Models

OpenAgent is model agnostic by design. It supports multiple LLM providers, and every model gets the same MCP tools, memory behavior, channels, and client surfaces.

Smart Router

The recommended default. Classifies each incoming message into a tier (simple / medium / hard) using a cheap classifier model, then routes to the tier-appropriate model. Also falls back to a cheaper model automatically when your monthly budget is nearly exhausted.

model:
  provider: smart
  monthly_budget: 50              # USD; router downgrades when near/over budget
  classifier_model: gpt-4o-mini   # cheap model that picks the tier
  routing:
    simple: gpt-4o-mini
    medium: gpt-4.1-mini
    hard: gpt-4.1
    fallback: gpt-4o-mini         # used when budget is exhausted

Every request gets logged to a usage_log table with input / output tokens and computed cost. The REST endpoint GET /api/usage returns the running totals; GET /api/usage/pricing returns the price-per-million table used for cost computation.

Model prices come from a bundled default pricing file (gpt-4.1 / gpt-4o-mini / gemini-2.5-flash / claude-sonnet-4-6 / etc.). User overrides in providers.<name>.models[].input_cost_per_million take precedence.

Claude CLI (Claude Agent SDK)

Uses your Claude Pro/Max membership — flat rate, not pay-per-token.

model:
  provider: claude-cli
  model_id: claude-sonnet-4-6
  permission_mode: bypass     # auto-approve all tool calls

Requires claude CLI installed and authenticated (claude login).

Claude API (Anthropic SDK)

model:
  provider: claude-api
  model_id: claude-sonnet-4-6
  api_key: ${ANTHROPIC_API_KEY}

Z.ai GLM / Any OpenAI-compatible

model:
  provider: zhipu
  model_id: glm-5
  api_key: ${ZAI_API_KEY}
  base_url: https://api.z.ai/api/paas/v4

Works with Ollama, vLLM, LM Studio — just change base_url:

model:
  provider: zhipu
  model_id: llama3
  base_url: http://localhost:11434/v1
  api_key: ollama

OpenAI

Standard OpenAI-hosted models run through the same Agno provider as Z.ai; the only difference is you put the key under providers.openai and reference the model as openai:<id> (or use the Smart Router above).

providers:
  openai:
    api_key: ${OPENAI_API_KEY}
    models:
      - gpt-4o-mini
      - gpt-4.1-mini
      - gpt-4.1

Hot Reload

Change model: or providers: in openagent.yaml and the gateway rebuilds the active model on the next incoming message — no restart. Other sections (mcp, scheduler, channels, system_prompt) still need a restart to take effect.