Anthropic Claude
AIRun Anthropic's Claude models across the full API surface: text, vision, PDF analysis with citations, tool use, structured JSON output, Message Batches, the Files API, and Managed Agents. Bring your own Anthropic key.
What This Integration Enables
Claude gives an agent language and vision reasoning as a callable step. Ask a single question, run a full multi-turn conversation with tool use, analyze an image, or read a hundred-page PDF and get an answer with citations back to the source. When output has to be machine-readable, schema-enforced JSON turns a language model into a reliable data producer that downstream systems can trust. For work that does not finish in one call, Claude's Managed Agents run as stateful sessions inside a sandboxed container. You define the agent once, start a session, and let it work asynchronously. FlowRunner's On Session Idle trigger fires when the agent finishes, so the flow reads the answer and downloads output files without polling. Batch processing and token counting keep large or cost-sensitive workloads predictable.
Without FlowRunner
With FlowRunner
Use Case Scenarios
Contract Review With Citations
A vendor emails a contract PDF. The agent calls Analyze Document with citations enabled and asks Claude to extract the payment terms, termination clause, and any change-of-control provision, returning schema-enforced JSON. Because citations are on, each extracted term links back to the exact passage it came from. The agent files the structured summary in the record system and posts a digest to the deal team. Instead of reading every page, the team reviews a short summary with the source text one click away. Reading shrinks to the clauses that carry risk.
Support Triage at Volume
Inbound support messages need to be classified and routed before anyone opens the queue. Rather than call Claude once per message during a spike, the agent builds a Message Batch, submits up to a large batch of classification requests at reduced token cost, and polls Get Message Batch until processing ends. It then reads the results, matches each answer to its original message by custom ID, and routes each ticket to the right team. Triage that used to sit in a shared inbox happens on a schedule, at a predictable cost.
A Research Agent That Knows When to Stop
An analyst needs a competitive brief that takes real work: reading sources, pulling numbers, and drafting a summary. The agent creates a Managed Agents session with a rubric using Define Outcome, and Claude works the task autonomously in its sandbox. When it reaches a step configured to require approval, the session goes idle with a stop reason of requires_action and waits. The On Session Idle trigger fires, FlowRunner routes the pending action to the analyst, and the agent resumes only after Send Tool Confirmation approves it. The long task runs on its own, and the human stays in control of the moments that matter.
Human-in-Loop Highlight
Claude's Managed Agents are built to stop and ask, which is exactly the behavior FlowRunner exists to orchestrate. When a tool is configured with an always-ask policy, the session emits a tool-use event, goes idle with a stop reason of requires_action, and does nothing further until a human responds. FlowRunner treats that pause as a [human-in-the-loop](/concepts/human-in-the-loop/) step: the agent pauses autonomously, assembles the pending action and its context, routes it to the right person through Slack or email, and resumes the moment they answer with Send Tool Confirmation. Approve, and the agent continues. Deny with a reason, and it adjusts its approach. The agent runs the long task, and a person owns the gate.
Agent Capabilities
40 actionsMessages
5- Ask Claude Sends a single prompt to a Claude model and returns the generated text plus token usage. The simplest way to use Claude for a one-shot generate, summarize, rewrite, or classify step inside a flow.
- Send Messages Sends a full conversation to the Messages API with every advanced option: multi-turn history, system prompt, sampling controls, tool definitions (function calling), schema-enforced structured JSON output, and thinking configuration. Returns all content blocks, stop reason, token usage, and a convenience text field. Use this when output has to be reliable and machine-readable.
- Analyze Image Analyzes an image with a Claude vision model and returns the text answer. Accepts a FlowRunner file, a public URL, or an Anthropic Files API file ID. Supports JPEG, PNG, GIF, and WebP.
- Analyze Document Analyzes a PDF or text document and returns the answer, reading both the text and any charts, tables, or images inside a PDF. Optionally enables citations so the answer references the exact passages it draws from. Use it for contract review, report extraction, and any document that needs provenance.
- Count Tokens Counts how many input tokens a request would consume for a specific model without running it. Useful for estimating cost or checking that content fits the context window before sending.
Message Batches
5- Create Message Batch Creates a Message Batch that processes a large volume of Messages API requests asynchronously at reduced token cost. Each request carries a unique custom ID for matching results back to inputs. Use it for classification, extraction, or enrichment across many records at once.
- Get Message Batch Retrieves a batch's processing status and per-request counts. Poll this until processing has ended, then fetch results.
- List Message Batches Lists Message Batches in the workspace, most recent first, with pagination.
- Cancel Message Batch Cancels a batch that is currently processing. Requests already completed still return results; unprocessed requests are canceled and not billed.
- Get Message Batch Results Downloads and parses the results of an ended batch, matching each result to its input by custom ID. Results indicate whether each request succeeded, errored, was canceled, or expired.
Files
5- Upload File Uploads a file from FlowRunner storage (or any accessible URL) to the Anthropic Files API and returns a file ID. Reference it in Analyze Image, Analyze Document, or Send Messages without re-sending the bytes each time.
- List Files Lists files stored in the Files API workspace with pagination, including each file's type, size, and whether it can be downloaded.
- Get File Metadata Retrieves a file's filename, MIME type, size, creation time, and downloadable flag.
- Download File Downloads a file's content to FlowRunner storage. Only files Claude created (via skills or the code execution tool) are downloadable.
- Delete File Permanently deletes a file from the Files API. Deleted files can no longer be referenced.
Models
2- List Models Lists the Claude models available to the API key, most recent first, with each model's ID, display name, and creation date.
- Get Model Retrieves details for a specific model, including context window, output limits, and capability flags. Accepts model aliases.
Agents
6- Create Agent Creates a persisted, versioned Managed Agent (beta) defining its model, system prompt, and tools. Create it once and reference its ID for every run rather than creating a new agent per run.
- Get Agent Retrieves an agent's full configuration: name, model, system prompt, tools, MCP servers, current version, and timestamps.
- List Agents Lists Managed Agents in the workspace with pagination.
- Update Agent Updates an agent's configuration. Every update creates a new immutable version; running sessions keep their version, and new sessions use the latest unless pinned. Supply the complete desired configuration.
- Archive Agent Archives an agent permanently. The agent becomes read-only and new sessions can no longer reference it. Use with care, since there is no unarchive.
- List Agent Versions Lists an agent's immutable versions. Pin a session to a known-good version for reproducibility or rollback.
Agent Environments
4- Create Environment Creates a reusable cloud environment: the template for the sandboxed containers where an agent's tools run. Networking is either unrestricted or a deny-by-default allowlist.
- List Environments Lists Managed Agents environments with pagination, including each container's type and networking policy.
- Get Environment Retrieves an environment's name, description, container configuration, and timestamps.
- Delete Environment Permanently deletes an environment. New sessions can no longer reference it.
Agent Sessions
7- Create Session Starts a Managed Agents session: a stateful, asynchronous run of a pre-created agent in a sandboxed container. Optionally pin a version, mount file or repository resources, and attach vaults with MCP credentials.
- Get Session Retrieves a session's current status (idle, running, rescheduling, or terminated), title, attached resources, token usage, and any outcome evaluations.
- List Sessions Lists Managed Agents sessions with pagination.
- Archive Session Archives a finished session, making it read-only while keeping its event history readable. The routine cleanup step once a session goes idle.
- Delete Session Permanently deletes a session, including its event history, container, and checkpoints.
- Get Session Result Reads a session's answer: concatenates the agent's messages since the last user message, and returns the status, stop reason, and token usage. Typically used right after On Session Idle fires.
- List Session Output Files Lists the output files the agent wrote in its container, ready to fetch with Download File.
Agent Session Events
6- Send Message To Session Sends a user message to a session to kick off or steer work. The agent runs, then returns to idle when done. Messages sent during a run are queued and processed in order.
- Send Custom Tool Result Returns the result of a client-side custom tool call. When the agent invokes a declared custom tool, the session waits with requires_action until this action supplies the result.
- Send Tool Confirmation Approves or denies a pending tool call when a tool uses an always-ask permission policy. This is the action a human-in-the-loop step calls to release or block an agent's next move, optionally with a message explaining a denial.
- Interrupt Session Interrupts a running session, jumping ahead of queued messages so the agent stops at the next safe boundary and goes idle. Follow up with a message to redirect it.
- Define Outcome Starts a rubric-graded work loop: the agent works the task while an independent grader scores each iteration against explicit criteria and feeds gaps back until the rubric is satisfied. Send this instead of a first message when the task has a checkable definition of done.
- Get Session Events Lists a session's event history: agent messages, thinking, tool use and results, status transitions, errors, and outcome evaluations, with optional filtering by event type.
Triggers
1 triggersEvent Triggers
1- On Session Idle Fires when a Managed Agents session transitions to idle, meaning the agent finished its current turn or is waiting on a gate. Polls session state and delivers the completed session so the flow can read the answer, download output files, or route a pending action to a human. Optionally filtered by a Title Contains match so one flow watches only its own sessions.
Start building with Anthropic Claude
$100 in credits. No card required. Connect in minutes.