Agent context and tools

Give agents native memory, instructions, credentials, and external tools.

Ara agents work best when they have the same context a teammate would need. Three primitives cover most of it.

Memory

Native Codex memory compounds per repository and is visible in Settings.

Secrets

Encrypted credentials injected into the sandbox. Values are write-only.

MCP servers

External tools the agent can call during a session.

What goes where

NeedUse
The agent should carry forward durable repository context.Native Codex memory
The sandbox needs a token or private registry key.Secret
The agent needs another system, database, or internal tool.MCP server

How they fit a session

A session starts from the prompt and repo. Ara then adds the context available to that run: the repository’s native Codex memory, scoped secrets, and enabled MCP tools.

Codex generates and compacts memory asynchronously. Inspect it in Settings → Memory; use Skills & Rules for instructions you want to author directly.

Secrets are different

Secret values can be set, overwritten, and deleted. They cannot be read back.

Use user scope for credentials tied to one operator. Use repo scope for credentials that should only be available on one repository.