Overview
LEX AI is a local MCP server that gives your IDE assistant ten purpose-built workflow modes, exposing one tool surface at a time.
Overview
LEX AI is a local FastMCP server that gives your IDE assistant (GitHub Copilot, Claude Code, Cursor, Codex, Windsurf) a structured, tool-driven way to build, document, change, audit, test, adapt, and deploy LEX App projects. Instead of free-form prompting, the assistant works through purpose-built modes, each with its own guardrails, workflow, and outputs.
The core idea: one server, ten modes, one surface at a time
The recommended entry point is a single unified process, lex-mcp. It loads every mode at boot but mounts and exposes exactly one mode's tools to the IDE at any moment. Switching modes flips the visible tool surface in place, without restarting the process, and refreshes the assistant's tool list. This keeps every workflow deterministic: the assistant only ever sees the tools that belong to the job you chose. With no recorded choice, the server starts in forward.
Mode | What it is for |
|---|---|
Brief ( | Interview you and write the project contract, |
Forward ( | Build a new LEX App end to end (20-step workflow). |
Backward ( | Reverse-document existing code into a technical map and business docs (18 steps). |
Edit ( | Make one focused, well-checked change to an existing app. |
Review ( | Run an on-demand audit (convention or business compliance). |
Test ( | Write the test suite, ground it in user stories, prove it catches regressions. |
Input ( | Adapt an existing app to a changed input-data format. |
Deploy ( | Get an existing app running in a target environment. |
MVP generator ( | Scaffold a small, capped first version fast. |
MVP completion ( | Expand an approved MVP into full-product scope. |
Only forward and mvp_completion push to GitHub. mvp_generator commits locally and never adds a remote; the rest leave your git history alone.
What every mode carries
read_input_file— converts a data file to markdown on the server:.csv,.tsv,.txt,.xlsx,.xlsm,.xls,.pdf. The alternative is the assistant writing a throwaway pandas script and reading your column names out of a terminal, which fails outright on some hosts.render_pdf_pages— rasterises PDF pages as images, so a multimodal model can read a scan or a chart that carries no text layer.ask_user_questionandrecord_user_answer— put one question to you through a host dialog, a local form, or plain text, and record who authored the answer, so an assistant's assumption is never filed as your decision.
Why a workflow layer instead of raw prompting
- Ordering and gates are enforced by the server, not the model — the assistant coordinates; it does not improvise the process.
- Every run leaves durable artifacts — contracts, plans, technical maps, and reports become handoff memory for humans and later AI sessions.
- Errors are structured — a failure returns
ok: falsewith a troubleshooting payload written for you rather than a stack trace, and the assistant's standing instruction is to stop and show it. - Git and delivery are handled for you in the modes that own them, with deterministic, non-interactive git operations.
The coordinator-agent model
Your IDE assistant is the coordinator: it calls an MCP tool, receives the next brief plus the name of a specialised agent, delegates the work, and reports completion back to the server. The agents — lex-step-04 for architecture, lex-deploy-verifier to establish that a deployed app actually works — do the reading and writing. The coordinator never does step work itself when the server has named an agent for it, which keeps context lean and results consistent.
Seven coding tools are supported: GitHub Copilot in JetBrains IDEs, in VS Code, and in the CLI, plus Cursor, Claude Code, OpenAI Codex, and Windsurf. The agents are authored once and rendered into each one's native layout — see Supported coding tools.
How to get started
- Run setup in your LEX project folder:
lex setup-with-ai. - Open the control surface:
lex ai-dashboard, confirm the project and pick a mode. - Start a fresh assistant request describing what you want.
See Setup and dashboard and Choose the right path next.
Was this helpful?
Your feedback shapes what we write next.
