realtime
The realtime skill enables WebSocket subscriptions for live database changes filtered by row-level security policies. Use it to stream INSERT/UPDATE/DELETE events to connected clients, ensuring each user receives only rows they can read according to their RLS permissions. Prerequisites include an existing table with a primary key and configured RLS policies to prevent unintended data exposure.
git clone --depth 1 https://github.com/butterbase-ai/butterbase-skills /tmp/realtime && cp -r /tmp/realtime/skills/realtime ~/.claude/skills/realtimeSKILL.md
# Butterbase Realtime
Live database change notifications over WebSocket, with per-row RLS enforcement. Once a table is enabled, INSERT/UPDATE/DELETE events stream to subscribed clients **filtered by the same RLS policies** that gate reads.
One tool: **`manage_realtime`** with two actions: `configure` and `get`.
---
## 1. The mental model
```
Postgres (data plane) Control API Browser
───────────────────── ──────────── ────────
INSERT/UPDATE/DELETE ──trigger──► realtime.changes ──WAL listener─► WebSocket ──► client
│
└── RLS check per (role, user) ──► filter rows
```
When you `configure` a table:
1. A Postgres trigger is installed → writes every change to `realtime.changes`.
2. A LISTEN connection in `RealtimeManager` reads those changes.
3. For each connected client, the change is RLS-checked **as that user** before broadcasting.
4. Clients only receive events for rows they could read with a regular SELECT.
---
## 2. Prerequisites
Before calling `configure`, the table must:
1. **Exist.** Run `manage_schema` (`action: "apply"`) first. Realtime won't auto-create it.
2. **Have RLS configured (if you care about isolation).** Realtime respects whatever policies exist via `manage_rls`. **No policies = all events flow to all users of that role.** This is the #1 silent leak.
3. **Have a primary key.** RLS checks query by PK. Tables without one can't be realtime-enabled cleanly.
---
## 3. Configure tables
```js
manage_realtime({
app_id: "app_abc123",
action: "configure",
tables: ["messages", "presence", "documents"]
})
// → [{ table: "messages", status: "enabled" }, ...]
```
- Idempotent — already-enabled tables are skipped.
- All three events (INSERT / UPDATE / DELETE) are enabled together; per-event filtering happens client-side via subscription `filter`.
- Validation: every named table must exist or you get `VALIDATION_TABLE_NOT_FOUND`.
### Inspect current state
```js
manage_realtime({ app_id: "app_abc123", action: "get" })
// → {
// tables: [{ table_name, enabled, trigger_installed, drift, created_at, updated_at }, ...],
// active_connection: true,
// websocket_url: "wss://api.butterbase.dev/v1/app_abc123/realtime"
// }
```
`drift: true` means the control-plane config says enabled but the data-plane trigger is missing — typically after a schema migration that dropped/recreated the table. Re-run `configure` to repair.
---
## 4. Connect from a client
```
wss://api.butterbase.dev/v1/{app_id}/realtime?token={JWT_or_API_KEY}
```
Browsers can't set custom headers on WebSocket upgrade, so the JWT goes in the query string. Server clients can use `Authorization: Bearer ...` instead.
```js
const ws = new WebSocket(
`wss://api.butterbase.dev/v1/${appId}/realtime?token=${userJwt}`
);
ws.onopen = () => {
ws.send(JSON.stringify({ type: "subscribe", table: "messages" }));
};
ws.onmessage = (e) => {
const msg = JSON.parse(e.data);
if (msg.type === "change") handleChange(msg); // { type, table, op, record, old_record, timestamp }
};
```
The Butterbase SDK wraps this:
```ts
const realtime = client.realtime(appId, userJwt);
realtime.subscribe("messages", (change) => console.log(change.op, change.record));
```
### Welcome and protocol
On connect, the server sends:
```json
{ "type": "connected", "app_id": "app_abc123", "role": "butterbase_user" }
```
Then a heartbeat every 30s:
```json
{ "type": "heartbeat", "timestamp": "..." }
```
### Client → server messages
| Type | Body | Purpose |
|------|------|---------|
| `subscribe` | `{ table, filter? }` | Subscribe to changes; optional client-side filter `{ col: value }` |
| `unsubscribe` | `{ table }` | Stop receiving |
| `presence_track` | `{ metadata }` | Announce yourself with arbitrary metadata (cursor, status) |
| `event` | `{ event, payload }` | Trigger a function with `trigger: { type: "websocket", config: { event } }` |
### Server → client messages
| Type | Body |
|------|------|
| `change` | `{ table, op: "INSERT"\|"UPDATE"\|"DELETE", record, old_record, timestamp }` |
| `presence_state` | `{ clients: [{ client_id, user_id, metadata }] }` |
| `heartbeat` | `{ timestamp }` |
---
## 5. RLS enforcement (the critical pitfall)
For each broadcast, the server runs (roughly):
```sql
SET LOCAL ROLE butterbase_user;
SET LOCAL request.jwt.claim.sub = '{user_id}';
SELECT 1 FROM "{table}" WHERE "{pk}" = {record_pk} LIMIT 1;
```
If the row is **not visible** under RLS, the change is **silently dropped** for that client. There is no error.
**Common consequences:**
- Client connects, sees `connected`, subscribes — but receives no events. → RLS too restrictive (or no policies at all + access mode `authenticated`).
- Client receives some events but not others. → RLS works for those rows; others are filtered out (often correct).
- Service key clients see everything. → Service bypasses RLS. Don't use this to "verify realtime works" if testing user-scoped behaviour.
**Always test with a real end-user JWT**, not the service key.
### Anonymous clients
If `manage_app` access mode is `authenticated`, anonymous WebSocket connections are rejected with close code `1008 (Policy Violation)`. To allow anon, the app must be in `public` mode AND the table must have a permissive policy for `butterbase_anon`.
---
## 6. Connection lifecycle
| Close code | Meaning |
|------------|---------|
| `1008` | App requires authentication, no token provided |
| `1013` (try again later) | Plan limit hit (`maxRealtimeListenersPerApp`) — upgrade |
| `1013` ("Realtime disabled by plan") | Free / starter tiers may have realtime off entirely |
| Normal close | Heartbeat missed, client disconnected, or server eviction |
The server caches table primary keys for 60s and batches RLS checks per `(role, user)` group, so connection cost is amortised.
---
## 7. Common patterns
### LiveClaude Code plugin for Butterbase — 30+ guided skills and auto-configured MCP for the AI-native backend-as-a-service.
Use when calling the app's AI gateway from agent tools — chat completions, embeddings, listing models, configuring defaults or BYOK, reading token/cost usage
Configure OAuth providers, auth hooks, JWT lifetimes, and service keys for a Butterbase app
Use when building a new Butterbase app from scratch, creating a full-stack application, or when the user asks to set up a complete backend with database, auth, and deployment
Use when users report access denied errors, see wrong data, RLS policies are not working, or when troubleshooting Row-Level Security issues in Butterbase
Deploy a frontend (React, Next.js, or static HTML) to a live URL on Butterbase
Use when building stateful per-key actors — chat rooms, multiplayer rooms, rate limiters, long-running agents, leaderboards — that need persistent in-memory + storage state across requests
Develop, deploy, or debug a Butterbase serverless function