Mythical Pro Agents

Anthropic (Claude) Integration Plan

GitHub Edit page

Anthropic (Claude) Integration Plan

Status: shipped (November 20, 2025) Scope: API support for Claude 4.5 Sonnet and Claude 4.1 Opus in Oracle CLI.

#Models & Pricing (public list prices)

  • claude-sonnet-4-6 (CLI alias: claude-4.6-sonnet) — 200k context, ~$3 / 1M input tokens, ~$15 / 1M output tokens.
  • claude-opus-4-1 (CLI alias: claude-4.1-opus) — 200k context, ~$15 / 1M input tokens, ~$75 / 1M output tokens.
  • Prompt-caching premium (not modeled in CLI costs): cached input portion >200k is billed higher (Sonnet ~$6 / 1M; Opus ~$18.75 / 1M).

#Requirements

  • Environment: ANTHROPIC_API_KEY (required), ANTHROPIC_BASE_URL (optional; defaults to https://api.anthropic.com).
  • Engine: API only. Browser mode is blocked for Claude.
  • Tokenizer: @anthropic-ai/tokenizer (wrapped to accept Oracle’s array inputs). Estimates are approximate; rely on API usage for actual billing.

#Planned CLI Behavior

  • Add models to --model/--models: claude-4.6-sonnet, claude-4.1-opus. Aliases: “sonnet”, “opus” map to those IDs.
  • Background runs: disabled for Claude (supportsBackground=false). Even if --background is set, the run streams normally and logs a note.
  • Search / tools: web_search_preview is ignored for Claude; --search is effectively off with a warning.
  • Base URL: --base-url / apiBaseUrl applies per provider; falls back to ANTHROPIC_BASE_URL for Claude, OPENAI_BASE_URL for GPT.
  • Cost display: uses the prices above; prompt-caching billing is not modeled (estimates are upper/lower bounds only).

#Usage Examples

  • Single model (Sonnet):

    • ``bash oracle --engine api --model claude-4.6-sonnet --prompt "Summarize the design doc" --file docs/design.md ``

  • High-reasoning (Opus) with files report:

    • ``bash oracle -m claude-4.1-opus --files-report --prompt "Analyze risk register" --file docs/risk.md ``

  • Multi-model compare (GPT + Claude):

    • ``bash oracle --models gpt-5.1-pro,claude-4.6-sonnet --prompt "Propose mitigation steps" --file docs/plan.md `` Background stays off for Claude; GPT may still use background.

#Implementation Notes (for maintainers)

  • Types/config: Claude entries use apiModel mapping to Anthropic IDs (claude-sonnet-4-6, claude-opus-4-1); Opus stays in ProModelName; pricing + 200k inputLimit; Anthropic tokenizer wrapper; supportsBackground=false. Opus gets reasoning: high.
  • Client factory: branch on claude* to the Anthropic adapter (messages.stream/create); pass provider-specific baseUrl.
  • Env selection: ANTHROPIC_API_KEY and ANTHROPIC_BASE_URL; log masked key per provider.
  • Token estimates: wrapper flattens Oracle message arrays into text before calling countTokens.
  • Multi-model: shared runOptions.background is gated by per-model supportsBackground; Claude never enters the background polling path.
  • Docs to update alongside code: this file, README.md model list, docs/configuration.md, docs/multimodel.md.

#Limitations / Caveats

  • Token estimates for Claude are approximate; API may still reject >200k inputs even if estimate passes.
  • Prompt-caching cost deltas are not reflected in CLI estimates.
  • No tool use/search for Claude in v1.

#Troubleshooting

  • Missing key: “Missing ANTHROPICAPIKEY…” — set the env var or pass --api-key.
  • Background ignored: expected; Claude does not support the Responses-style job API.
  • Search ignored: expected; Claude adapter currently drops web_search_preview.

#Next Steps (post-v1)

  • Optional: add tool/use support with a provider-agnostic tool layer.
  • Add dated model-id mapping if Anthropic starts versioned IDs (similar to Gemini resolver) — currently aliases map to the undated IDs above.
  • Improve cost estimation if API exposes cached-token counters.