Agents and Protocols
Connect external tools and delegate work to agent-compatible systems
MCP
The MCP node connects Innflow to external Model Context Protocol servers. Use it when a workflow needs to discover or call tools exposed by another system.
Modes:
| Mode | Description |
|---|---|
| Direct tool execution | Select a server tool and call it directly from the workflow |
| Agentic use | Let a model discover and call tools from the server |
| Read resource | Retrieve a resource by URI |
| Get prompt | Retrieve a named server prompt with its arguments |
Operations:
- Agentic Mode (
AGENTIC) — LLM discovers and calls MCP tools autonomously - Call Tool (
DIRECT) — Call a specific MCP tool with mapped inputs - Read Resource (
READ_RESOURCE) — Read a resource exposed by the MCP server - Get Prompt (
GET_PROMPT) — Render a prompt exposed by the MCP server
Configuration:
| Field | Description |
|---|---|
| Server URL | Streamable HTTP endpoint for the MCP server |
| Authentication | Credentials required by the server, if any |
| Tool | Tool to call in direct execution mode |
| Input | Tool input payload, with support for template variables |
For a node named mcp, a direct tool call exposes {{mcp.content}} and optional {{mcp.structuredContent}}; agentic execution exposes {{mcp.text}}. Resource and prompt modes return the corresponding server payload. Inspect the selected mode’s result for its nested fields.
All operations and configuration fields.
A2A Agent
The A2A Agent node delegates work to an external agent that supports the Agent-to-Agent protocol.
Common uses:
- Send a task to a specialized external agent
- Poll for an agent result before continuing the workflow
- Connect Innflow workflows to agent systems owned by another team or vendor
Operations:
- Polling Mode (
POLLING) — Send task and poll for result - Streaming Mode (
STREAMING) — Receive real-time SSE updates from the agent
Configuration:
| Field | Description |
|---|---|
| Endpoint | A2A-compatible agent endpoint |
| Mode | Polling or task submission mode |
| Task | Instructions or payload to send to the external agent |
| Authentication | Credentials required by the endpoint |
The node supports polling and streaming modes. Use polling when Innflow should wait for a completed task result, and streaming when the external agent exposes incremental progress.
All operations and configuration fields.
Browser Agent
Browser Agent is a separate workflow node that uses Stagehand and Browserbase for interactive web tasks. It is not an MCP or A2A transport. See Browser Automation.
Operations:
- Run Browser Task (
RUN_TASK) — Complete one multi-step task in a fresh browser session - Run Browser Sequence (
RUN_SEQUENCE) — Run deterministic browser actions in one shared session
All operations and configuration fields.
Tool-Only Nodes
Some nodes can be connected to AI Agent nodes as callable tools instead of running as standalone workflow steps.
| Tool node | Use for |
|---|---|
| Slack Tool | Let an agent send or read Slack messages |
| Gmail Tool | Let an agent send or read Gmail messages |
| HTTP Tool | Let an agent call an HTTP API |
| Exa Tool | Let an agent search or retrieve web content |
These legacy tool-only node types are hidden from universal search. Existing workflows can retain them; for new workflows use the tool-connection controls exposed by the AI Agent editor.