Models and providers

Connect your preferred model provider, then choose a model for each conversation. You can save several connections—for example, a personal API account and a local model server—and switch between them in TUI, Desktop, or Web.

Configure a connection

To add a connection in Desktop/Web, open Settings → AI providers and choose Add provider. Fill in your provider and API key, then click Save and check.

From the terminal, this example adds OpenRouter under the name my-openrouter and lists its models:

deepcode provider set my-openrouter --template openrouter --api-key
deepcode provider models my-openrouter --refresh
deepcode provider list

Enter your API key when prompted; it will stay hidden as you type. If you keep the key in an environment variable, use --api-key-env NAME instead, replacing NAME with the variable’s name. Your saved connections are available across projects on this computer.

Replace MODEL_ID with an exact ID in this connection’s catalog:

deepcode provider test my-openrouter --model MODEL_ID

This sends a short request to check that the model responds. Desktop/Web offer the same check through Save and verify model.

When trying a new model server, you can also check streaming and tool calls:

deepcode provider test my-openrouter --model MODEL_ID --agent

This asks the model to use a small local test tool; it does not run shell commands or edit your files. Read the result for each check. If one is skipped, that capability has not been tested. These checks use your provider’s API quota.

Select a model, then switch future turns

Use /model to open the TUI picker or specify both connection and model:

/model my-openrouter MODEL_ID

At launch, use deepcode --connection my-openrouter --model MODEL_ID with your workspace/trust options. Desktop/Web use the composer model picker. Configuring or testing a connection does not select it for every Session.

You can switch models and keep the conversation. The new selection applies to messages you send afterward; work already running or queued keeps the model it started with.

Explicit protocols and local services

If you use a local server or custom gateway, set its address and API protocol. Choose openai_chat, openai_responses, or anthropic_messages to match the API your server provides. For standard providers, the template’s auto setting usually handles this for you.

When you explicitly select openai_responses, your server must support the Responses API; DeepCode will keep using that protocol if a request fails.

For a local OpenAI Chat-compatible service without authentication, replace LOCAL_MODEL_ID and the example endpoint with your deployment’s values:

deepcode provider set local-vllm --template vllm --api-base http://127.0.0.1:8000/v1 --protocol openai_chat --auth none --catalog manual --model LOCAL_MODEL_ID
deepcode provider test local-vllm --model LOCAL_MODEL_ID --agent

Use --auth none only when that endpoint does not require authentication. For an authenticated gateway, use --auth api_key with --api-key or --api-key-env. For additional gateway settings or OpenRouter browser login, see the Provider configuration reference.

Amazon Bedrock

Bedrock’s OpenAI-compatible endpoint is region-specific. Supply its /openai/v1 base URL and expose an Amazon Bedrock API key through the standard AWS_BEARER_TOKEN_BEDROCK environment variable:

deepcode provider set work-bedrock --template bedrock \
  --api-base https://bedrock-runtime.us-east-1.amazonaws.com/openai/v1 \
  --api-key-env AWS_BEARER_TOKEN_BEDROCK
deepcode provider models work-bedrock --refresh
deepcode provider test work-bedrock --model us.anthropic.claude-sonnet-4-6

Replace us-east-1 with the Region that serves your account and choose a model ID returned by the refresh command. Bedrock may return an in-Region ID such as anthropic.claude-sonnet-4-6, a geography-scoped ID such as us.anthropic.claude-sonnet-4-6 or eu.anthropic.claude-sonnet-4-6, or a global. ID depending on model and Region.

This first integration sends the configured value as a bearer token through Bedrock’s OpenAI-compatible Chat Completions API. It does not implement AWS SigV4 signing, IAM role/profile discovery, SSO, or the native Converse API.

Declared models and capacities

If your server’s model is missing from the list, add it manually. You can also set its context window and output limit to match your deployment.

Open ~/.deepcode/deepcode_config.json and add the following connection under providers.profiles, keeping your other settings:

{
  "local-vllm": {
    "template": "vllm",
    "apiBase": "http://127.0.0.1:8000/v1",
    "protocol": "openai_chat",
    "auth": "none",
    "modelCatalog": "manual",
    "manualModels": [
      {
        "id": "LOCAL_MODEL_ID",
        "contextWindow": 65536,
        "maxOutputTokens": 4096
      }
    ]
  }
}

Replace LOCAL_MODEL_ID with your server’s model ID and the example limits with those supported by your deployment. Use these settings to tell DeepCode how much context and output the model can handle.

Reasoning effort

For models that offer reasoning levels, use /effort to choose how much thinking to request. For example, if your model supports high:

/effort high

Use a level offered for the selected model; /effort auto leaves the choice to the provider, and /effort off disables extended thinking where supported. Choose from the levels shown for your model. Use Ctrl+O separately when you want to change how much activity the TUI displays.

Context window cap

To keep conversations within a smaller context budget, set a cap such as 64k. Use auto to return to the model’s known window:

/context 64k
/context auto

The first narrows the Session’s context budget; the second follows the selected model’s known window. The cap must leave room for the generation limit and cannot expand the model’s declared/catalog window. Desktop/Web expose the same setting in the model picker. A smaller budget can trigger compaction sooner.

Usage and troubleshooting

After a response, the footer shows elapsed time and token usage when your provider reports it. Check your provider’s billing page for charges. If a model fails to connect or use tools, follow Troubleshooting.