FlowForge for coding agents

Give Claude Code, Codex or Claude Desktop full control of the Realbotics FlowForge voice-agent platform: build and test AI phone agents, place calls, run campaigns, manage numbers and knowledge bases, and read usage, from a chat with your agent.

Install

Paste this into your coding agent:

Install https://github.com/RealboticsCorp/flowforge-mcp

The agent follows INSTALL.md. (Coding agents: read and follow INSTALL.md step by step.) It will ask for your FlowForge API key (dashboard → Settings → API Keys). Prefer to do it yourself:

git clone https://github.com/RealboticsCorp/flowforge-mcp.git ~/.flowforge-mcp/src
~/.flowforge-mcp/src/install.sh

Needs macOS or Linux with uv or Python 3.10+. Restart your agent afterwards.

Related MCP server: CallCenter.js MCP

What you get

• MCP server (flowforge-mcp, 48 tools). Curated tools cover daily work, and flowforge_api_reference + flowforge_api_request reach all 160 External API endpoints using the API spec bundled with the server.

• flowforge skill teaches the agent the platform's rules and traps: which actions ring real phones or spend money, branch-then-test before publishing, campaign start requirements, knowledge-base flags, webhook-tool shapes, and how to write code against the REST API.

• Slash commands for Claude Code: /ff-call, /ff-batch, /ff-call-review, /ff-campaign, /ff-campaign-status, /ff-assistant, /ff-test, /ff-tool, /ff-usage, /ff-status.

Tool areas

• Assistants: list, get (prompt variables, flow, knowledge base, stats, export), create, update, delete, clone/publish, import, branches, pronunciations.

• Voice: TTS/STT catalog, presets, voice cloning.

• Calls: start, batch (≤100), list, get (transcript, metrics, events, recordings), active calls, end.

• Campaigns: create, control, settings, leads, dispositions and outcomes, campaign calls.

• Phone numbers: list, search, assign, purchase/release, pools.

• Knowledge: semantic and full-text search, document upload/URL import, knowledge bases.

• Tests: run, manage, read results.

• Org and spend: organization settings, MCP servers attached to agents, usage reports, balance.

Safety

• Deletes and actions that dial, spend money or replace org-wide config return a dry_run preview until the agent repeats the call with confirm=true. The skill tells the agent to show you the preview and wait for your go-ahead.

• Requests other than GET are never retried automatically: FlowForge has no idempotency keys, so a retried call request would dial the person twice.

• Path IDs are validated, the API key never follows a redirect to another host, and file uploads refuse hidden paths such as ~/.ssh or .env.

• Your key is stored only in each client's own MCP config.

Several organizations

Each API key belongs to one organization. Install the first key as flowforge and add more with:

FLOWFORGE_API_KEY=ff_live_… ~/.flowforge-mcp/src/install.sh --name flowforge-acme --server-only

The installer records which server belongs to which organization in ~/.config/flowforge/servers.md, and the skill reads that file.

Configuration

Development

uv run python -c "import asyncio; from flowforge_mcp.server import mcp; print(len(asyncio.run(mcp.list_tools())))"
FLOWFORGE_API_KEY=… uv run flowforge-mcp --check
uv run --with httpx python scripts/refresh_spec.py   # pull the latest API spec

flowforge_mcp/
  app.py          # FastMCP instance + constants
  client.py       # authenticated async HTTP client + error translation
  formatting.py   # JSON/Markdown rendering, truncation, pagination
  models.py       # shared Pydantic input mixins
  server.py       # entry point (--check verifies the install)
  spec.py         # index over data/openapi.json (bundled API spec)
  tools/          # one module per resource group; _dispatch.py = confirm gate
skills/flowforge/ # agent skill (SKILL.md + references/)
commands/         # Claude Code /ff-* slash commands
install.sh        # installer (see INSTALL.md)

Contributor rules: AGENTS.md.

License

MIT, see LICENSE.