Providers

deepcode ships with three built-in backends (DeepSeek / GLM / Kimi·Moonshot), plus an OpenAI-compatible custom slot that can point at any self-hosted or third-party OpenAI-compatible endpoint. All four share the same main loop, tools, and permission system — switching backends just swaps the key and baseURL, everything else works the same.

The four backends

provider	env key	baseURL	smart (main loop)	fast (sub-operations)
deepseek (default)	DEEPSEEK_API_KEY	https://api.deepseek.com	deepseek-v4-pro	deepseek-v4-flash
glm	ZHIPUAI_API_KEY	https://open.bigmodel.cn/api/paas/v4	glm-5.2	glm-5-turbo
kimi	MOONSHOT_API_KEY	https://api.moonshot.cn/v1	kimi-k2.7-code (thinking-only)	kimi-k2.5
custom	user-defined (defaults to DEEPCODE_API_KEY)	your own	your own	your own

If provider is unset it defaults to deepseek.

Configuring a provider

Option 1: environment variable

Simplest, works out of the box. Each provider only reads its own env var:

export DEEPSEEK_API_KEY=sk-...
# or
export ZHIPUAI_API_KEY=...
# or
export MOONSHOT_API_KEY=...

Option 2: ~/.deepcode/settings.json

Use provider to pick the active backend and providers.{} to hold each provider's key (or a custom endpoint definition). Three examples:

Switch to GLM:

{
  "provider": "glm",
  "providers": {
    "glm": { "apiKey": "sk-..." }
  }
}

Switch to Kimi:

{
  "provider": "kimi",
  "providers": {
    "kimi": { "apiKey": "sk-..." }
  }
}

Point at a self-hosted / third-party OpenAI-compatible endpoint (custom):

{
  "provider": "custom",
  "providers": {
    "custom": {
      "baseURL": "https://your-endpoint.example.com/v1",
      "apiKeyEnv": "MY_ENDPOINT_API_KEY", // defaults to DEEPCODE_API_KEY
      "dialect": "openai",                // deepseek | glm | kimi | openai
      "models": { "fast": "your-fast-model", "smart": "your-smart-model" }
    }
  }
}

custom.baseURL plus custom.models.fast/models.smart are required — missing either one means the config is treated as incomplete and silently falls back to deepseek. dialect only accepts deepseek/glm/kimi/openai; anything else is discarded. Both providers and provider only take effect at the user layer (~/.deepcode/settings.json) — project-layer settings are stripped, so a shared repo's config can't quietly hijack which provider your key gets sent to.

Switching at runtime: /model

Type /model in the interactive TUI to open the picker (or /model <name> to switch directly). The list shows the current provider's tiers first, followed by tiers from other providers that already have a key configured.

• Switching within the same provider (e.g. deepseek fast ↔ smart): instant, no restart.
• Switching to a different provider (e.g. deepseek → GLM): needs a process restart, which deepcode handles for you automatically — it saves the setting, exits, and relaunches with --resume, restoring your session history exactly, just on a different backend.

If the target provider has no key configured, the picker marks it as unconfigured and refuses the switch, so you never restart into a backend that can't actually connect.

smart / fast tiers

Every provider has two tiers: smart for the main loop — planning, writing code, making decisions — and fast for cheaper, quicker sub-operations (subagents, summarization, etc). Defaults are in the table above.

Full built-in model list per provider (models marked "thinking-only" can't have thinking disabled — see below):

DeepSeek: deepseek-v4-flash, deepseek-v4-pro

GLM: glm-5.2, glm-5.1, glm-5, glm-5-turbo, glm-4.7, glm-4.6, glm-4.5, glm-4.5-air, glm-4.6v, glm-4.6v-flash

Kimi: kimi-k3 (thinking-only), kimi-k2.7-code (thinking-only), kimi-k2.7-code-highspeed (thinking-only), kimi-k2.6, kimi-k2.5

Thinking toggle & the thinking-only guard

Thinking defaults to disabled — saves tokens, responds faster. /think toggles it on/off, /effort low|medium|high adjusts how hard it thinks.

Some models are thinking-only — sending thinking:{type:"disabled"} gets flatly rejected by the endpoint (the Kimi tiers marked above). deepcode guards against this: for those models, turning thinking off doesn't send disabled at all, it omits the field entirely so the model falls back to whatever thinking behavior it always uses — turning thinking off never breaks the request.

Pricing

See each vendor's own website for current pricing.

---

Next: Settings & env vars.