DevFlow MCP Server
MCP server for integrating DevFlow with AI code assistants. Enables structured development workflows with enforced planning, task tracking, and code review gates.
What is DevFlow?
DevFlow brings structure to AI-assisted development. Instead of letting AI agents write code without oversight, DevFlow enforces a workflow:
idea -> planning -> approval -> ready -> in_progress -> review -> done
Every feature or bugfix goes through planning, gets approved, is tracked with tasks, and must pass review before completion. The MCP server connects your AI coding assistant to DevFlow, making this workflow seamless.
Installation
As a Claude Code Plugin (recommended)
DevFlow is listed in the official Claude Code community marketplace. Install in one line:
/plugin marketplace add anthropics/claude-plugins-community
/plugin install devflow
This installs everything in one shot: MCP server (auto-registered via mcpServers in the plugin manifest), enforcement hooks, state-specific skills, and slash commands (/devflow-start, /devflow-status, /devflow-next, /devflow-tasks, /devflow-create, /devflow-list, /devflow-review). No separate MCP setup step needed.
Since 4.23.0: the MCP server no longer modifies your CLAUDE.md. The plugin's skills, hooks, and MCP tool responses cover the rules that the old <!-- DEVFLOW-RULES-START --> block used to inject. If you have an existing block from a pre-4.23 install and want it cleaned up, run npx @dev-flow-tech/mcp-server uninstall --client claude once.
Since 4.26.0 (DF-339): Two new auto-review skills fire when transitioning state. devflow-plan-critic reminds at planning → approval to deep-review the plan against 7 dimensions. devflow-code-critic does the same at in_progress → review for the implementation. Both support iterative loop (max 3), trivial-flow-skip, and auto-approve-gate based on project_configs.allow_agent_self_approval.
Since 4.30.0 (DF-374/DF-377): The DevFlow backend is the primary gate for state-transitions (unified flowGate service with conditions like plan-required, tasks-required, agent-summary, testing-instructions, knowledge-gaps-resolved, pr-state-consistent, adr-compliance-clean). The plugin's strictness checks are now informational warnings — the backend's 409 gate.failures[] response is the source of truth. This makes Multi-Client setups (Claude Code, Codex, Cursor, Gemini) consistent — the gate behavior is the same regardless of which client connects.
Updates: /plugin update devflow.
Manual Setup (fallback)
If you'd rather skip the plugin and wire the MCP server directly:
Claude Code
npx github:KlausFreiberufler/devflow-mcp setup
Cursor
npx github:KlausFreiberufler/devflow-mcp setup --client cursor
Gemini CLI
npx github:KlausFreiberufler/devflow-mcp setup --client gemini
Windsurf
npx github:KlausFreiberufler/devflow-mcp setup --client windsurf
Restart your AI client after setup.
Server-Split (since DF-334)
For clients with per-server tool-caps (e.g. Cursor 40-tool/server cap), DevFlow ships three entry-points:
| Bin | Domain | Tool-Count |
|---|
devflow-mcp | combined (default, BC) | ~65 |
devflow-mcp-flows | workflow (flow_, task_, agent_session_, devflow_, release_*) | ≤40 |
devflow-mcp-wiki | knowledge (adr_, wiki_, doc_page_, knowledge_, planning_context) | ≤40 |
Cursor (or any cap'd client) registers BOTH split-bins to get full tool-coverage:
{
"mcpServers": {
"devflow-flows": { "command": "npx", "args": ["-y", "@dev-flow-tech/mcp-server", "devflow-mcp-flows"] },
"devflow-wiki": { "command": "npx", "args": ["-y", "@dev-flow-tech/mcp-server", "devflow-mcp-wiki"] }
}
}
Claude Code uses the combined devflow-mcp bin (no cap, simpler config).
Getting Started
- Run
devflow_status() to check the connection
- The browser opens for authentication on first use
- Link your project — a
.devflow.json file is created in your project directory
- Start working:
flow_list() to see available flows, or flow_create() to create one
- Initialize your session:
devflow_init({ flowId: "..." })
Compatible Clients
Flow States
