Ollima in VS Code

Use the Ollima catalog as the models behind your AI coding assistant. Because the gateway is OpenAI-compatible, any VS Code extension that takes a custom OpenAI-shape endpoint works — point it at Ollima, drop in your key, and code against kimi-2.7 and friends.

Install

The Ollima catalog works in any VS Code extension that supports a custom OpenAI-compatible endpoint. Pick one and install it from the Extensions panel (Ctrl+Shift+X / ⇧⌘X):

• Continue, Cline, or Roo Code — all on the VS Code Marketplace. Any of them works; the rest of this page uses Continue's config format.

Then add your key and the models as shown below. A preconfigured Ollima extension with this catalog built in is on the way — until it lands, use any of the above.

How it works

Every extension needs the same four values. The field names differ; the values don't:

Mint a key

Sign in at app.ollima.com, open the Keys page, click Create key, and copy the sk-t0-… value once — we never show it again. Use one key per machine so a lost laptop is a one-line revoke. Set a monthly spend cap while you're there.

Configure the models

For Continue-based extensions (including the Ollima build), models live in a YAML config. Click the gear icon in the chat panel to open it, or edit it directly:

• Windows: %USERPROFILE%\.continue\config.yaml
• macOS / Linux: ~/.continue/config.yaml

Paste the catalog — replace sk-t0-… with your real key in each entry:

# ~/.continue/config.yaml
models:
  - name: Kimi K2              # default — agentic coder (chat + agent)
    provider: openai
    model: kimi-2.7
    apiBase: https://api.ollima.com/v1/
    apiKey: sk-t0-…
    roles: [chat, edit, apply]
    capabilities: [tool_use]     # required for Agent mode — see below
    defaultCompletionOptions: { maxTokens: 8192 }   # reasoning model — give it headroom

  - name: Qwen3 Coder Next     # fill-in-middle autocomplete
    provider: openai
    model: qwen3-coder-next
    apiBase: https://api.ollima.com/v1/
    apiKey: sk-t0-…
    roles: [autocomplete]

  - name: DeepSeek V4          # flagship coder, 1M context
    provider: openai
    model: deepseek-v4
    apiBase: https://api.ollima.com/v1/
    apiKey: sk-t0-…
    roles: [chat, edit]
    capabilities: [tool_use]

Save, and the models appear in the chat model dropdown immediately — no restart. Pick Kimi K2 to start.

Models & roles

Roles tell the extension which model to use for each job: chat (the conversation panel), autocomplete (inline ghost-text), edit (in-place rewrites), and apply (applying a suggested diff).

Strongest coding picks

Beyond the core models above, the catalog has a dedicated coding lane. The current open-weights benchmark leaders are deepseek-v4 (80.6% SWE-bench Verified) and minimax-m3 (80.5%) — both priced under $3 / 1M output. Drop these into the same config.yaml:

# strong coders — add alongside the models above
  - name: DeepSeek V4          # flagship coder, 1M context
    provider: openai
    model: deepseek-v4
    apiBase: https://api.ollima.com/v1/
    apiKey: sk-t0-…
    roles: [chat, edit, apply]

  - name: DeepSeek V4 Flash    # cheapest strong coder — great for autocomplete
    provider: openai
    model: deepseek-v4-flash
    apiBase: https://api.ollima.com/v1/
    apiKey: sk-t0-…
    roles: [autocomplete, chat]

  - name: MiMo v2.5 Pro        # agentic coding, long context
    provider: openai
    model: mimo-v2
    apiBase: https://api.ollima.com/v1/
    apiKey: sk-t0-…
    roles: [chat, edit, apply]

Full list with INR pricing and context windows is at /models; deeper guidance at Choosing a model.

Agent mode

In Agent mode the model doesn't just talk — it reads your files, creates new ones, edits code, and runs terminal commands by calling tools. To use it: switch the mode dropdown at the bottom of the chat panel from Chat to Agent, open a folder (the agent needs a workspace to write into), then ask for what you want — e.g. "create a FastAPI project with a health endpoint and a Dockerfile". It will scaffold the files for you.

By default every file-write or command pauses for your approval — you click Accept on the tool card before it runs. To skip that for a given tool, open Tools settings (the tools icon in the chat toolbar, or the ⚙ gear → Tools) and change its policy:

Verify it routed

Don't trust that text merely appeared — confirm it went through Ollima. Send a prompt, then open app.ollima.com → Logs. The inference shows up with the model name and token counts. That's proof the request hit our gateway and not some other provider the editor had configured.

Other editors

The same four values work in any extension with a custom OpenAI-compatible provider:

• Cline / Roo Code / Kilo Code — settings → API Provider → OpenAI Compatible → Base URL https://api.ollima.com/v1/, paste the key, type the model ID.
• GitHub Copilot Chat (BYOK) — model picker → Manage Models → an OpenAI-compatible option, then the same base URL, key, and model ID. Availability of a custom base-URL field depends on your VS Code + Copilot version; if you only see fixed named providers, use a Continue/Cline-style extension instead.

Local models (Ollama) — free & private

Want fully-free, fully-private inference with no key at all? Run models locally with Ollama and point the extension at it. Because this runs on your machine, it's configured directly in the editor — not through the Ollima cloud gateway (a cloud service can't reach your localhost).

1. Install Ollama and pull a coding model: ollama pull qwen2.5-coder.
2. Add it to the same config.yaml with the ollama provider:

# local — no apiKey, runs on your machine
  - name: Qwen2.5 Coder (local)
    provider: ollama
    model: qwen2.5-coder
    apiBase: http://localhost:11434

Mix and match: keep Ollima cloud models for the heavy lifting and a local Ollama model for offline or sensitive work — both appear in the same model picker. Local runs are free and never leave your machine, but quality and speed depend on your hardware.

Troubleshooting