Skip to main content

AI Agents

An AI Agent is a managed deployment that runs next to your data and connects to an ekoDB database. It can answer questions about your collections, build backends from a template or from scratch, run multi-step workflows, and keep your database tuned, all without you wiring up a separate service.

Part of the platform, not bolted on

The agent is part of ekoDB: you deploy it from the same dashboard, and it uses your database's built-in RAG and vector search — there is no second vector database, no separate RAG pipeline, and no glue code to maintain. An agent always connects to a database; a database runs perfectly well on its own.

How it runs​

ekoDB runs in the cloud, and so does the agent. From the dashboard you deploy a database and, separately, an AI Agent — each is its own managed deployment. The agent connects to a database you've deployed (you pick which one), and both run as managed services. You never host or operate them yourself.

Once an agent is deployed, you can work with it three ways:

  • The dashboard, in your browser, with a full chat and management UI.
  • The desktop app, a native client for macOS, Linux, and Windows.
  • The terminal client, an interactive agent in your shell.

The desktop app and the terminal client are thin clients. They authenticate, discover your deployments, and connect to your cloud agent. They never run the backend or hold your API keys. See Get Started for the client install commands.

What an agent can do​

  • Chat with your data. Ask questions in plain English. The agent answers from your own collections using ekoDB's native RAG and vector search.
  • Build your backend. The agent assembles collections, server-side functions, and scheduled jobs. It can start from ready-made templates for common backends (APIs, data models, auth, automation), or build the whole thing from scratch when no template fits what you need.
  • Run real workflows. Give an agent a system prompt, scoped tool permissions, goals, and scheduled tasks that run on their own.
  • Use built-in tools. File operations, web search, HTTP fetch, persistent memory, and reusable templates all come built in.
  • Tune and debug itself. The agent can explain a slow query, spot a missing index and create it, surface which data is hot, and read server logs to find problems.

Agents for you, or for your users​

An agent can serve two audiences:

  • Internal. Run it for yourself and your team, from the dashboard or a client.
  • Customer-facing. Expose an agent through a public customer endpoint so your own users can chat with it. You control CORS allowlists, anonymous access, rate limits, and token lifetime. A customer token can never escalate past the agent's permission ceiling, so it is safe to put in front of the public.

For customer-facing agents, prefer stored functions over raw read/write tools so behavior stays auditable and versioned.

Models​

Bring your own model key. The agent supports the tool-calling providers: OpenAI, Anthropic (Claude), and Google (Gemini). Your key stays server-side and is never exposed to clients.

Concepts​

  • Agent. A scoped assistant with its own system prompt, allowed collections, tool permissions, and model. You can run many agents on one deployment.
  • Goal. A higher-level objective the agent works toward across multiple steps.
  • Task. A unit of work, including scheduled tasks that run on a cron-like cadence without you in the loop.
  • Tools. The capabilities an agent can invoke: querying data, calling stored functions, file and web operations, and more. Permissions are per-agent, so you can mark write-capable tools as confirm-only or deny. Irreversible operations — deleting, moving, or copying files and running shell commands — always ask for a per-call confirmation in the interactive client, even when the tool is otherwise allowed, so a single broad "allow always" can never be reused to silently delete or overwrite something later.
  • Memory. Persistent context the agent carries across sessions.
  • Templates. Complete backend recipes (collections plus functions plus schedules) the agent applies in one step.

Next steps​