Claude · Anthropic

Access and limits

Authentication

Workspace API key (plus optional Admin API key)

Plans and access

Anthropic API (Claude Console) organisations with credits or billing. Keys are scoped to a workspace; model access and limits follow the usage tier. Admin API reads need an Admin API key (sk-ant-admin...) from an organization admin and are unavailable to individual accounts.

Test environment

Use a separate workspace with a spend limit for testing.

Rate limits

Requests, input tokens and output tokens per minute per model; token counting, batches, files and skills have separate limits. HTTP 429 with retry-after. Usage and cost reports support polling about once per minute.

Pagination

Message batches, workspaces, members and API keys page with after_id while has_more is true; files, skills, rate limits, Managed Agents lists, usage and cost reports page with next_page tokens.

Provider events

None for the Messages API. Flows run on demand, on a schedule, or when a scheduled check finds new or changed records.

Before you build a flow

• Authenticated model discovery and bounded text generation. Select a model available to your key. Generation sends mapped content to the provider and incurs provider charges. Automatic retries are disabled; inspect usage before repeating an uncertain request. No agents, tool calls, image generation or embeddings.
• Extraction and batch creation are billable and never retried automatically; check usage before repeating an uncertain request.
• Extraction uses native structured outputs (output_config.format json_schema); schemas must use the JSON Schema subset Anthropic documents. Check the valid flag for refusals or truncated output.
• Batch creation accepts up to 500 requests per Nexra call (Anthropic allows 100,000). Cancelling batches is not implemented.
• File and skill uploads (multipart) and file downloads are not implemented: the recipe engine sends JSON bodies and reads JSON.
• Anthropic's API reference lists the Admin API (usage and cost reports, workspaces, API keys, rate limits) under its beta organization namespace; treat these reads as beta. They require the optional Admin API key.
• Usage and cost reports accept one workspace, API key, model, service tier and group-by value per request. Amounts are in cents as decimal strings and exclude Priority Tier.
• The Claude Code usage report is not implemented because it returns user email addresses. Organization users and invites are not read for the same reason.
• Read only for administration: creating, renaming or archiving workspaces, updating or archiving API keys, member and role changes, and deleting files, skills or batches are intentionally excluded.
• Managed Agents reads (agents, agent versions, environments, sessions, deployments, deployment runs, memory stores) are beta and send anthropic-beta managed-agents-2026-04-01 (memory stores send agent-memory-2026-07-22). They are metadata only: agent system prompts, tool schemas, MCP server URLs, deployment initial prompts, outcome descriptions, session events and memory contents are never mapped. Creating, updating, running, pausing, archiving or deleting agents, sessions, environments, deployments, vaults and memory stores is intentionally excluded; vaults and credentials are not read.
• Message Batch results (/v1/messages/batches/{id}/results) are read as JSON Lines with message_batch_results.list: one record per request with its result type, stop reason, text output and token usage. Tool-use, thinking and other non-text content blocks are not mapped. Each page re-reads the results file up to that page, so very large batches page slowly.
• Service accounts and federation (WIF) endpoints require an org:admin OAuth token rather than an Admin API key, and are not implemented.
• The session status filter sends one statuses[] value; Anthropic documents the parameter as a repeatable array, and the bracketed form follows the convention used by the usage report.

Developer reference