About pacer-mcp
A retired standalone MCP server for Pacer revenue-management data. Its tools no longer reach the API; the README directs users to remove it and use the in-process Pacer connector instead.
Key features
- 01Retired; API calls no longer work
- 02Historical portfolio and reservation tools
- 03Historical pacing and market metrics
- 04Historical client-health briefs
- 05Historical Guesty pricing tools
Use cases
- 01Reviewing the server's former tool coverage
- 02Identifying the replacement Pacer connector
MCP support
It formerly exposed Pacer API endpoints as MCP tools over stdio. The standalone server is retired and should be removed from MCP client configuration.
Configuration
Local server config is the same for every client. Copy it into your client’s MCP config file.
docker run -i --rm -e PACER_CORE_URL -e PACER_CORE_TOKEN ghcr.io/str-consulting/pacer-mcp:1.0.0PACER_CORE_URLBase URL of the pacer/core app (PAT JSON API). Defaults to production (portal.pacerrev.io).PACER_CORE_TOKENSecretPersonal access token (pat_...) issued via `pacer pat create`. Requires the employee role or higher.Local servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.
ghcr.io/str-consulting/pacer-mcp:1.0.0stdioTools
Local servers expose tools only once running on your machine, and the maintainer has not provided a tool list.
Details
From the project READMEpacer-mcp
🪦 RETIRED. This standalone PAT-over-stdio server no longer reaches the Pacer API. Every tool returns a notice to switch to the in-process ("inline") Pacer MCP connector built into core, offered as the hosted "Pacer" connector on claude.ai / Claude Cowork (tools appear as
mcp__…Pacer__…). The standalone binary called the API over the network atportal.pacerrev.io, which now redirects unauthenticated API calls to an HTML login page; the in-process connector dispatches to the same/api/v1handlers with no network hop and can't drift from the API host. Remove this server from your MCP client config. The setup and tool docs below are kept for historical reference only.
An MCP server that exposes pacer/core API endpoints as native tools for Claude Code (and any other MCP-aware client). One binary, stdio transport, no fuss.
Install
macOS (Homebrew)
brew install STR-Consulting/tap/pacer-mcp
Windows (Scoop)
scoop bucket add pacer https://github.com/STR-Consulting/scoop-bucket
scoop install pacer-mcp
Non-programmer? See docs/windows-setup.md for a copy-paste prompt to hand to your AI assistant.
Linux (binary tarball)
For server/automation use. Replace <version> and pick amd64 or arm64:
curl -fsSL https://github.com/STR-Consulting/mcp/releases/latest/download/pacer-mcp_<version>_linux_amd64.tar.gz | tar -xz
sudo install pacer-mcp /usr/local/bin/
Or use the OCI image: ghcr.io/str-consulting/pacer-mcp:latest (linux/amd64 + linux/arm64).
From source
go install github.com/STR-Consulting/pacer-mcp@latest
Setup
Add to your MCP config (e.g. .mcp.json in your project or your Claude Code user config):
{
"mcpServers": {
"pacer": {
"command": "pacer-mcp",
"env": {
"PACER_CORE_URL": "https://portal.pacerrev.io",
"PACER_CORE_TOKEN": "pat_..."
}
}
}
}
| Env var | Description | Default |
|---|---|---|
PACER_CORE_URL | Base URL of the pacer/core app (PAT JSON API is mounted here) | https://portal.pacerrev.io |
PACER_CORE_TOKEN | Personal access token, format pat_... | (unset) |
Run health_check after install to confirm the server can reach core.
PATs are minted by a core admin via pacer pat create --user <email> --label <name>; they require an employee-or-higher role and are sent as Authorization: Bearer pat_....
Tools
Each tool's full description (caveats, when to use, args) is returned to the MCP client at registration time — agents see them at tool-selection time. The list below is a quick index.
Operational
| Tool | What it does |
|---|---|
health_check | Pings the Pacer API and reports config + reachability |
Portfolio fundamentals
| Tool | What it does |
|---|---|
list_briefable_portfolios | Enumerate active portfolios (optional q name filter) |
list_portfolio_teams | Bulk: notification team (RM/RD/Jon) for every portfolio |
get_portfolio_team | Notification team for a single portfolio |
list_portfolio_units | Unit roster: bedrooms, type, managed/active, location |
list_portfolio_reservations | Reservations in a date range (by check_in/check_out/booked_on) |
Performance
| Tool | What it does |
|---|---|
get_portfolio_pacing | Recent reservations w/ YoY rent/ADR/ABW/LOS + anomaly score |
get_portfolio_metrics_ytd | CY vs PY YTD: revenue, ADR, occupancy, RevPAR, LOS, count |
get_portfolio_market_metrics | One-month CY vs PY + market benchmark deltas; optional decomposition |
Guesty PMS
| Tool | What it does |
|---|---|
guesty_pricing_config | Per-unit pricing intent: base price, fees, min/max nights, factors, channel settings |
guesty_reservation_promotions | Channel-applied promos on reservations in a month (Airbnb, Vrbo, etc.) |
Client health
| Tool | What it does |
|---|---|
get_client_health_brief | Latest (or dated) sentiment brief for a portfolio |
upsert_client_health_brief | Log a lightweight sentiment brief (1-5 + stage + payload) |
list_client_health_briefs | Dashboard view: latest brief per portfolio as of a date |
get_client_health_scoring_config | Scoring weights, labels, and tier thresholds |
upsert_intel_brief | Publish a full intel brief (Postgres + ClickUp task + BigQuery mirror) |
KeyData
| Tool | What it does |
|---|---|
list_managed_keydata_units | Pacer-managed unit UUIDs for a KeyData customer account |
Not wrapped:
POST /portfolios/{p}/intel-brief/attachments(multipart binary upload). MCP has no natural multipart story.
Development
go build -o pacer-mcp .
go test ./...
golangci-lint run --fix ./...
Releases are automated — push a v* tag and GitHub Actions builds darwin-arm64, windows-amd64, and linux-{amd64,arm64} binaries (plus a ghcr.io/str-consulting/pacer-mcp OCI image), then updates the Homebrew tap and Scoop bucket and publishes to the MCP Registry.
License
MIT
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.