Layerz

Layerz MCP

Financial models your AI agent builds and edits: items, formulas, dependencies, branches. Exports to Excel.

Layerz is a structured spreadsheet for financial models, ready for AI. This repository is its plugin: the hosted MCP server plus five skills that teach an agent how to drive it (first session, business plan, monthly close, scenarios and cash, review and fix). Claude, ChatGPT, Codex, Cursor or any MCP client works on the model structure itself, not on loose cells, so formulas hold when the agent changes an assumption and the model is still there next session.

• Endpoint: https://layerz.cc/mcp (streamable HTTP)
• Auth: OAuth 2.1 with PKCE and dynamic client registration, or a Layerz API key
• Docs: layerz.cc/for-agents · llms.txt
• Registry: io.github.layerzlabs/layerz on the official MCP Registry

The server is hosted. There is nothing to run locally: this repository holds the plugin manifests (Agent Plugins 1.0, Claude Code, Cursor), the five skills under skills/, the setup instructions and the metadata MCP directories read. Installing the plugin adds one connection to https://layerz.cc/mcp and nothing else; SECURITY.md says exactly what is sent.

---

Why

Ask an AI to build a financial model and you get a workbook in 30 seconds. Then one assumption changes: links break, values get hardcoded, and the next session starts from scratch.

Layerz keeps the structure of a model (items, formulas with named references, the dependency graph, timelines) apart from its data (values, scenarios, actuals). The agent edits that structure through MCP tools, every change is versioned with a readable summary, and Excel is an export format: clean, standard .xlsx with live formulas, the file you send to people who never see the agent.

Install

Where a plugin can be installed, install the plugin: it is the same server with the skills on top. The bare URL remains the fallback for every other MCP client.

Sign in to Layerz the first time a tool is called. Free for one person working alone.

Server only (any MCP client)

Every client below speaks to the same server at https://layerz.cc/mcp (streamable HTTP, OAuth 2.1 with PKCE and dynamic client registration, or a Layerz API key as a bearer token). It gets the tools, the conventions and the modeling guide, not the skills.

Cursor, without the plugin

{

"mcpServers": { "layerz": { "url": "https://layerz.cc/mcp" } } }

Claude Code, without the plugin

claude mcp add -t http layerz https://layerz.cc/mcp

OpenAI Codex, without the plugin

codex mcp add layerz --url https://layerz.cc/mcp

codex mcp login layerz

Claude, without the link: Settings → Connectors → Add custom connector → https://layerz.cc/mcp

ChatGPT, without the directory: Settings → Security and login → Developer mode, then Plugins → + → https://layerz.cc/mcp

Mistral Le Chat: Settings → Connectors → Add custom MCP connector → https://layerz.cc/mcp

OpenCode: add to opencode.json, then run opencode mcp auth layerz

{

"$schema": "https://opencode.ai/config.json", "mcp": { "layerz": { "type": "remote", "url": "https://layerz.cc/mcp", "oauth": {} } } }

Grok (grok.com): Connectors → New Connector → Custom → https://layerz.cc/mcp

Antigravity CLI (formerly Gemini CLI): add to mcp_config.json

{

"mcpServers": { "layerz": { "serverUrl": "https://layerz.cc/mcp" } } }

Any other MCP client: point it at https://layerz.cc/mcp over streamable HTTP. Clients without OAuth can send a Layerz API key, created in the Layerz app, as a bearer token.

The skills

Five skills, each named after a job the user asks for. The SKILL.md is a short sequence of tool calls for that job; the matter comes from the Layerz modeling guide, copied verbatim into references/ by a script in the Layerz code repo, so the skills can never drift from what the server documents.

Every skill opens with the same rule: explicit instructions from the user come first, then the model's FINANCE.md and the workspace conventions (both returned by the tools), then the user's own skills, then the skill's defaults. The skills state no fact about any business and never quote a workspace's text. Finance judgement (is this plan credible) is not in the plugin: the cookbook has skills for that, yours to fork.

Where instructions come from

Four places, one owner each. The plugin skills (this repository, Layerz) say how the tools work and in which order. Your own skills and instructions, in your agent, say how you work. The workspace Context, written in the Layerz app and shared by the workspace's members, says what this workspace expects of a model (depth, structure, presentation); every client receives it in tool results. Each model's FINANCE.md says what is true of that model (currency, units, signs, statement structure, objective). When they disagree, the narrower one wins: your request in the conversation, then the FINANCE.md, then the Context, then your skills, then the plugin's defaults.

Updates

Two rhythms. The server (tools, conventions, modeling guide, engine messages) is live for everyone as soon as it is deployed. The skills change with a commit here: Claude Code refreshes marketplace plugins at startup or with claude plugin update layerz; Cursor re-reads the repository; ChatGPT and Anthropic's directory take a new version through their review. The skills are written to stay stable: they point at the server for everything that moves.

What the agent can do

29 tools, grouped by job. Read-only tools are annotated as such, so clients can run them without asking.

The server also exposes reference resources the agent reads on its own: modeling-guide://reference (session workflow, modeling patterns, validated patch examples), formula-spec://reference (the formula language) and format-spec://reference (number formats).

Using it well

The facts live on the server and travel with the model; the method is in the skills above and in the guide the server serves:

• Conventions travel with the model. Each model carries a FINANCE.md (currency, units, sign convention, how the statements are built), and each workspace carries standing instructions for agents. The agent receives both when it opens a model.
• The modeling guide comes from layerz_get_conventions { include_guide: true }. Agents read it at the start of a session.
• Recipes (prompts for budgets, three-statement models, monthly closes, LBOs) are in the cookbook.

A good first prompt:

Build a three-statement model for a B2B SaaS company: 2 M€ ARR, 40 % growth, 18-month runway. Ask me what you need to know first.

Pricing

Free for one person working alone: unlimited models, MCP and REST API, Excel and JSON export. A paid plan starts when a second person edits the same models. Details at layerz.cc/pricing.

Related

• finance-md: the open standard for financial conventions an agent reads
• cookbook: recipes for financial work with AI agents
• layerz.cc: the web app, where people review, share and collaborate on the same models

Versioning

One version, the server's. The number published to the official MCP Registry (io.github.layerzlabs/layerz, from the server.json of the Layerz code repo) is the only version; plugin.json, .claude-plugin/plugin.json, .cursor-plugin/plugin.json, lhm.plugin.json, the X-Layerz-Client header of the Claude and Cursor MCP entries (the Agent Plugins mcp.json read by ChatGPT and Codex carries none: OpenAI's public submission rejects server headers) and the layerz_version of every skill mirror it, never one of their own. A change to this repository alone (a skill reworded, a tool declared) ships under the current version. The version check workflow (.github/workflows/version-check.yml, on every push and weekly) fails when they disagree or run behind the registry, and only warn