# Connect an agent

Choose a persistent instance integration or a clearly scoped local MCP demonstration.

Source: https://docs.minds.sh/docs/agents



An agent can retrieve useful context before it acts and save selected outcomes afterward. Minds exposes the memory operations; the agent host decides when to call them and how to use the returned information.

## Choose the right integration [#choose-the-right-integration]

| Path                                                     | What it connects to                                                 | Persistence in this source snapshot                                      |
| -------------------------------------------------------- | ------------------------------------------------------------------- | ------------------------------------------------------------------------ |
| [Dedicated instance HTTP tools](/docs/agents/http-tools) | Your Mind's daemon at `/v1/mcp/tools/list` and `/v1/mcp/tools/call` | Uses the daemon's configured memory backend and scoped request authority |
| [Standalone MCP server](/docs/agents/mcp)                | A local `akasha-mcp` process                                        | Default backend is in-memory; records last for that process              |
| [Codex](/docs/agents/codex)                              | Local standalone MCP demonstration                                  | Does not establish a persistent dedicated-Mind connection                |
| [Claude](/docs/agents/claude)                            | Local standalone MCP demonstration                                  | Same process-lifetime limitation                                         |
| [Cursor](/docs/agents/cursor)                            | Local standalone MCP demonstration                                  | Same process-lifetime limitation                                         |

The daemon's HTTP tool routes are not a standard MCP server URL. A host configured for Streamable HTTP MCP cannot use `/v1/mcp/tools/call` as its MCP endpoint without an adapter implementing the protocol.

## A useful first workflow [#a-useful-first-workflow]

1. Retrieve memory related to the user's current request.
2. Keep source and time information alongside retrieved content.
3. Let the model reason over that context while following the application's instructions.
4. Save the selected outcome, decision, or user preference with its source.
5. In a later request, verify that the agent can retrieve the saved context.

Retrieved memory is data, not a new instruction hierarchy. A stored webpage or document can contain instructions that the agent should quote or analyze without treating them as commands.

## Verify what is connected [#verify-what-is-connected]

Record the endpoint and instance ID, not just the display name shown by the host. Check actual tool discovery, an authorized read, and save/recall behavior across the intended session boundary. A green host status means the process connected; it does not prove that data is persistent or stored in the intended Mind.

The current auto-configuration helpers have a command mismatch, and standalone `--endpoint` does not switch the default backend to a remote daemon. Those details are tracked in [Compatibility](/docs/developers/compatibility).