Build, change, run, host and scale your product on Fractal.

Setting up agents once, and using them in any process.

• AI Nodes

  An AI node calls a model in a process; the answer goes on to the next nodes, and the process runs the tools it asks for.

• AI Providers

  Connect Anthropic, OpenAI or Google once and reuse the credentials; which model runs is set apart from the process.

• Reusable Agents

  An Agent bundles a model, instructions, skills and settings. Define it once and use it in any process.

• Agent Skills

  A Skill is a reusable set of instructions, examples, response style and guardrails you can add to any Agent.

• Agent Settings

  Limits on tokens and tool rounds, sampling, and each provider’s own options: Anthropic’s thinking budget, OpenAI’s reasoning effort, Google’s safety thresholds.

What the agent knows when it answers.

• Task Context

  The task’s data is put into the agent’s instructions, so it answers with the current facts.

• Conversation Memory

  Earlier messages of a conversation are loaded and new ones saved, so the agent remembers what was said.

• Knowledge Sources

  Any node can be a knowledge source — a database query, an API call — and the agent looks things up in it as it works, instead of everything going into the prompt.

What an agent is allowed to do.

• Tool Definitions

  Describe each tool with a name, inputs and outputs, and list exactly which process nodes and MCP tools the agent may use.

• Tool Execution

  The agent picks a tool and the process runs it, round after round, until the agent is done or reaches your limit.

• Agents as Tools

  One AI node can use another AI node as a tool; both share one time limit.

Using what the agent returns.

• AI Results

  The answer and its token usage are added to the task, ready for conditions, next nodes and replies.

• AI Errors

  Failures come back with a cause — rate limit, timeout, too many rounds, input too large — so each can take its own path.

Using other MCP servers from your processes and agents.

• MCP Discovery

  Preview

  Connect an MCP server and see what it offers: tools, resources and prompts.

• MCP Authorization

  Preview

  Connect servers that are public, that use a bearer token, or that use OAuth — tokens are refreshed for you.

• Pinned Tool Definitions

  Preview

  When a process is published, the name, description and inputs of each MCP tool it uses are saved with it.

• MCP Calls

  Preview

  A node can call a tool, read a resource or get a prompt from an MCP server, using the task’s data.

• MCP Tools

  Preview

  Choose exactly which tools of an MCP server an agent may use — the rest stay out of its reach.

• MCP Results

  Preview

  The server’s answer is added to the task; if the call fails, the task takes its error path.