Models & endpoints

Runtime LLM behavior comes from ~/.anycode/config.json (and per-agent overrides in routing.agents). Cargo feature openai on anycode-llm adds an OpenAIClient code path; OpenAI-compatible gateways (z.ai, OpenRouter, etc.) still typically use the Zai OpenAI-shaped client unless provider is exactly openai with that feature enabled.

OpenClaw alignment

• Provider list: The CLI and config validator use the static catalog in crates/llm/src/provider_catalog.rs (PROVIDER_CATALOG), aligned with the OpenClaw Provider Directory. Canonical upstream ids live in openclaw/openclaw.
• Model refs: OpenClaw often writes provider/model (e.g. anthropic/claude-opus-4-6). In anyCode, split that into provider + model in config.json (same meaning). You may also put a qualified string in model alone (e.g. anthropic/claude-3-5-sonnet); it is validated independently of the global provider. Resolution helpers live in anycode_llm (build_qualified_chat_model_value, resolve_chat_model_ref, mirroring OpenClaw chat-model-ref.ts).
• anycode status: prints primary_chat_ref, model_routes aliases, and the resolved provider / model per RuntimeMode so you can verify mode routing without starting a session.
• Naming: Config provider values are snake_case (e.g. cloudflare_ai_gateway, vercel_ai_gateway). OpenClaw kebab-case names are accepted and normalized (e.g. cloudflare-ai-gateway → cloudflare_ai_gateway).
• Aliases: Examples: claude → anthropic, zai / bigmodel → z.ai, kimi → moonshot, github-copilot → copilot, amazon-bedrock → bedrock, glm → z.ai.
• AWS Bedrock: Set provider to amazon_bedrock (alias bedrock), choose a model id for your region, and rely on the AWS credential chain (e.g. AWS_PROFILE, instance role). The stack uses Bedrock Converse (Converse / streaming).
• GitHub Copilot: Set provider to github_copilot (alias copilot), pick a Copilot Chat–compatible model id, then run anycode model auth copilot (device flow) so tokens are stored under ~/.anycode/credentials/.
• Placeholders: Some catalog entries remain OpenClaw parity only (e.g. media-only APIs). Use custom with your own OpenAI-compatible base_url when the catalog entry is not wired in anyCode.

Run anycode model to pick a provider interactively; the menu follows the same catalog.

config.json fields (summary)

• provider: A known catalog id (see PROVIDER_CATALOG above), plus aliases such as z.ai / bigmodel / zai, anthropic / claude, or kebab-case OpenClaw-style ids.
• plan: coding or general (affects default z.ai base URL when base_url is empty).
• base_url: optional override.
• model: model id for the active provider.
• api_key: vendor key.
• provider_credentials: extra keys for other vendors when routing mixes providers.
• session (optional): TUI session behavior. auto_compact (default true): before sending your next user message, if the last agent turn reported input token usage above a threshold, anyCode runs automatic compaction (same pipeline as /compact). context_window_auto (default true): derive the context window from provider + model (built-in heuristics in anycode_llm::resolve_context_window_tokens, e.g. Claude ≈200k, GLM/z.ai ≈128k, Gemini ≈1M). Set context_window_auto to false and set context_window_tokens to a fixed size when you want a manual override. Tune auto_compact_ratio (default 0.88) or auto_compact_min_input_tokens (absolute threshold, overrides the ratio). Set auto_compact to false to disable. The runtime flag context-compression (**anycode enable context-compression**) is tracked in runtime.features (see Releases & flags); threshold behavior remains driven by session.auto_compact_* fields above.

z.ai (BigModel)

Default endpoints (when base_url omitted):

• General: https://api.z.ai/api/paas/v4/chat/completions
• Coding plan: https://api.z.ai/api/coding/paas/v4/chat/completions

Client uses OpenAI Chat Completions shape: tools / tool_calls and multi-turn history.

Anthropic

Set provider to anthropic (or claude), provide api_key and a valid model id for the Messages API. Optional base_url overrides the default endpoint.

Retries

Retries with backoff on HTTP 429, 5xx, and transport errors. 401/403 and other non-retryable codes are not retried.

OpenAI official API (optional)

With cargo build -p anycode --features openai, if global provider normalizes to openai, the stack may use OpenAIClient instead of ZaiClient for that profile.

---

More detail (Chinese, including feature matrix and env vars) lives in 模型与端点（中文）.