Proxies to Cloudflare Workers AI for inference and model access via the /cfproxy endpoint.
Provides an AI proxy endpoint to Cloudflare Workers AI, enabling access to Cloudflare's AI models.
Routes requests to OpenAI models (e.g., GPT) through the LiteLLM proxy for AI completions.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@Skein Toolkit MCP Server run the orchestrator to analyze and update the open questions ledger"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
A repo-agnostic clone of the AI development toolchain originally built inside Electron-Splines: a Cloudflare Workers AI proxy, a multi-step task orchestrator, and a local agentic MCP server, fronted by a unified LiteLLM proxy.
This repo is the result of the "LLM Tech Stack Standalone Repository" spin-off
(see architecture-docs/global/ai-task-queue.md, AT-1116-1139, in the
Electron-Splines repo for the full task history and rationale). Further
refinement of this agentic system happens here, not in Electron-Splines.
Path | Purpose |
| The core: a local MCP SSE server (port 3100) that also proxies/instruments Cloudflare Workers AI requests at |
| A lighter MCP server variant for a remote GPU devserver. |
| LiteLLM unified proxy config routing local/CF/Groq/DeepSeek/Anthropic/OpenAI models. |
| Launcher/diagnostic scripts for running Cline against this stack on Windows. |
| Cloudflare Workers AI configuration (env-var only -- no zone/firewall config). |
| Containerized mcp-server + LiteLLM stack. |
Related MCP server: Remote MCP Server (Authless)
Create a Python virtualenv at the repo root and install dependencies:
python -m venv .venv
.venv\Scripts\pip install -r mcp-server\requirements.txtCopy mcp-server\litellm.env.example to mcp-server\litellm.env and fill
in your API keys (see cloudflare/README.md for the Cloudflare token).
Run mcp-server\toolchain-doctor.ps1 to diagnose and (where possible)
auto-start LiteLLM and local-mcp.py.
Run mcp-server\run-cline.ps1 -Task "..." to launch Cline against the
configured model.
By default, local-mcp.py operates on the parent of the mcp-server/
directory (i.e. this repo's checkout). Set WORKSPACE_ROOT to point it at a
different project checkout instead.
cp docker/.env.example docker/.env # fill in API keys
docker compose up --buildThis starts mcp-server (port 3100) and litellm (port 4000, dashboard at
/ui). By default mcp-server operates on ./workspace -- use
docker-compose.override.yml to mount a different project checkout.
Variable | Default | Purpose |
| parent of | Project checkout local-mcp.py reads/writes/runs commands in. |
| -- | Cloudflare Workers AI proxy target + token. See |
|
| Path (relative to |
|
| CF spend-review accounting (optional). |
This is an early-stage clone (Phase 1 of the migration plan in
planning_document.md): the toolchain runs standalone, but the orchestrator's
"open questions" / "actionable tasks" governance integration
(CF_PROXY_OQ_LEDGER_PATH and friends) still assumes an Electron-Splines-style
architecture-docs/ layout when enabled. Generalizing that integration into
reusable create_open_question / create_actionable_task MCP tools is tracked
as AT-1137-1139 in the source repo.
Full mirroring of the consuming project's app/, engine/, and scripts/
directories (AT-1117/1118/1119) has not been done here -- this clone currently
contains only the AI-toolchain pieces (MCP server, orchestrator, LiteLLM,
Cloudflare/Docker config). If the original AT-1117/1118 scope (mirroring the
entire Electron-Splines app and engine source trees into this repo) is still
wanted, that is a separate, much larger effort and should be re-scoped with
the architect first.
Licensed under the Apache License, Version 2.0 -- see LICENSE and NOTICE.
This is a deliberate relaxation from the proprietary Electron-Splines source
repository's license, applying only to this standalone toolkit: the goal is
to let other projects and contributors adopt and extend the MCP
server/orchestrator/CF-proxy toolchain, and to make it mergeable into
complementary open-source projects (e.g. Odysseus, AGPL-3.0 -- permissively
licensed code can be incorporated into an AGPL project).
This server cannot be deployed
MCP server for AI agents to plan, verify, and deploy Cloudflare-native apps.
Cloudflare Workers MCP server: agent-workflow-engine
Cloudflare Workers MCP server: ai-model-router
Cloudflare Workers MCP server: ai-gateway
