ChatCompletion sends a prompt to a Claude model and returns the response. Set model to choose the model (e.g., claude-opus-4-7, claude-sonnet-4-6) and maxTokens to cap the response length. It returns the completion text, any tool uses, the stop reason, and cache-token counts as outputs, and emits input/output/cache token-usage metrics.

Enable thinking with budgetTokens of at least 1024 and less than maxTokens to request extended thinking. The task sets temperature to 1 when thinking is enabled and warns if a different value was configured. Reasoning returned by Claude is available in thinking, and redacted reasoning remains in rawResponse. Set stream: true to log complete text lines as they arrive and flush the final partial line. Killing the task cancels the active request in both streaming and non-streaming modes and closes the client. The final outputs and usage metrics have the same shape as a non-streaming response. An interrupted or incomplete stream fails the task.

Message content accepts either the existing plain string or a list of blocks with type: TEXT, IMAGE, or DOCUMENT. Image and PDF sources accept type: URL with url, type: BASE64 with data and mediaType, or type: FILE with fileId from UploadFile. The lowercase API spellings are accepted too. Image media types are image/jpeg, image/png, image/gif, and image/webp, and PDFs use application/pdf. The API spellings media_type and file_id are also accepted. Pass betas: [files-api-2025-04-14] for uploaded file sources. To request document citations, set citations: {enabled: true} on the document block. Response references are returned in citations, including document indexes and page, character, or block positions.

Use toolChoice: AUTO to let Claude decide, ANY to require a tool, TOOL with toolChoiceName to select one, or NONE to disable tools. Extended thinking supports AUTO and NONE. stopSequences accepts a list of stop strings, and stopSequence returns the matched string. Optional metadata: {userId: workflow-analysis} attributes requests to a user or workflow, and serviceTier accepts AUTO or STANDARD_ONLY. betas also accepts additional Anthropic beta identifiers without a plugin release.

For explicit prompt cache breakpoints, set cacheControl: EPHEMERAL on a custom tool, a content block, or a message (applied to its final content block). Use systemPromptBlocks with text and optional cacheControl instead of the string system property to cache individual system blocks. These breakpoints can be combined with the existing global promptCaching flag, subject to Anthropic's cache breakpoint limits.

Content block types, source types, and cache control use uppercase enum names in the UI, such as TEXT, BASE64, and EPHEMERAL.

All Anthropic tasks accept an optional baseUrl, connectTimeout, readTimeout, and maxRetries. Durations use ISO-8601 strings, for example connectTimeout: PT30S and readTimeout: PT5M. Unset timeouts use SDK defaults, and zero disables that timeout. maxRetries defaults to 2, and 0 disables SDK retries. Requests, including the API key, go to baseUrl, so when the API key is set in plugin defaults, set baseUrl there too with forced: true.

CountTokens estimates input tokens for the same messages, system, tools, and model you would send to ChatCompletion. It calls the token-count endpoint and returns inputTokens only. It does not create a message or call the model. A later task can read {{ outputs.count_tokens.inputTokens }} and skip ChatCompletion when the count is too high. CountTokens and MessageBatch reject FILE sources because they cannot send the Files API beta header.

MessageBatch creates, retrieves, or cancels an Anthropic Message Batch. Set mode to CREATE and pass requests (each with a unique customId, a model, and messages) to start a batch and return batchId. customId is letters, digits, _, or -, 1–64 characters, and is sent unchanged. Set mode to RETRIEVE with that batchId to read status and requestCounts. Once status is ended, the same call also returns resultsUrl. By default fetchType is STORE: per-request rows are written to Kestra internal storage as an ION file (uri, with size) instead of being kept on the task output. Set fetchType to FETCH to return every row on results, FETCH_ONE for the first row only, or NONE to skip the download. If the results file is no longer available, RETRIEVE still returns the batch status and logs a warning. Repeat RETRIEVE from a LoopUntil task until the batch ends, and set retry on that task for transient download errors. CANCEL stops a batch that is still processing. Batches run asynchronously (at most 24 hours) and are billed at half the real-time Messages API rate. RETRIEVE emits one total per token-usage metric for the succeeded rows it actually read.

ListModels lists the Claude models available to your API key.

The Files API tasks manage files for use with Claude, all via Anthropic's beta Files endpoint: UploadFile uploads a file from Kestra internal storage, GetFile retrieves a file's metadata, ListFiles lists stored files, and DeleteFile removes one.

ManagedAgentSession runs a Claude Code session on Claude Managed Agents, so nothing is installed or hosted on the Kestra side. It authenticates with apiKey only. Set model and prompt, and optionally systemPrompt, a Git repository (url, branch, token), mcpServers, skills and maxCostUsd (a spending limit above 0). The task creates an agent and an environment for the run, follows the session events in the execution logs, and archives the temporary agent and environment at the end. Set environmentId to reuse an existing environment; otherwise networking sets the access of the created one, and LIMITED with allowedHosts is the safer choice for untrusted prompts or repositories. It returns sessionId, the final message, the status and costUsd. status is FINISHED, SPENDING_LIMIT_REACHED or WAITING_FOR_INPUT; any other ending fails the task. Killing the execution, or setting the task timeout, interrupts the session.

ClaudeCodeCLI runs the Claude Code CLI in headless mode (claude -p) in a task runner (Docker, Kubernetes, cloud batch or process), so the session executes on your own infrastructure and can work on namespace files. It authenticates with apiKey only, and installs a pinned version of the CLI unless installCli is false; the key is hidden from the install step, but any beforeCommands you add can read it. Set prompt, and optionally systemPrompt, model, allowedTools, permissionMode, maxTurns and maxCostUsd. Files created or changed in the working directory are returned as outputFiles, and the other outputs match ManagedAgentSession: sessionId, message, status (adding MAX_TURNS_REACHED), costUsd and numTurns. The task fails when the session ends with an error. The session transcript can contain secrets read by the agent, so it is neither logged nor stored unless keepTranscript is true. Prefer ACCEPT_EDITS with allowedTools: BYPASS_PERMISSIONS runs any command with the privileges of the task runner.