https://github.com/timescale/ghostgres
There are no dumb queries, only dumb databases
https://github.com/timescale/ghostgres
Last synced: about 1 month ago
JSON representation
There are no dumb queries, only dumb databases
- Host: GitHub
- URL: https://github.com/timescale/ghostgres
- Owner: timescale
- License: apache-2.0
- Created: 2026-03-26T16:55:50.000Z (4 months ago)
- Default Branch: main
- Last Pushed: 2026-04-23T00:31:24.000Z (3 months ago)
- Last Synced: 2026-05-26T14:46:16.690Z (about 2 months ago)
- Language: Go
- Homepage: https://ghostgres.com
- Size: 103 KB
- Stars: 39
- Watchers: 0
- Forks: 2
- Open Issues: 2
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Ghostgres
Are you tired of maintaining your relational database? Tired of having to write "syntactically correct" SQL queries? Tired of having queries fail because tables "don't exist"?
Look no further. [Ghostgres](https://ghostgres.com) is the database of the future. There's nothing to maintain. No query language. It doesn't return errors. The only limits on what you can query are the limits of your imagination.
The best part? It's Postgres wire-compatible, so you can plug it in wherever you use Postgres.
Just grab an API key for your favorite model provider and connect:
```bash
psql "postgres://openai:@try.ghostgres.com/gpt-5.4"
```
Then start querying:
```
gpt-5.4=> What is the best database?;
```
Just remember to end your queries with a semicolon (`;`).
Oh, and be careful with single (`'`) and double quotes (`"`).
**Built with <3 by [Ghost](https://ghost.build).**
## Connecting
Connect using `psql` or any Postgres client. The username is the LLM provider (`openai` or `anthropic`), the password is your API key, and the database is the model name.
```bash
psql "postgres://:@try.ghostgres.com/"
```
## Running Locally
You can run the Ghostgres server locally with:
```bash
go run github.com/timescale/ghostgres/cmd/ghostgres@latest
```
Then connect to localhost:
```bash
psql "postgres://openai:@localhost/gpt-5.4"
```
### Flags:
The server accepts the following flags:
| Flag | Default | Description |
|------|---------|-------------|
| `-host` | `""` (all interfaces) | Hostname/interface to bind to |
| `-port` | `5432` | Port to listen on |
| `-log-level` | `info` | Log level (`debug`, `info`, `warn`, `error`) |
| `-prompt` | Built-in prompt | Path to a file containing a custom system prompt |
| `-tls-cert` | (disabled) | Path to TLS certificate PEM file (requires `-tls-key`) |
| `-tls-key` | (disabled) | Path to TLS private key PEM file (requires `-tls-cert`) |
## Supported providers
### OpenAI
Username: `openai`.
Password: your [OpenAI API key](https://platform.openai.com/api-keys).
Database: any OpenAI model name (e.g. `gpt-5.4`, `gpt-4o`, `o3`).
**Options:**
| Option | Description | Default |
|--------|-------------|---------|
| `reasoning_effort` | Reasoning effort level (e.g. `none`, `minimal`, `low`, `medium`, `high`). Valid values depend on the model. | Lowest supported for the model |
### Anthropic
Username: `anthropic`.
Password: your [Anthropic API key](https://platform.claude.com/settings/keys).
Database: any Anthropic model name (e.g. `claude-sonnet-4-6`, `claude-opus-4-6`).
**Options:**
| Option | Description | Default |
|--------|-------------|---------|
| `effort` | Output effort level (`low`, `medium`, `high`, `max`) | `low` |
| `thinking` | Extended thinking budget in tokens (minimum 1024) | Disabled |
### Setting options
Options are passed via the `options` connection parameter (`PGOPTIONS` env var or `options=` in the connection string). Multiple options are space-separated.
```bash
psql "postgres://openai:sk-...@localhost/gpt-5.4?options=reasoning_effort%3Dhigh"
# Or:
PGOPTIONS="reasoning_effort=high" psql "postgres://openai:sk-...@localhost/gpt-5.4"
```
## How it works
Each client connection gets its own LLM chat session. The server translates Postgres wire protocol messages into LLM API calls and converts the structured JSON responses back into proper Postgres result sets. Chat history is maintained per connection, so the LLM can stay consistent across queries within a session.
## SSL/TLS
Ghostgres supports optional SSL/TLS encryption via the PostgreSQL wire protocol's `SSLRequest` handshake. When enabled, clients that request SSL (e.g. `sslmode=require`) will have their connections upgraded to TLS before any credentials are sent.
To enable TLS, provide a certificate and private key:
```bash
ghostgres -tls-cert /path/to/cert.pem -tls-key /path/to/key.pem
```
When TLS is not enabled, SSL requests from clients are denied and the connection continues in plaintext. Clients using `sslmode=prefer` (the default for most Postgres clients) will fall back to an unencrypted connection automatically.
## Security note
Authentication uses cleartext passwords because the server needs the raw API key to call the LLM provider. When running remotely, enable TLS (see above) to encrypt the connection. Without TLS, connect over localhost or use SSH tunneling.