Give every agent its own database.
Ghost is hosted Postgres you spin up, fork, and throw away. One per task, production stays untouched, no blast radius.
curl -fsSL https://install.ghost.build | sh
curl -fsSL https://install.ghost.build | sh
- Agents need to explore.Coding agents need to be able to test various ideas and see what works.→Letting agents experiment unlocks their full potential.
- Experimentation needs to be safe.You can't mess up a production system because an agent did something stupid.→Agents can't touch your production database.
- Experiments need to work on real data.Otherwise you never know if the results are valid.→Ghost forking lets you duplicate your database quickly, at any size.
- Humans need to get out of the way.During experimentation, an agent that needs human approval for every operation moves too slow.→Isolated environments are the only solution.
- Most experiments will fail.They need to be cheap to run and cheap to throw away.→Both code and databases need to be disposable.
- Humans need to review the final result.For correctness, style, and taste.→Ghost lets you introspect the database and share it with others for review.
Give every agent its own Postgres environment.
Create databases in seconds for new tasks and experiments
Fork existing databases instantly before changes or testing
Discard environments freely when work is complete
Use CLI and MCP from day one for agent-native workflows
Set hard spending caps to avoid surprise usage
Built for real agent workflows.
Safe schema experimentation
Fork a database before your agent touches migrations, data transformations, or structural changes. Test freely without risking prod.
Database per task
Give each agent task its own isolated Postgres environment instead of sharing one fragile staging database.
Persistent memory for agents
Use real Postgres as long-lived memory for agent systems that need state, retrieval, history, and structure.
Rapid product prototyping
Spin up disposable databases for prototypes, internal tools, and experiments without infrastructure overhead.
Auto-pause for experiments. Always-on for production.
Charged only when something actively queries the database, in 15-minute chunks. Spin one up for a task, throw it away when done — built for agent runs and experimentation.
- Each instance: 2 GB RAM + 0.5 CPU burstable
- 100 hrs/mo free, then $0.075/CPU-hr
- 1 TB storage free, then $0.25/GB-mo
- Charged only when queried (15-min chunks)
- Unlimited databases & forks
- Shareable forks across accounts
Reserved Postgres for production workloads. Never pauses. Same platform as your on-demand forks — fork production into disposable copies without leaving the system.
- 2 GB RAM / 0.5 CPU, 10 GB storage included
- Always running — no auto-pause
- Scales to 16 GB RAM / 4 CPU ($10 per 2 GB step)
- $0.25 per GB storage above 10 GB
- Unlimited databases & forks
- Shareable forks across accounts
People using it already.
“We had three agents fighting for one staging database. Now each one gets its own, spun up in the time it takes the first tool call to return. The whole category of 'agent stepped on another agent's work' just disappeared.”
“We fork main before every migration. If it works on the fork, we swap. If it doesn't, we delete the fork. We haven't had a migration roll back in production since we moved to Ghost.”
“Databases used to be expensive enough that you reused them for everything. Now I create one for a single afternoon of work and throw it away at the end. It feels wasteful. It's not.”
“I stopped writing cleanup scripts. That alone paid for it.”
What's landed and what's in flight
MCP server
First-class tool surface for Claude Code, Cursor, Windsurf, Codex, Gemini CLI, VS Code, and more. One command to install: ghost mcp install.
Schema introspection
ghost schema displays your database structure in a format optimized for agents. Tables, views, constraints, indexes, all formatted so Claude can write accurate queries without guessing.
MCP prompts
Reusable guidance packages agents can pull on demand. Eight shipped so far, covering pgvector semantic search, hybrid text search, TimescaleDB setup, and Postgres schema design.
Read-only mode
A read_only config option prevents destructive operations like delete, pause, and sql writes. Give agents access to your databases without worrying about accidental data loss.
Drop into psql
ghost psql opens an interactive psql session connected to your database. No more copying connection strings.
Headless login
ghost login --headless uses OAuth device flow for environments without a browser. Sandboxes, VMs, containers, remote servers. Enter a code on any device and you're in.
pgvector + pg_textsearch baked in
No extensions to install. Forks of a vector-heavy db stay consistent. Agents can search on day one.
Scheduled deletes
Auto-delete a database after N minutes. For agents that forget to clean up after themselves.
Your first Ghost database in minutes.
Install Ghost
Get set up with the CLI and connect your workflow.
Create or fork a database
Start fresh or clone an existing environment for safe experimentation.
Use it with your agent
Connect through MCP or CLI and let your agent build against real Postgres.
Infrastructure for agent-native software starts here.
Ghost is the database layer built for how agents actually work: fast, parallel, disposable, and stateful.
curl -fsSL https://install.ghost.build | sh