generated: '2026-08-27'
method: searched
source: openapi/*.yml, https://platform.claude.com/docs/en/api/overview, https://platform.claude.com/docs/en/api/errors
base_url: https://api.anthropic.com
authentication:
  style: api-key-header
  header: x-api-key
  admin_header: x-api-key (Admin API key, sk-ant-admin...)
  alternatives: Workload Identity Federation (short-lived OIDC bearer tokens); SigV4
    on Claude Platform on AWS.
  ref: authentication/anthropic-authentication.yml
versioning:
  style: date-header
  header: anthropic-version
  current: '2023-06-01'
  ref: lifecycle/anthropic-lifecycle.yml
beta_opt_in:
  header: anthropic-beta
  style: comma-separated dated identifiers
idempotency:
  supported: false
  notes: No idempotency-key header on the Messages API. Message Batches provide at-most-once
    semantics via per-request custom_id.
pagination:
  style: cursor
  params: [before_id, after_id, limit]
  response_fields: [data, has_more, first_id, last_id]
  managed_agents: Sessions/events use next_page and prev_page cursors passed via the
    page parameter.
request_tracing:
  header: request-id
  body_field: request_id
  notes: Every response includes a unique request-id header, mirrored as request_id
    in error bodies. On Claude Platform on AWS a second x-amzn-requestid header is
    returned for CloudTrail.
error_envelope:
  shape: '{ "type": "error", "error": { "type": "...", "message": "..." }, "request_id":
    "..." }'
  ref: errors/anthropic-problem-types.yml
rate_limit_signaling:
  headers: [retry-after]
  model: Token-bucket with cache-aware ITPM (cache reads excluded on most models);
    RPM/ITPM/OTPM limits per usage tier.
  ref: rate-limits/anthropic-rate-limits.yml
streaming:
  style: server-sent-events
  events: [message_start, content_block_start, content_block_delta, content_block_stop,
    message_delta, message_stop, error, ping]
  notes: Errors can occur mid-stream after a 200; handle error events separately.
content_type: application/json
request_size_limits:
  messages: 32MB
  token_counting: 32MB
  batches: 256MB
  files: 500MB
reversibility:
  grade: verified
  grade_basis: >-
    At least one write surface publishes BOTH a reversal operation and a stated
    window: cancelMessageBatch is documented as callable "any time before
    processing ends", and the batch results window is stated as 29 days from
    creation with a 24-hour processing expiry. Most other write surfaces publish a
    reversal path with NO stated window, and several publish no reversal at all —
    recorded honestly below rather than averaged away.
  docs: https://platform.claude.com/docs/en/build-with-claude/batch-processing
  surfaces:
  - resource: message batch
    write: createMessageBatch
    reversal: cancelMessageBatch
    window: Any time before processing ends. Cancellation is best-effort — the batch
      enters a `canceling` state and the system may complete in-progress,
      non-interruptible requests first, so a cancel may result in zero canceled
      requests. Check request_counts.canceled.
    window_stated: true
    grade: verified
    docs: https://platform.claude.com/docs/en/api/messages/batches/cancel
  - resource: message batch (expiry, not a reversal)
    write: createMessageBatch
    reversal: null
    window: Batches expire if processing does not complete within 24 hours; results
      are downloadable for 29 days after creation, after which the batch remains
      readable but its results do not. Expired requests are not billed.
    window_stated: true
    grade: verified
    note: A time bound, not an undo. Recorded because it is the window an agent must
      plan against.
    docs: https://platform.claude.com/docs/en/build-with-claude/batch-processing
  - resource: message batch (hard delete)
    write: createMessageBatch
    reversal: deleteMessageBatch
    window: null
    window_stated: false
    grade: documented
    note: Permanent. Delete is the opposite of a reversal here — it destroys the
      results, it does not restore anything.
  - resource: file
    write: createFile
    reversal: deleteFile
    window: null
    window_stated: false
    grade: documented
    note: >-
      Permanent delete with no published restore and no grace window. The soft
      alternative is expiry: set expires_in_seconds at upload and the file object
      reports expires_at.
    docs: https://platform.claude.com/docs/en/build-with-claude/files
  - resource: managed agent
    write: createAgent
    reversal: archiveAgent
    window: null
    window_stated: false
    grade: documented
    note: Archive is a soft state; Anthropic's docs do not state an un-archive
      operation or a restore window.
  - resource: managed agent session
    write: createSession
    reversal: archiveSession
    window: null
    window_stated: false
    grade: documented
    note: >-
      deleteSession also exists and is permanent — the webhook catalog states
      session.deleted leaves no object to fetch. Archive and delete are different
      actions; archive terminates the session (session.status_terminated).
  - resource: environment
    write: createEnvironment
    reversal: archiveEnvironment
    window: null
    window_stated: false
    grade: documented
    note: deleteEnvironment is permanent, including deleting an already-archived
      environment.
  - resource: workspace
    write: createWorkspace
    reversal: archiveWorkspace
    window: null
    window_stated: false
    grade: documented
    note: No published un-archive operation or window. Treat as one-way.
  - resource: organization invite
    write: createOrganizationInvite
    reversal: deleteOrganizationInvite
    window: Before acceptance only. Once the invite is accepted there is no invite
      left to revoke; removeOrganizationMember is the action instead.
    window_stated: true
    grade: verified
  - resource: workspace member
    write: addWorkspaceMember
    reversal: removeWorkspaceMember
    window: null
    window_stated: false
    grade: documented
    note: Removes workspace access only; the user remains in the organization.
      removeOrganizationMember is the wider, separate action.
  - resource: skill
    write: createSkill
    reversal: deleteSkill
    window: null
    window_stated: false
    grade: documented
    note: >-
      Versioned — createSkillVersion adds a version rather than overwriting, so
      rolling back to an earlier getSkillVersion is the practical reversal for a
      bad publish. Deleting the skill is not.
  - resource: message (inference)
    write: createMessage
    reversal: null
    window: null
    window_stated: false
    grade: na
    note: >-
      Inference is not reversible and Anthropic does not claim it is. A generated
      message cannot be un-generated, un-billed, or recalled. This is the honest
      answer for the API's highest-volume write, and an agent should treat spend as
      the irreversible consequence — cap it with a workspace spend limit or a
      Managed Agents session budget rather than expecting an undo.
  irreversible:
  - createMessage (inference and its token spend)
  - deleteFile
  - deleteMessageBatch
  - deleteSession / deleteEnvironment / memory store delete (webhook docs state these
    leave no object to fetch)
  spend_controls:
  - {control: workspace spend limit, docs: https://platform.claude.com/docs/en/api/rate-limits}
  - {control: Managed Agents session budget, docs: 'https://platform.claude.com/docs/en/managed-agents/budgets',
     note: 'A hard cap on session spend at public list rates; the session pauses with stop reason budget_reached instead of starting new model requests, and changing or removing the budget resumes it.'}
dry_run_mode:
  supported: partial
  mechanism: countMessageTokens
  ref: sandbox/anthropic-sandbox.yml