Anthropic (Claude): Add Models

The Anthropic provider fetches the model list live from the API on every startup, so new Claude models appear automatically once Anthropic publishes them. Code changes are only needed for per-model cost overrides or capability flags — and neither is currently required for any Claude model.

Audience: coding agents and contributors adding or adjusting Claude models in NodeTool.

TL;DR

New Claude models need zero code changes in almost every case. The provider calls GET https://api.anthropic.com/v1/models at startup (getAvailableLanguageModels, line 113 of anthropic-provider.ts), and any model Anthropic publishes there is immediately available. The only cases that require code:

Cost tracking is missing: add a pricing entry to @pydantic/genai-prices upstream, or open an issue if genai-prices already covers it.
A new capability cannot be inferred at runtime: extend hasToolSupport or add a capability flag (currently all Claude models return true for tool support).

Where things live

Concern	Path
Provider class	`packages/runtime/src/providers/anthropic-provider.ts`
Dynamic model fetch	`AnthropicProvider.getAvailableLanguageModels()` — line 113
Tool support check	`AnthropicProvider.hasToolSupport()` — line 109
Cost calculation	`packages/runtime/src/providers/cost-calculator.ts`
Token pricing catalog	`@pydantic/genai-prices` (npm dependency)
Provider ID constant	`PROVIDER_IDS.ANTHROPIC = "anthropic"` in `packages/protocol/src/api-types.ts` line 864
Provider registration	`packages/runtime/src/providers/index.ts` line 203
Claude Agent SDK provider	`packages/runtime/src/providers/claude-agent-provider.ts`

How Claude models are defined

AnthropicProvider.getAvailableLanguageModels() (line 113–200) does a direct HTTP call:

GET https://api.anthropic.com/v1/models
Headers:
  x-api-key: <ANTHROPIC_API_KEY>
  anthropic-version: 2023-06-01

It retries up to three times (exponential backoff: 1 s, 2 s, 4 s) on HTTP 429/5xx, times out after 10 seconds per attempt, and returns [] on auth errors (401/403) without retrying. Each entry from payload.data becomes a LanguageModel:

{ id: "claude-opus-4-5", name: "claude-opus-4-5", provider: "anthropic" }

There is no static fallback list. If the API is unreachable, no Anthropic models appear in the model selector until the next successful fetch.

Tool support: hasToolSupport() (line 109) unconditionally returns true for all Claude models. Anthropic’s tool-use API is available across the whole current model family.

Vision / image input: handled in convertMessage() (line 323) and convertImagePart() (line 269). Any model can receive image content; the provider always encodes images as base64 inline blocks regardless of model ID.

Extended thinking: the thinkingBudget parameter (used in generateMessages line 539 and generateMessage line 725) enables thinking: { type: "enabled", budget_tokens: N }. No model-specific check exists today — callers set the budget; the API rejects it for models that don’t support thinking.

Add / adjust a model

When nothing is needed

A new Anthropic model (e.g. claude-opus-5) appears in NodeTool automatically once Anthropic adds it to their /v1/models list. No code change required.

Verify with:

npm run dev:nodetool -- node run nodetool.text.Concat \
  --props '{"a":"test","b":""}' --no-secrets
# Then check the model selector in the UI or:
npm run dev:nodetool -- info

When cost tracking needs updating

Claude token costs are priced via @pydantic/genai-prices. Check if the new model ID is already in that package:

node -e "const {calcPrice}=require('@pydantic/genai-prices'); \
  console.log(calcPrice({input_tokens:1000,output_tokens:500},'claude-opus-5','anthropic'))"

If calcPrice returns null, the model is not in the catalog. Options:

Open a PR to pydantic/genai-prices (upstream), then bump the package version here.
If the model is token-based and you need an interim workaround, note that the local MODEL_TO_TIER table in cost-calculator.ts is for non-token modalities only (image/audio/3D). Adding a Claude model there will not produce correct per-token costs — go upstream.

Cost tracking flows through trackUsage() in base-provider.ts (line 352), called by the Anthropic provider at message_stop (line 580 for streaming, line 743 for non-streaming). Anthropic’s API reports uncached input separately from cache read/write tokens; the provider sums them before passing to the cost calculator to match genai-prices’ expected contract (see comment at line 577).

When capability detection needs updating

hasToolSupport() (line 109) returns true for all models. If Anthropic releases a model where this is not true:

// packages/runtime/src/providers/anthropic-provider.ts
async hasToolSupport(model: string): Promise<boolean> {
  // example: a hypothetical embedding-only model
  if (model.startsWith("claude-embed-")) return false;
  return true;
}

Run npm run typecheck && npm run lint after any edit to this file.

ClaudeAgentProvider (separate path)

packages/runtime/src/providers/claude-agent-provider.ts is a distinct provider registered as PROVIDER_IDS.CLAUDE_AGENT_SDK = "claude_agent_sdk" (line 210 of index.ts). It authenticates through a local claude CLI subscription rather than an API key. It also fetches its available models dynamically. Changes to AnthropicProvider do not affect it and vice versa.

Verify

Run these after any change to the Anthropic provider:

# 1. TypeScript — must pass
npm run typecheck

# 2. Lint — must pass
npm run lint

# 3. Full check (typecheck + lint + tests)
npm run check

# 4. Confirm models are returned (requires ANTHROPIC_API_KEY)
npm run dev:nodetool -- info --json | grep -A5 '"anthropic"'

# 5. Smoke-test a single-node call
npm run dev:nodetool -- node run nodetool.llm.Claude \
  --props '{"prompt":"hello","model":"claude-haiku-4-5","max_tokens":50}'

How past commits did it

d1491abf (“add claude agent package”, 2026-06-26) introduced the ClaudeAgentProvider and wired it as CLAUDE_AGENT_SDK. The AnthropicProvider itself was already present; that commit shows the registration pattern at index.ts line 210 and the relationship between the two Anthropic-backed providers.

abea04ec (“Add AI-driven 3D scene editing…”, 2026-06-26, PR #3913) extended the tool message contract to support image content blocks in tool results (convertMessage lines 335–362). This is an example of a capability extension that only touched anthropic-provider.ts — no model list changes were involved.

Both commits follow the pattern: model availability is never hardcoded; behavioral changes are isolated to the provider class.

Contributing

PRs welcome at https://github.com/nodetool-ai/nodetool. Before pushing:

npm run check   # typecheck + lint + test

Join the discussion on Discord.

Edit this page on GitHub