agentsflow

AgentsFlow

Agentsflow is the production-grade control plane for coding agents. It provides a harness that controls workflow setup, routing, and guardrails so coding agents stay on rails. Full operator and contributor docs live at docs.agentsflow.dev.

Quick start (MCP)

The easiest way to run AgentsFlow is as an MCP server using uvx (part of uv).

Prerequisites

• uvx (install uv to get the uvx runner)
• Temporal CLI/dev server (for the local Web UI at http://localhost:8233): https://docs.temporal.io/cli
• One coding agent CLI: Claude Code, Codex CLI, or Gemini CLI (pick whichever assistant you use)

Option 1: Use with AI Assistant (Recommended)

Run the following commands to add AgentsFlow as an MCP server.

Claude Code

claude mcp add agentsflow --scope user --env OPENAI_API_KEY=sk-... --env PROCESS_CODING_AGENT_PROVIDER=claude -- uvx agentsflow

Gemini CLI

gemini mcp add agentsflow --scope user --env OPENAI_API_KEY=sk-... --env PROCESS_CODING_AGENT_PROVIDER=gemini -- uvx agentsflow

Codex CLI

codex mcp add agentsflow --env OPENAI_API_KEY=sk-... --env PROCESS_CODING_AGENT_PROVIDER=codex -- uvx agentsflow

Cursor (.cursor/mcp.json or via Settings > MCP)

{
  "mcpServers": {
    "agentsflow": {
      "command": "uvx",
      "args": ["agentsflow"],
      "env": {
        "OPENAI_API_KEY": "sk-...",
        "PROCESS_CODING_AGENT_PROVIDER": "codex"
      }
    }
  }
}

PROCESS_CODING_AGENT_PROVIDER lets you choose which coding agent (claude / codex / gemini) AgentsFlow will route coding actions to.

Option 2: Standalone HTTP Server

To run the stack independently (e.g. for debugging or remote access):

uvx agentsflow --transport streamable-http

Tracking Progress

Regardless of how you run it (Option 1 or 2), you can track workflow execution and status in the Temporal Web UI at http://localhost:8233.

3. Use it

Once connected, you can use natural language to interact with the process workflow. A few concrete prompts:

• "Call start_agentsflow_process on this repo with issue https://jira.example.com/browse/TEAM-123."
• "Kick off start_agentsflow_process with task text 'Fix 500 error when cookie is missing' for the current repo."
• "The workflow paused; send these answers via provide_agentsflow_clarification and then await_agentsflow_result with the workflow id in a list (it returns after the first workflow finishes or pauses)."
• "Check the latest status for workflow process-1a2b3c using await_agentsflow_result([\"process-1a2b3c\"]) and call again for additional workflows if needed."

You can supply either an issue URL or a free-text task when calling start_agentsflow_process, but not both.

Architecture (high level)

flowchart LR
  Dev[CLI / MCP client] -->|start process workflow| Temporal[Temporal Server]
  Temporal --> Worker[AgentsFlow worker]
  Worker --> Activities[Git / Issue / ACP activities]
  Activities --> Repo[(Git worktree)]
  Activities --> Providers[Jira or GitHub issues]
  Activities --> ACP[Claude / Gemini / Codex ACP]

Supported issue providers

Docs

• New operators: Quickstart + deployment: https://docs.agentsflow.dev
• Process workflow: https://docs.agentsflow.dev/latest/workflows/process/
• Providers + env reference: https://docs.agentsflow.dev/latest/guides/providers/