CoreBase

Description

CoreBase is a platform for building and shipping governed AI agents on your own data. It gives an AI agent real access to real systems without giving it unlimited access, with permissions scoped per caller, human approval where it matters, and every call on the record. Agents can query, reason, and act over your databases, REST and GraphQL APIs, and 50+ SaaS apps.

You build with an OpenAI-compatible chat API, so any existing OpenAI SDK works by pointing its base URL at CoreBase, or you drop in an embeddable chat widget to ship a governed assistant inside your own product. Conversation rails enforce allow, require-approval, or deny on every tool call rather than merely suggesting them in a prompt, and an approval gate holds the exact call with its arguments until a person approves it, with an audit trail of executed and blocked calls alike.

CoreBase is read-only by default, with per-source write governance, per-tenant isolation via PostgreSQL Row-Level Security and AES-GCM encryption keys, per-end-user rate limits via signed JWTs, and a full audit log. It reads the source of truth on every call with no ETL or embedding pipeline. Live connectors include SQL Server (2000+), PostgreSQL, REST and GraphQL APIs, Slack, Microsoft 365, and 50+ SaaS apps such as Jira, Gmail, GitHub, Notion, and HubSpot via managed OAuth. The open-source CoreMCP bridge reaches on-prem and air-gapped systems over outbound 443 only, with no inbound ports. Deployment options span cloud-direct, cloud-relay, self-hosted, and a fully air-gapped enterprise container, and CoreBase never trains models on your data. Free, Pro, and Team plans are available with Enterprise custom-priced, starting free with no credit card.

CoreBase's Core Features

• OpenAI-compatible chat API and embeddable chat widget

• Agents that query and act on databases, REST/GraphQL APIs, and 50+ apps

• Conversation rails: allow, require-approval, or deny per tool call

• Human approval gates bound to the exact held call

• Read-only by default with per-source write governance

• Per-tenant isolation via PostgreSQL Row-Level Security and full audit log

• Open-source CoreMCP bridge for on-prem and air-gapped systems

• Cloud, cloud-relay, self-hosted, and air-gapped deployment options

How to use CoreBase?

1. 1

   Get an API key: Sign up with no card and create a scoped cb_live_ key in the panel.

2. 2

   Connect a source: Add a database, API, or SaaS app; use CoreMCP for on-prem or air-gapped systems.

3. 3

   Set the rails: Define which sources and actions are allowed, require approval, or are denied.

4. 4

   Build or embed: Point your OpenAI SDK at CoreBase or drop in the chat widget.

5. 5

   Query and act: Let the agent answer questions and take approved actions, with every call audited.

CoreBase's Use Cases

• Governed data agents
• In-product assistant
• On-prem and air-gapped access
• Approval-gated actions
• Automated workflows

FAQ from CoreBase

• Can CoreBase reach an on-prem or air-gapped database?

  Yes. You run the open-source CoreMCP agent inside your network and it dials out over 443 only, with no inbound ports, VPN, or exposed database. In air-gap mode, rows never leave your network.

• Can agents take actions or only answer questions?

  They can act. An agent can write to a connected API or database where you allow it, inside rails you define, and actions that need approval are held with the exact arguments they would run with until a person approves them.

• Can an agent change or break our data?

  Databases are read-only at the session, so mutating SQL is rejected by the database itself. For APIs and SaaS apps, write actions are off by default and enabled per connection, and every call is scoped, timed, result-size capped, and audited.

• How is data isolated between customers?

  Per-tenant by default, with each tenant having its own AES-GCM encryption keys and PostgreSQL row-level security enforced at the database, making cross-tenant access structurally impossible.

• Do you train models on our data?

  No. CoreBase never trains any model on your schemas, queries, or results, and with bring-your-own-key, inference runs through your own LLM provider account.

• Can I use my existing OpenAI client?

  Yes. The chat endpoint is OpenAI-compatible, so any OpenAI SDK works by pointing its base URL at the CoreBase API and passing your cb_live_ key. Official Node and Python SDKs are also rolling out.