Design Flow MCP

A thin MCP bridge between ChatGPT-compatible MCP clients and Design Flow System v0.2.

The repository also contains a private Custom GPT Action REST API under design_flow_action_api. Both transports call the same adapter and engine; neither owns semantic authority.

The adapter exposes a small capability-based tool surface while all decisions, provenance, TRACE, validation, persistence, and commit boundaries remain inside the Design Flow engine.

What this is

• A transport and orchestration layer around design-flow-system.
• Adapter version 0.1.0.
• Engine version 0.2.0, pinned to merge commit 30bedf2032179fcb46186f30cc86c33b070c17d1.
• One active Design Flow project/session per server process; multiple project paths may exist below the configured root.

What it is not

• Not the Design Flow semantic engine.
• Not an LLM provider.
• Not a database.
• Not a replacement for the Design Flow CLI.
• Not a shell, filesystem browser, Python evaluator, or raw-state editor.

Authority boundary

import_draft changes working draft state but creates no decision authority. preview_round is explicitly non-authoritative. Only lock_round can change decision authority, and it calls PersistentProject.lock_draft() rather than editing persistence files. No tool directly edits decisions, concepts, supersession, TRACE, manifests, hashes, or project JSON.

Tools

Read-only tools:

readiness, get_state, get_unresolved, get_round, preview_round, compile_context_handoff, compile_living_document, recommend_next_round, get_decision_ledger, get_concepts, get_trace, and get_session_brief.

Working/session or authority-changing tools:

• new_project and resume_project create or open a project and engine session.
• import_draft changes non-authoritative working state.
• lock_round is the only decision-authority-changing tool.
• end_session ends the Design Flow session without marking the project complete.

See Tool contracts for preconditions and exact boundaries.

Project-root security

Set DESIGN_FLOW_PROJECT_ROOT to the only directory projects may occupy. Relative paths are resolved below it. Parent traversal, absolute escapes, and detectable symlink escapes are rejected. Draft file imports are subject to the same confinement.

Local setup on Windows PowerShell

cd $HOME\Documents
git clone https://github.com/fargo161/design-flow-mcp.git
cd design-flow-mcp

python -m venv .venv
.\.venv\Scripts\python.exe -m pip install -e .

$env:DESIGN_FLOW_PROJECT_ROOT="$HOME\Documents\DesignFlowProjects"
New-Item -ItemType Directory -Force $env:DESIGN_FLOW_PROJECT_ROOT | Out-Null

.\.venv\Scripts\python.exe -m design_flow_mcp.server

The default transport is local stdio, where the MCP client launches this command as a subprocess.

For local Streamable HTTP development:

.\.venv\Scripts\python.exe -m design_flow_mcp.server --transport streamable-http --host 127.0.0.1 --port 8000

The MCP endpoint is http://127.0.0.1:8000/mcp. Streamable HTTP is a deployment surface, not an authentication or tunneling solution. Do not expose it publicly without HTTPS, authentication, network policy, and the MCP SDK's production host/origin configuration.

Development

For an editable local engine checkout, install it first and install this adapter without resolving dependencies:

python -m pip install -e ..\design-flow-system
python -m pip install "mcp>=2,<3" build
python -m pip install -e . --no-deps
python -m unittest discover -s tests -v
python -m build

Normal installation uses the exact Git dependency declared in pyproject.toml; engine source is not copied or vendored here.

Private Custom GPT Action API

$env:DESIGN_FLOW_PROJECT_ROOT="$HOME\Documents\DesignFlowProjects"
$env:DESIGN_FLOW_API_KEY="replace-with-at-least-32-random-characters"
.\.venv\Scripts\python.exe -m design_flow_action_api

The local API is available at http://127.0.0.1:8080/docs. A Custom GPT requires deploying the included Dockerfile behind HTTPS and configuring the X-API-Key Action credential. Project paths remain hidden behind stable opaque IDs.

Supply the API key only through environment or hosting secret management. Do not put secrets in project files or the Docker build context. The REST boundary returns stable sanitized errors and never intentionally exposes physical server paths.

Documents

• Architecture
• ChatGPT setup
• Tool contracts
• Private plugin architecture
• Private plugin testing
• Phase 1 data handling draft
• Submission readiness
• Phase 2 owner decisions
• Custom GPT setup
• Custom GPT instructions
• Action API authority model
• Local Action API development