ollamaClientAdapter
Points the OpenAI SDK at a local (or remote) Ollama instance and layers on the same tool-transform + coordinate-normalization pipeline that powers OpenRouter's computer-use integration.
Signature
ollamaClientAdapter({
ollama: OpenAI,
})
Install
npm install supercompat openai
Make sure Ollama is running and you've pulled the model you want:
ollama serve
ollama pull gemma4:e4b
Example
import OpenAI from 'openai'
import {
supercompat,
ollamaClientAdapter,
completionsRunAdapter,
memoryStorageAdapter,
} from 'supercompat/openai'
const ollama = new OpenAI({
apiKey: 'ollama', // Ollama accepts any non-empty value.
baseURL: 'http://localhost:11434/v1',
})
const client = supercompat({
clientAdapter: ollamaClientAdapter({ ollama }),
storageAdapter: memoryStorageAdapter(),
runAdapter: completionsRunAdapter(),
})
const response = await client.responses.create({
model: 'gemma4:e4b',
input: 'Hello.',
})
Remote Ollama
Swap the base URL to reach a different machine:
const ollama = new OpenAI({
apiKey: 'ollama',
baseURL: 'http://gpu-box.local:11434/v1',
})
Computer use
The adapter intercepts
computer_use_preview tools, rewrites them as a regular computer_call function tool, and post-processes model output — denormalizing 0-1000 normalized coordinates (Gemma 4, GLM-4.6v) back to pixels, fuzzy-extracting malformed JSON (Qwen), and relaying screenshots through a synthetic user message (Ollama drops tool-role images). See Computer use → Ollama.Per-model quirks are matched by prefix on the Ollama model id:
gemma4, gemma3 — 0-1000 normalized coords, box_2d fuzzy fallbackglm-4.6v, glm4.6v — 0-1000 normalized coords, native tool-format artifact cleanupqwen — fuzzy extraction for malformed JSONkimi-k2.5, kimi2.5 — rescale from 1920×1080 reference or 0.0-1.0 relative coords