call the API again.
  5. Repeat until the model returns a final answer without tool calls.
The Agent does not execute tools on the server. It only returns tool calls; execution and file access happen in your environment.

​
Base URL

Use the production API base: 
https://api.llmtune.io/api/agent/v1
Example: list models for the agent: 
GET https://api.llmtune.io/api/agent/v1/models
Chat: 
POST https://api.llmtune.io/api/agent/v1/chat/completions

​
Authentication

Same as the rest of the API: include your API key in the request: 
Authorization: Bearer sk_live_YOUR_KEY
Balance is checked and usage is deducted per request the same way as for inference.

​
Model support

Only models that are enabled for inference and listed for the Agent can be used. Call GET .../api/agent/v1/models to see which models are available. Use one of those model IDs in the model field of chat requests.

​
Limitations

  • Client-side execution — You must implement tool execution in your client. The API does not run commands or read your files.
  • Model list — Only models returned by the Agent models endpoint are supported; others may return 400 or 404.
  • Rate limits and billing — Same rate limits and balance rules as inference; 402 when balance is insufficient.
For request/response examples, see Agent API examples.