Models and providers

DeepCode is model-agnostic: connections are yours, defined once, switchable mid-conversation without losing history.

Connections

A connection is one route to a provider — a template plus your credentials:

deepcode provider set my-openrouter --template openrouter --api-key
deepcode provider set office-vllm  --template vllm --api-base http://gpu-box:8000/v1
deepcode provider list
deepcode provider test my-openrouter --model deepseek/deepseek-v4-pro

• --api-key prompts without echo; --api-key-env NAME reads an environment variable instead — the config file never holds a literal secret.
• deepcode provider models <id> --refresh fetches the connection’s live model directory.
• Removing is deepcode provider remove <id>.

Connections live under providers in ~/.deepcode/deepcode_config.json (named entries under providers.profiles; the classic per-vendor blocks still work). Project-level config deliberately cannot add connections — credentials stay user-scoped.

Declared models

When a gateway serves a model the catalog doesn’t know — or knows wrongly — declare it yourself:

"profiles": {
  "my-openrouter": {
    "template": "openrouter",
    "modelCatalog": "manual",
    "manualModels": [
      { "id": "deepseek/deepseek-v4-pro",
        "contextWindow": 1048576,
        "maxOutputTokens": 384000,
        "reasoningEfforts": ["off", "low", "medium", "high"] }
    ]
  }
}

A declaration is authoritative everywhere the model appears: the /model picker shows your numbers, and the runtime budgets context with them. Declare only what you need to correct — absent fields fall through to the built-in catalog.

Switching in the TUI

› /model

opens a picker over every configured connection’s full directory — type to filter, Tab switches connection scope, Shift+Tab cycles the highlighted model’s reasoning effort, Enter commits. Or be explicit:

› /model my-openrouter deepseek/deepseek-v4-pro

Switching affects future turns only — history is never rewritten, and an active turn finishes on the model it started with.

Reasoning effort

› /effort high     # or: auto · off · low · medium · high

Effort is a property of the model: the picker only offers levels the selected model actually supports (declared or catalog-known). auto lets the provider decide; off disables extended thinking where the model allows it.

While the model thinks, the status line shows the effort and the newest thought; Ctrl+O’s verbose mode prints full reasoning summaries when the provider exposes them.

What the turn footer tells you about cost

· 41s · 38.2k in · 512 out

Those are the provider’s own reported token counts, not estimates — the same numbers that drive automatic compaction. If a conversation compacts earlier or later than you expect, this footer is the ground truth to reason from.