Peil MCP Server

Peil MCP server

mcp-name: app.peil/peil-mcp

Connect Peil to Claude (or any MCP client): log hours, draft invoices from unbilled hours, and check where you stand — from a prompt.

• Docs: https://peil.app/en/docs/mcp
• Source: https://github.com/Luminc/peil-mcp

The server is a pure client of Peil's public API. It authenticates with a scoped API key you create in Peil under Settings → Developer (Pro).

Draft-by-default

draft_invoice only ever creates a draft — nothing is sent to your clients. Sending is a separate tool (send_invoice) that also requires the separate invoices:send permission on your key. A key without that permission can never email anything on your behalf.

Tools

Reads (read)

Hours (timesheet:write)

Clients (clients:write)

Invoices (invoices:write)

Client-facing email (invoices:send — irreversible, always confirm first)

---

1. Create your key

In Peil: Settings → Developer → create a key with the permissions you want. Start with read + Log hours + Draft invoices; leave Send invoices off unless you truly want an assistant emailing clients. You'll paste this key into your assistant's config as PEIL_API_KEY below.

2. Install the server

The easy way — no clone, no install (once peil-mcp is published to PyPI):

uvx peil-mcp          # runs the latest release on demand
# or, with pipx:
pipx run peil-mcp

→ Your launch command is uvx peil-mcp (or pipx run peil-mcp). Skip to step 3. Everything below is only needed if you're running from source (e.g. before the first release, or to hack on it).

From source

You need Python 3.11 or newer and a local copy of this mcp-server/ folder. Check your Python with python3 --version.

Which method? Run which uv pipx first. If you already have uv, use A — it's the least work. If not, pipx (B) gives you a clean global command. If you have neither and don't want to install tooling, the plain-venv path (C) works with nothing but the Python that's already on your machine.

Everywhere below, replace /ABS/PATH/TO/mcp-server with the real absolute path to this folder (run pwd inside it to get it).

A. With uv (no install step)

uv builds and runs on demand — nothing to install first:

uv run --directory /ABS/PATH/TO/mcp-server peil-mcp

→ Your launch command is: uv run --directory /ABS/PATH/TO/mcp-server peil-mcp

Don't have uv? curl -LsSf https://astral.sh/uv/install.sh | sh (macOS/Linux) or pip install uv.

B. With pipx (isolated global command)

pipx installs the server into its own isolated environment and puts a peil-mcp command on your PATH:

pipx install /ABS/PATH/TO/mcp-server

→ Your launch command is simply: peil-mcp

Don't have pipx? python3 -m pip install --user pipx && python3 -m pipx ensurepath.

C. Plain venv + pip (works with only stock Python)

No extra tooling — just the python3 you already have:

cd /ABS/PATH/TO/mcp-server
python3 -m venv .venv
.venv/bin/pip install .

→ Your launch command is: /ABS/PATH/TO/mcp-server/.venv/bin/peil-mcp (equivalently /ABS/PATH/TO/mcp-server/.venv/bin/python -m peil_mcp).

Use a plain pip install . (not -e/editable) for running. An editable install relies on a .pth path hook that can silently fail to load on some setups, giving ModuleNotFoundError: No module named 'peil_mcp'. Editable is only needed if you're modifying the server itself — see Local development.

3. Connect your assistant

The only thing that changes between assistants is where the config lives. Every MCP client needs the same three things:

• command — your launch command from step 2
• env — PEIL_API_KEY set to the key from step 1
• (optional) PEIL_API_URL — only if you're pointing at a non-production Peil (see Local development); defaults to https://api.peil.app/api/v1.

The canonical config block (used by Claude Desktop, Cursor, Windsurf, Cline, and most others) looks like this — command + args are just your launch command split on spaces:

{
  "mcpServers": {
    "peil": {
      "command": "peil-mcp",          // or "uv", or the venv's python path
      "args": [],                     // e.g. ["run","--directory","/ABS/PATH/TO/mcp-server","peil-mcp"] for uv
      "env": { "PEIL_API_KEY": "your-key" }
    }
  }
}

Claude Desktop

Edit claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\), add the block above, and restart Claude Desktop.

Claude Code (CLI)

# pipx / venv (single-command launcher):
claude mcp add peil -e PEIL_API_KEY=your-key -- peil-mcp

# uv:
claude mcp add peil -e PEIL_API_KEY=your-key -- uv run --directory /ABS/PATH/TO/mcp-server peil-mcp

Anything after -- is the launch command. Verify with claude mcp get peil (look for Status: ✔ Connected) and use /mcp in a session to reconnect.

Cursor

Add the canonical block to .cursor/mcp.json (this project) or ~/.cursor/mcp.json (all projects), then enable peil in Settings → MCP.

Windsurf

Add the canonical block to ~/.codeium/windsurf/mcp_config.json, then hit Refresh in the Cascade MCP panel.

Cline / Roo (VS Code)

Open the extension's MCP Servers → Configure panel and add the canonical block to cline_mcp_settings.json.

VS Code (native Copilot agent mode)

VS Code uses a slightly different shape — servers (not mcpServers) and an explicit type — in .vscode/mcp.json:

{
  "servers": {
    "peil": {
      "type": "stdio",
      "command": "peil-mcp",
      "args": [],
      "env": { "PEIL_API_KEY": "your-key" }
    }
  }
}

Any other MCP client

Give it the same command + args + PEIL_API_KEY env. The server speaks MCP over stdio; if a client can launch a stdio command, it can run Peil.

First prompts

Once connected, try:

• "Where do I stand?" → orientation_snapshot
• "Log 6 hours to De Correspondent today for editing work." → log_hours
• "Draft an invoice from my unbilled hours for De Correspondent." → draft_invoice

With Send invoices left off your key, an assistant can prepare everything but physically cannot email a client — you send from Peil yourself.

---

Local development

Point the server at a local backend with PEIL_API_URL:

PEIL_API_URL=http://localhost:8000/api/v1 PEIL_API_KEY=your-local-key peil-mcp

If you're modifying the server, an editable install picks up your changes without reinstalling:

.venv/bin/pip install -e ".[dev]"

Tests (mocked HTTP, no backend needed — pythonpath = ["src"] in pyproject.toml makes them independent of the install mechanism):

.venv/bin/pytest