Work in Progress • Active Development

Spec-driven development
for AI coding agents

An MCP server plugin that gives AI agents a structured development lifecycle. Requirements → Design → Tasks → Implementation, with human approval gates at every transition and a real-time dashboard to manage it all.

The problem

AI coding agents are powerful, but they skip the process that makes software reliable.

🏃

Straight to code

Agents jump to implementation without a requirements phase, design review, or approval step. The result: rework, scope creep, and features that don’t match intent.

🛑

No approval gates

Without structured checkpoints, there’s no place for a human to review direction before code is written. You catch problems in PR review instead of during design.

📄

Spec drift

Requirements live in chat messages that scroll away. Design decisions aren’t recorded. When the agent loses context mid-session, there’s nothing to resume from.

A structured lifecycle

Every non-trivial feature follows the same path. Each transition requires explicit human approval through the dashboard. No phase-skipping.

Optional
Discovery
research first
Phase 1
Requirements
what to build
Phase 2
Design
how to build it
Phase 3
Tasks
break it down
Phase 4
Implement
write the code

Specs are structured markdown files that live on disk — not in any agent’s memory. They survive context loss, session boundaries, and agent switches.

Dashboard approvals

Human in the loop

A real-time web UI where you review specs, approve phase transitions, request changes, or reject proposals. The agent blocks until you decide.

Implementation logging

Artifact tracking

Every implementation step records what was built: functions, endpoints, tests, components. The dashboard shows a complete audit trail of what shipped.

Parallel execution

Subagent dispatch

Tasks with no file overlap can execute concurrently via isolated subagents. Each task goes through implementation, compliance review, and quality review.

Steering documents

Project context

Product vision, technical constraints, and codebase structure captured in steering docs. Agents reference these instead of relying on chat history.

What ships in the box

The npm package provides MCP tools and the dashboard. The GitHub repo additionally ships lifecycle skills and slash commands as plain markdown.

MCP Tools & Prompts
Tools (7) spec-workflow-guide Lifecycle orchestration instructions steering-guide Steering document creation guidance spec-status Phase, completion, implementation audit approvals Dashboard approve / reject / request changes log-implementation Record artifacts (functions, endpoints, tests) write-spec-doc Create or update spec documents spec-list List all specs with status Prompts (7) create-spec Create requirements, design, or tasks doc create-steering-doc Create product / tech / structure steering implement-task Dispatch implementation to subagents refresh-tasks Re-sync task state from spec files spec-status Phase and completion summary inject-spec-workflow-guide Inject workflow guide into session inject-steering-guide Inject steering doc guide into session
Shipped Skills (installed by copying from the repo)
Core lifecycle (12 skills shipped) /spec Full spec lifecycle orchestrator (Requirements → Implement) /discover Structured research before committing to a spec /chat Exploratory conversation with context /start Quick session resume with context (~15 seconds) /prime Full session boot with health checks /wrap End-of-session cleanup and summary /retro Extract prescriptive lessons from the session /audit On-demand project health check /issue Create and manage project issues /pr-cleanup Post-merge branch and worktree cleanup /publish-templates Promote a template into the plugin /migrate-skill Port a user-level skill into the plugin

Works with any MCP agent

SpecFlow is an MCP server — any agent that speaks the protocol gets the full spec lifecycle and dashboard. Spec state lives on disk, so you can start in one agent and continue in another.

Claude Code

Full support

MCP server via npm + optional skills plugin. Full lifecycle skills, slash commands, and subagent dispatch.

Gemini CLI

MCP tools

Add as an MCP server in Gemini’s config. Full access to spec tools and the approval dashboard. Uses GEMINI.md for agent-specific behavior.

Codex CLI

MCP tools

Configure in Codex’s TOML config. Same MCP tools, same workflow, same dashboard. Uses CODEX.md for agent-specific behavior.

Any MCP agent

Protocol-level

Also verified with OpenCode. Any agent that implements MCP gets the full spec lifecycle — no vendor lock-in.

Obsidian-compatible knowledge structure

Spec documents, steering docs, and templates are stored as structured markdown. If you use Obsidian, you get graph visualization and wikilinks for free. If not, it’s just a folder of markdown files.

Vault layout
your-vault/specflow/ templates/ # global spec templates (auto-populated on first run) {Project}/ steering/ # product.md, tech.md, structure.md specs/ # requirements, design, tasks, implementation logs approvals/ # approval records

Each project needs a thin config file (.specflow/config.json) pointing to your vault location. Multi-project setups share one vault.

Extending the stack

SpecFlow handles the spec lifecycle — it doesn’t try to be a memory system, code search engine, or knowledge base. Those are separate concerns. The author’s workflow pairs SpecFlow with these companion tools, but none are required.

Session memory

SessionFlow

Indexes conversation turns into a local vector database for semantic recall across past sessions.

Code search

claude-context

Semantic code search via self-hosted Milvus. Hardened fork of zilliztech/claude-context.

Episodic memory

mem0

Cross-session memory for retro lessons, decisions, and handoffs. Cloud API (self-hosted fork planned).

These integrations are part of the author’s workflow — not SpecFlow requirements. The MCP server and dashboard work standalone with just Node.js.

Get started

One line to install the MCP server. Copy skills from the repo for the full workflow.

Install
# Add to your MCP config (Claude Code example) "specflow": { "command": "npx", "args": ["-y", "@lbruton/specflow@latest", "."] } # Optional: install lifecycle skills git clone https://github.com/lbruton/specflow.git ~/specflow-source cp -r ~/specflow-source/skills/* ~/.claude/skills/ cp -r ~/specflow-source/commands/* ~/.claude/commands/ # Launch the dashboard npx @lbruton/specflow@latest --dashboard --port 5051

Requires Node.js 18+. Works with any MCP-compatible agent.

Give your AI agent a development process

SpecFlow is open source, in active development, and used daily by the author across multiple projects. It’s a work in progress — contributions, issues, and feedback are welcome.