populi-mcp
# populi-mcp
MCP server for the [Populi](https://populi.co/api) API v2. It lets Claude, Codex, Cursor and any other [Model Context Protocol](https://modelcontextprotocol.io) client look up people, admissions, academics and billing data in your Populi instance, and, if you opt in, make changes.
**Package:** `@kisasa/populi-mcp` on npm. **Command:** `populi-mcp`. **Transports:** stdio (local, the default) or Streamable HTTP with sign-in (`--http`, for [Claude and ChatGPT connectors](#hosted-server-claude-and-chatgpt-connectors)). **Protocol:** MCP 2026-07-28, plus the 2025-era `initialize` handshake for older clients, on both transports.
## How it works
Populi's API has 697 routes across 180 models. Exposing each as its own tool would flood a model's context, so this server gives the model a small toolset plus the API reference itself:
| Tool | What it does |
|------|--------------|
| `populi_search_api` | Keyword search over the bundled route catalog (`applications`, `invoice`, `enrollments`, ...). |
| `populi_describe_api` | Full reference for one route: path ids, required and optional fields with types, `expand` options, filter conditions, required Populi roles, docs link. |
| `populi_get` | `GET` any route. Supports `filter`, `expand`, `page`, `limit`, and `all_pages` to merge every page into one list. |
| `populi_download` | `GET` a route and **save the response to a file** instead of returning it. For PDF/XLS/CSV exports, transcripts, Data Slicer results, file and backup downloads, and for keeping a large export out of the conversation. |
| `populi_find_people` | Look up people by exact email and/or name. |
| `populi_status` | Shows the instance, key type (sandbox or live), whether writes are on, the files folder and the timeout. `verify: true` makes one live request to check the key. |
| `populi_write` | `POST` / `PUT` / `DELETE` any route, including multipart **file uploads**. **Only exists when `POPULI_ALLOW_WRITES=true`.** |
Together they cover the whole API: all 492 documented `GET` routes, all 205 write routes, file uploads and file downloads. Routes that aren't in the bundled catalog work too, since the tools take any path.
A typical exchange: the model searches for the route, reads its parameters, then calls `populi_get` with a server-side filter.
The catalog is generated from the public API docs and ships inside the package, so searching and describing work offline and cost no API requests.
## Configuration
| Variable | Required | Description |
|----------|----------|-------------|
| `POPULI_BASE_URL` | yes | Your instance, e.g. `https://yourschool.populiweb.com`. A trailing `/api2` is fine and optional. Must be `https`. |
| `POPULI_API_KEY` | yes | An API key token. In Populi: **Account & Settings > API > Keys** (Account Administrator). Sent as a bearer token to `POPULI_BASE_URL` only. |
| `POPULI_ALLOW_WRITES` | no | Set to `true` to enable `populi_write`. Default: read-only. |
| `POPULI_FILES_DIR` | no | The one folder downloads are saved to and uploads are read from. Default: `~/populi-mcp-files`. |
| `POPULI_TIMEOUT_MS` | no | How long one Populi request may take, 1000 to 600000. Default: `120000` (two minutes). Each call can also pass `timeout_ms`. |
The key's Populi roles decide what it can see. A route the key cannot reach returns a 403 naming the problem; `populi_describe_api` lists the roles each route needs.
### Sandbox first
Populi issues separate `sk_sandbox_...` keys that only work against sandbox instances (live keys only work on live). Point this server at a sandbox while you evaluate it.
### Read-only by default
With no `POPULI_ALLOW_WRITES`, the write tool is not registered at all, so the model cannot even attempt a change. When you enable it, `populi_write` is marked destructive, so clients that prompt for approval will ask before each call. `DELETE` in Populi is permanent.
## Install
Requires Node.js 20 or newer. Every client below runs `npx -y @kisasa/populi-mcp`, so there is nothing to install by hand.
### Claude Code
```bash
claude mcp add populi \
-e POPULI_BASE_URL=https://yourschool.populiweb.com \
-e POPULI_API_KEY=sk_sandbox_... \
-- npx -y @kisasa/populi-mcp
```
### Claude Desktop
Edit `claude_desktop_config.json` (macOS: `~/Library/Application Support/Claude/`, Windows: `%APPDATA%\Claude\`):
```json
{
"mcpServers": {
"populi": {
"command": "npx",
"args": ["-y", "@kisasa/populi-mcp"],
"env": {
"POPULI_BASE_URL": "https://yourschool.populiweb.com",
"POPULI_API_KEY": "sk_sandbox_..."
}
}
}
}
```
On Windows, if Claude Desktop cannot find `npx`, use `"command": "cmd"` with `"args": ["/c", "npx", "-y", "@kisasa/populi-mcp"]`.
### Codex
```bash
codex mcp add populi \
--env POPULI_BASE_URL=https://yourschool.populiweb.com \
--env POPULI_API_KEY=sk_sandbox_... \
-- npx -y @kisasa/populi-mcp
```
or in `~/.codex/config.toml`:
```toml
[mcp_servers.populi]
command = "npx"
args = ["-y", "@kisasa/populi-mcp"]
[mcp_servers.populi.env]
POPULI_BASE_URL = "https://yourschool.populiweb.com"
POPULI_API_KEY = "sk_sandbox_..."
```
### Cursor
In `.cursor/mcp.json` (project) or `~/.cursor/mcp.json` (global). Cursor expands `${env:NAME}`, which keeps the key out of the file:
```json
{
"mcpServers": {
"populi": {
"command": "npx",
"args": ["-y", "@kisasa/populi-mcp"],
"env": {
"POPULI_BASE_URL": "https://yourschool.populiweb.com",
"POPULI_API_KEY": "${env:POPULI_API_KEY}"
}
}
}
}
```
### Any other MCP client
Run `npx -y @kisasa/populi-mcp` as a stdio server with `POPULI_BASE_URL` and `POPULI_API_KEY` in its environment.
### More than one instance
Add one entry per instance under different names (`populi-sandbox`, `populi-live`), each with its own key.
## Hosted server (Claude and ChatGPT connectors)
The stdio server above runs on each person's computer. To let a whole staff use Populi from Claude or ChatGPT (including the web and mobile apps, which can only reach remote servers), run the same code as a hosted server with `--http` and add it as a **custom connector**.
Each person signs in with **their own Populi API key**, so every request runs with that person's Populi permissions and rate limit. A key never has to be shared, and the operator holds no credentials: the key lives only inside an encrypted token that the person's AI app stores.
```
Claude / ChatGPT ──OAuth + MCP over HTTPS──▶ populi-mcp --http ──▶ Populi
(person clicks Connect, pastes their key once) (stateless, no database)
```
### How sign-in works
1. The app discovers the server (`/.well-known/oauth-...`) and registers itself (RFC 7591), then sends the person to `/authorize`.
2. That page asks for their Populi API key (and instance address, unless the deployment pins one), checks it with one live request, and optionally lets them opt in to writes.
3. The server returns an authorization code; the app trades it (with PKCE) for a one-hour access token and a 30-day refresh token. Both are AES-256-GCM sealed copies of the person's credentials, so the server stores nothing and any instance can serve any request.
### Run it
```bash
docker build -t populi-mcp .
docker run -p 8080:8080 \
-e POPULI_MCP_SECRET="$(openssl rand -base64 48)" \
-e POPULI_BASE_URL=https://yourschool.populiweb.com \
-e POPULI_MCP_PUBLIC_URL=https://mcp.yourschool.edu \
populi-mcp
```
Without Docker: `POPULI_MCP_SECRET=... npx -y @kisasa/populi-mcp --http`. The connector URL is the public address plus `/mcp`. It must be served over HTTPS (put it behind a load balancer or reverse proxy, or use a host that terminates TLS for you).
| Variable | Required | Description |
|----------|----------|-------------|
| `POPULI_MCP_SECRET` | yes | 32+ random characters. Seals every token. Changing it signs everyone out, so keep a copy. |
| `POPULI_MCP_PREVIOUS_SECRETS` | no | Comma-separated older secrets that still open tokens, for rotation: put the new secret in `POPULI_MCP_SECRET` and the old one here. |
| `POPULI_MCP_PUBLIC_URL` | recommended | The `https://` address people use. If unset it is derived from the request (`x-forwarded-proto`/`host`), which works when the server is reached at its own address but not behind a proxy or custom domain that rewrites the host. |
| `POPULI_BASE_URL` | no | Pin everyone to one Populi instance; the connect page then asks only for a key. Tokens for any other instance stop working. |
| `POPULI_ALLOWED_HOSTS` | no | When the instance is *not* pinned, the hostname suffixes people may enter. Default `populiweb.com,populi.co`. This keeps the server from being pointed at arbitrary sites. |
| `POPULI_ALLOW_WRITES` | no | `true` lets each person tick an "allow changes" box when connecting. Reads only otherwise, and unticked people stay read-only. |
| `POPULI_ALLOWED_REDIRECTS` | no | Where sign-in may send people back to: `scheme://host` entries or `scheme:`. Default: claude.ai, claude.com, chatgpt.com, chat.openai.com, platform.openai.com, localhost, 127.0.0.1, and the cursor/vscode schemes. |
| `POPULI_TIMEOUT_MS` | no | As above. |
| `PORT` | no | Default `8080`. |
### Add it as a connector
- **Claude** (claude.ai or Desktop): Settings > Connectors > Add custom connector, paste `https://<your address>/mcp`. On Team and Enterprise plans an owner adds it once and each member then clicks **Connect** and signs in themselves.
- **ChatGPT**: Settings > Connectors (or Apps), enable developer mode if your workspace requires it, create a connector with the same URL, and choose OAuth. Workspace admins may need to allow custom connectors first.
- **Claude Code**: `claude mcp add --transport http populi https://<your address>/mcp`, then run `/mcp` to sign in.
### Deploying it
The image runs anywhere containers do, and `--http` needs only Node.js 20+ if you would rather not use one. Where and how you deploy (which platform, what terminates HTTPS, how secrets are supplied) is up to you. A few things matter whatever you choose:
- Pin a released version (`@kisasa/populi-mcp@x.y.z`) rather than building from `main`.
- Keep `POPULI_MCP_SECRET` in your platform's secret store, and keep a copy: losing it signs everyone out.
- Allow long requests. Some Populi reports take minutes, so a front end that cuts requests off at 30 seconds will break them.
- Allow request bodies of about 5 MB if people will upload files.
### What differs from the local server
- **Files.** There is no local files folder. `populi_download` returns text and JSON in full and attaches a file to the result (up to 4 MB); `populi_write` uploads take `content_base64` only (the request body is limited to about 5 MB).
- **No sessions.** Every request is self-contained, so scaling and restarts are invisible. A warm instance still keeps each person's one-at-a-time queue and 429 backoff.
- **Protocol versions.** Clients on MCP 2026-07-28 start with `server/discover` and send each request on its own, with no `initialize`. Older clients get the classic `initialize` handshake, answered statelessly. Both get plain JSON replies, never a stream, so buffered hosts work. The server advertises no tool-list changes, and turns down `subscriptions/listen` straight away rather than holding a stream open.
- **Revoking access.** Tokens can't be listed or revoked individually. Delete the person's key in Populi (the immediate fix), or rotate `POPULI_MCP_SECRET` to sign everyone out.
- **Authorization codes** are not single-use (no database to track them), but they last five minutes and need the PKCE verifier only the connecting app holds.
- **Logs** contain method, path, status and duration only: never bodies, tokens or keys.
## API behaviors worth knowing
- **Rate limits.** 50 requests/min per key from 3AM-7PM PST, 100/min overnight, plus per-IP and per-school limits. On HTTP 429 the call returns an error with how long to wait, and later calls fail fast with the remaining wait instead of hammering a limited key. Prefer one filtered call over many unfiltered ones.
- **One request at a time.** Populi allows a single "heavy" call (Data Slicer, big exports) at once, so the server runs requests sequentially even if the client sends several in parallel.
- **Pages.** Lists return at most 200 records. `all_pages` follows every page up to 500 pages, stopping once `max_records` (default 500, up to 1,000,000) is reached. The page that crosses the limit is kept whole, so nothing is dropped. The result says whether it `complete`d and otherwise gives a `next_page` to resume from. If a rate limit or timeout hits part-way, the pages already fetched are returned, not thrown away.
- **Filters.** Populi ignores filter field names it does not know and returns the *whole list*, which looks like "everyone matches". `populi_describe_api` lists the valid condition names for a route; `populi_find_people` uses the right ones for people.
- **GET parameters.** `filter`, `expand`, `page` and `limit` travel in a `parameters` query value, because most HTTP clients cannot send a GET body. If that would make the URL longer than about 7 KB (for example a filter listing thousands of ids), the request is sent as a `GET` with a JSON body instead, which Populi documents as the normal form. `POST`/`PUT`/`DELETE` send a JSON body.
- **Large responses** from `populi_get` are truncated at `max_chars` (default 40,000, up to 1,000,000) with a note on how to narrow the request. Use `populi_download` to save the whole thing to a file instead.
- **File downloads.** Routes that return a file (PDF/XLS/CSV exports, transcripts, Data Slicer results, file downloads) are fetched with `populi_download`, which writes into the files folder. It never overwrites (an existing name gets `-1`, `-2`, ...), and only uses names, never paths, so it can't write outside the folder. `populi_get` describes a file response but can't show it.
- **File uploads.** `populi_write` takes `files`, each either a `path` inside the files folder or `content_base64`. They are sent as `multipart/form-data` with the file as `file` and the other fields as a JSON `parameters` part, which is how Populi's upload routes (`/people/{id}/upload_file`, `/people/{id}/update_profile_picture`, `/libraryresources/{id}/update_image`, `/libraryresources/{id}/digitalresources/upload`) expect them. As a safety cap, this server refuses files over 100 MB.
- **The files folder is a boundary.** A model can only read uploads from, and save downloads to, `POPULI_FILES_DIR`. Paths that leave it, including through symlinks, are refused, so a prompt-injected "upload my SSH key" has nothing to read. Point it at a broader folder only if you want that.
- **Timeouts.** Populi publishes no timeout, and reports such as Data Slicer can be slow, so the default is two minutes. Your MCP client may enforce its own tool timeout as well, which this server can't raise; check its settings if a long report is cut off.
- **Student data is sensitive.** Records can include SSNs, birthdates, grades and financial aid. The server never logs request or response bodies, and the key is never echoed in errors, but whatever a tool returns goes to your model provider. Use the narrowest key role and `expand` you need.
## Development
```bash
git clone https://github.com/kisasa/populi-mcp.git
cd populi-mcp
npm install
npm run build # compile to dist/
npm test # unit + in-memory MCP tests, no network
```
Try it against a real instance:
```bash
POPULI_BASE_URL=https://yourschool.populiweb.com POPULI_API_KEY=sk_sandbox_... \
npx @modelcontextprotocol/inspector node dist/index.js
```
### Refreshing the API catalog
`data/catalog.json` is scraped from the rendered API docs (Populi publishes no OpenAPI file):
```bash
npm run build:catalog
```
Review and commit the diff when Populi ships API changes.
### Releasing
Publishing is automated by `.github/workflows/publish.yml` on every push to `main`, and only acts when the version is new. To release:
1. Bump `version` in `package.json` and both `version` fields in `server.json`.
2. Merge to `main`. The workflow publishes to npm (with provenance), the [MCP Registry](https://registry.modelcontextprotocol.io) as `io.github.kisasa/populi-mcp`, and creates a GitHub release.
npm authentication uses [trusted publishing](https://docs.npmjs.com/trusted-publishers) if configured for this repo, otherwise the `NPM_TOKEN` repository secret.
## License
MIT. Not affiliated with Populi. "Populi" is a trademark of its owner.
TDQS
Scored across 6 tools
Each tool has a distinct role: status/diagnostics, catalog search, route description, generic GET, file-saving GET, and people lookup. The main overlaps are populi_download vs populi_get (both fetch a route) and populi_find_people vs populi_get with filters, but the descriptions draw clear boundaries (files/big exports vs inline reads; name/email shortcut vs generic filtering).
All names share the populi_ prefix and snake_case, with mostly verb-based patterns (search_api, describe_api, get, download, find_people). populi_status is a noun rather than a verb, a minor deviation in an otherwise coherent scheme.
Six tools is well-scoped for a generic API gateway: discovery, execution, and two convenience wrappers, with no redundant entries. Nothing feels padded or missing at the count level.
The read path is well covered (search catalog, describe route, GET, download, plus a people shortcut), but there are no write/mutation tools despite populi_status explicitly reporting whether writes are enabled, implying a capability the surface doesn't expose. An agent asked to create or update records would hit a dead end.