AI that starts where your data lives.
ASK, embeddings, vector search and MCP ship in the RedDB engine today. Around it we are building models, agent memory, artifact storage and managed agents — each labelled with its real status.
Managed AI services are planned. The engine features are available now; we onboard early users of the rest by email.
AI is already in the database.
No extra service to run. Embeddings, retrieval and grounded answers are query features of the same engine that stores your rows, documents, graphs and vectors.
- ASK, with citations
- Retrieves context across tables, documents, vectors, graph edges and key-value config, then answers with [^N] citations tied to the source records. Factual questions are planned as read-only queries that run under row-level security.
- Embeddings on write
- An EMBED policy on a collection embeds the declared fields asynchronously over CDC, so writes never wait on the provider.
- Vector and hybrid search
- SEARCH SIMILAR, hybrid text + vector search, and HNSW indexes that apply filters before ranking.
- MCP for agents
- Agents read and write durable state in RedDB through the Model Context Protocol, over stdio.
CREATE TABLE articles (id INT, title TEXT, body TEXT)
WITH (
EMBED (fields = ('title', 'body'), provider = 'openai', model = 'text-embedding-3-small')
);
SEARCH SIMILAR TEXT 'suspicious login' COLLECTION logs USING openai;
ASK 'who owns passport AB1234567?' USING groq;
EXPLAIN ASK 'who owns passport AB1234567?';
Providers you can configure
Bring your own provider key; keys are stored in the encrypted vault. Without a configured provider, ASK returns an error — the engine never substitutes a managed key.
Source: engine query docs, as of 2026-10-04.
RedDB Models
Choose the model. Keep your application.
One OpenAI-compatible endpoint and one key for the models we carry — usable from your application, from RedDB itself and from our agents.
- 01
One endpoint, one key
Call chat and embedding models through a familiar OpenAI-compatible API.
- 02
Plugged into RedDB
A database without its own provider can use RedDB Models for ASK and embeddings. Bringing your own key stays first-class.
- 03
One balance across products
Redcode, Agent Memory and the coding agents are planned to draw on the same token balance.
You would pay for
- Input, cached-input and output tokens, per model
- Usage drawn from prepaid credit, with the balance as the spending cap
Not promised yet
- The supported model catalog and rates will be published at launch.
- Not available yet; the database works with your own provider keys today.
Source: planning record, as of 2026-09-24.
Agent Memory
Give the next session somewhere to start.
Managed long-term memory for your agents: messages go in, durable memories come out, and the next session starts with the context that matters.
- 01
Write messages
Your agent sends the conversation. One extraction call per write turns it into memories.
- 02
Store in RedDB
Memories are embedded and kept in RedDB alongside their graph and vectors.
- 03
Recall by meaning
A query is embedded and matched with vector search, returning the memories relevant to the new session.
You would pay for
- Memory tokens (extraction and embeddings)
- Retained memory storage, per GB-month
- Not metered: retrievals, agents, end users or seats
Not promised yet
- The API shape, retention controls and bring-your-own-key options are still being decided.
Source: planning record, as of 2026-09-24.
AI Storage
Keep the work your agents create.
A place for what agents produce. Publish a plan, a design or a report from an agent run and share it with a link.
- 01
Publish from the agent
A plan from a coding agent, a design prototype or an operations report is saved as an artifact.
- 02
Kept in object storage
Artifacts live independently of the chat session that created them.
- 03
Share a URL
Each published artifact gets a link you can send to your team.
You would pay for
- Stored GB-month
- Writes and reads
- Not billed: egress, links, viewers or seats
Not promised yet
- Planned as a convenience inside our agents, not a standalone storage product.
- Private and expiring links are not built yet.
Source: planning record, as of 2026-09-24.
Agent Hosting
A home for the agents you choose.
We plan to provision, operate, upgrade and recover a persistent environment for the agent runtime you choose — so it keeps running when your laptop is closed.
- 01
A dedicated environment
Starts as a dedicated VM with persistent state. Tools run in a separate sandbox.
- 02
Pinned and upgraded
We pin the runtime version, apply upgrades deliberately, back it up and restore it.
- 03
Your models, your channels
Bring your own model keys, or use RedDB Models once that adapter is validated. Nothing is configured on your behalf.
Cost drivers
- Persistent compute, disk and network
- Tool sandbox time
- Model consumption and backups
Not promised yet
- Idle suspension, GPUs and support for every channel are not promised.
- Paperclip keeps its own PostgreSQL database.
- Each runtime needs isolation and recovery qualified before onboarding.
Source: planning record, as of 2026-10-04.
Status
What you can use today, and what is next.
Start with the data
Start where the data lives.
Create a free nano database, configure a provider and run your first ASK. Tell us which planned service you need, and we will talk about fit before anything is deployed.