Description

n8n-nodes-anthropic-structured

An n8n community chat-model node that wraps Anthropic Claude and exposes Anthropic’s constrained-decoding structured-output API (output_config.format) as a first-class option. Set a JSON Schema, and the model is guaranteed to return schema-conformant JSON — no prompt-injected formatting instructions, no optional tool the model can ignore.

A drop-in alternative to n8n’s built-in Anthropic Chat Model for any workflow that needs reliable structured output.

!Anthropic (Structured) wired into an AI Agent alongside Memory and a Civic tool

Why this exists

n8n’s stock Anthropic node uses LangChain, which (as of early 2026) implements structured output via the legacy tool-use trick — define a synthetic formatfinaljson_response tool and hope the model calls it. This is unreliable on medium-complex schemas: the model can produce free-form text, or call the tool with malformed args, or ignore it entirely.

Anthropic shipped real constrained decoding — outputconfig.format with type: "jsonschema". The schema is enforced at the sampling layer: every emitted token has to match. LangChain hasn’t picked this up yet. This node does.

Installation

In your n8n instance:

1. Go to Settings → Community Nodes
2. Click Install a community node
3. Enter n8n-nodes-anthropic-structured and accept the risk acknowledgement

Setup

1. Create or open a workflow.
2. Add an AI Agent, Basic LLM Chain, or any node with an ai_languageModel input.
3. On the Chat Model input click +, search Anthropic (Structured).
4. Configure an Anthropic API credential (x-api-key header — get a key at console.anthropic.com).
5. Pick a Claude model (4.5+ — only those support output_config.format).
6. (Optional) Paste a JSON Schema into the JSON Schema field. When non-empty, every response is constrained-decoded against that schema.

!Node configuration showing the JSON Schema field

With and without tools

• Without tools (Basic LLM Chain): the single response is JSON-schema conformant. Standard use case.
• With tools (AI Agent + tool nodes): intermediate agent rounds emit tool_use blocks as normal; the model’s final answer is the schema-conformant JSON. Anthropic explicitly supports this composition.

Supported Claude models

• Claude Opus 4.7 / 4.6 / 4.5
• Claude Sonnet 4.6 / 4.5
• Claude Haiku 4.5

Earlier Claude models don’t support output_config.format and aren’t listed in the model picker.

How it works

Verified-community-node compliant — uses only @n8n/ai-node-sdk and standard fetch, no third-party SDKs in the published bundle:

• Extends BaseChatModel from @n8n/ai-node-sdk
• Implements generate() / stream() by POSTing to https://api.anthropic.com/v1/messages directly
• Maps n8n’s Message/MessageContent types to Anthropic’s request shape (system messages extracted to the top-level system field, tool-call/tool-result content blocks translated to Anthropic’s tooluse/toolresult)
• Tools bound by the agent via withTools() are sent as Anthropic tool definitions; input_schema type is backfilled to "object" when missing so Anthropic accepts schemas like {} from tools without parameters
• When jsonSchema is set, includes outputconfig: { format: { type: "jsonschema", schema } }

See nodes/AnthropicStructured/helpers.ts for the conversion functions (covered by vitest tests).

Credentials

The Anthropic API credential takes a single field:

• API Key — your Anthropic API key (sent as the x-api-key header along with anthropic-version: 2023-06-01)

The credential test issues a GET /v1/models against the Anthropic API.

Compatibility

• n8n >= 2.19
• Self-hosted and (once verified) n8n Cloud
• Zero third-party runtime dependencies

Development

pnpm install      # install deps
pnpm lint         # n8n's community-node lint suite
pnpm test         # vitest unit tests
pnpm build        # compile TypeScript and copy static files into dist/
pnpm dev          # run a local n8n with this node loaded (hot reload)

CI runs lint + tests + build on every push and PR. Releases are triggered by pushing a version tag (e.g. 0.1.1); GitHub Actions then publishes to npm with a provenance attestation.

Resources

• Anthropic structured outputs docs
• Anthropic Messages API reference
• n8n community node docs

License

MIT