Skip to content

Coding Agents

A coding agent can use a registered LGOS graph when the agent's OpenAI wire requests stay inside LGOS's supported Responses or Chat Completions subset. The agent owns client tools: it sends function definitions, receives calls, executes them locally, and returns matching outputs.

The demo simple-graph-external-tools is the smallest tool-enabled graph. It forwards client-provided tools to the upstream model and does not execute them.

Endpoint And Model

Connect directly to LGOS and use the exact model ID from GET /v1/models:

Base URL Model ID
https://lgos.example.com/v1 simple-graph-external-tools

An optional gateway may use a provider-qualified routing ID, but it must pass the native contract tests in the proxy guide. The pinned Bifrost demo's native Responses route preserves the tested data-plane contract, including phase; its normalized model-detail and error metadata remain lossy. The raw OpenAI pass-through route passes the complete tested subset.

Client Compatibility

LGOS implements a deliberately bounded Responses surface. A client that always sends hosted tools, reasoning configuration, include, prompt-cache options, or another unsupported field will receive an explicit OpenAI invalid_request_error. Do not place a Responses-to-Chat translator in front of LGOS to hide that mismatch.

Codex custom providers use the Responses wire API. Direct Codex compatibility has not been verified against LGOS's bounded subset, and Codex can configure Responses controls that LGOS does not accept. The configuration shape is documented by Codex custom model providers; add a runnable example here only after a direct integration test passes without request rewriting.

The examples below intentionally exercise the direct Chat compatibility route. They are demonstrations for clients whose provider adapters use Chat Completions; the maintained Chainlit and Open WebUI demos remain Responses-only.

OpenCode's official provider guide assigns @ai-sdk/openai-compatible to /v1/chat/completions providers. Add a direct LGOS provider to opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "lgos": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Direct LGOS",
      "options": {
        "baseURL": "https://lgos.example.com/v1",
        "apiKey": "DUMMY"
      },
      "models": {
        "simple-graph-external-tools": {
          "name": "LGOS external tools"
        }
      }
    }
  }
}

Start OpenCode, run /models, and select LGOS external tools. See OpenCode's custom provider documentation.

pi supports both OpenAI APIs. This compatibility example selects its openai-completions adapter in ~/.pi/agent/models.json:

{
  "providers": {
    "lgos": {
      "baseUrl": "https://lgos.example.com/v1",
      "api": "openai-completions",
      "apiKey": "DUMMY",
      "compat": {
        "supportsReasoningEffort": false
      },
      "models": [
        {
          "id": "simple-graph-external-tools",
          "name": "LGOS external tools"
        }
      ]
    }
  }
}

pi --provider lgos --model simple-graph-external-tools

DUMMY makes a keyless development model appear in pi's selector. Use a real deployment credential when the LGOS host enforces authentication. See pi's custom model documentation.

Other Agents

For a Responses client, compare its emitted request with the exact supported subset and verify streaming plus function continuation before adopting it. For a Chat Completions client, use the direct base URL and registered graph name. No LGOS-specific SDK or agent plugin is required.