Agent Backends

Agent Chat can run on four backends. They differ in what the agent can reach outside your vault. Here is an example. You ask the agent to read a log in your Downloads folder. Claude Code with Computer access reads the file. The built-in model connection cannot: it has the memrynote vault tools and nothing else.

This page lists what each backend can reach, how long a turn can run, and how its reasoning shows. For setup, see Agent Chat & MCP Server and Provider Setup.

What each backend can reach

The model menu in the prompt bar has an Access setting and a Web search switch:

• Vault only: the agent gets the memrynote vault tools. It reads and writes notes, tasks, journal entries and the rest of your vault, and every change asks you first unless you turned that off.
• Computer access: the agent also gets the backend's own tools, such as a shell and files outside the vault.
• Web search: the agent can search and fetch web pages for that turn.

Backend	Vault only	Computer access	Web search	Step limit set by memrynote	Reasoning control	Reasoning shown
Claude Code (claude)	Yes	Yes	Yes	None	Effort: low to max	Yes, in a "Thinking" block
Codex (codex)	Yes	Yes	Yes	None	Effort: low to extra high	Yes, reasoning summaries
Antigravity (agy)	Yes	Yes	Yes	None	Part of the model id	No
Built-in model connection	Yes	No	No	24 model calls	Default, High, Max	Yes, when the model streams it
External MCP clients (see below)	Read only	Their own	Their own	Their own	Their own	In their own window

Claude Code, Codex and Antigravity

These three run the CLI installed on your computer. memrynote starts it for each turn and connects it to the vault tools.

• Vault only turns get no shell and no file access outside the vault. Web search is added only when you turn it on.
• Computer access turns give the CLI its own tools, such as a shell and file edits, with access to your whole disk. Codex and Antigravity run them without asking. Claude Code runs without prompts too, so a tool its own permission settings do not allow is refused rather than asked about.
• memrynote sets no step limit. A turn runs until the CLI finishes or you press Stop.
• Claude Code and Codex take a reasoning effort from the prompt bar. Antigravity model ids carry their own tier, such as gemini-3.1-pro-high.

Built-in model connection

The built-in model connection talks to any OpenAI-compatible server: Ollama, LM Studio, llama.cpp server, or a hosted API such as DeepSeek through the Custom preset.

• It has the memrynote vault tools only: no shell, no files outside the vault, and no web search. With this backend selected, Computer access and Web search show as unavailable with the reason "Vault tools only with this model", and the turn runs vault-only. Your saved default does not change.
• Vault tools turn on only after a capability check passes for the model. If the check fails, the model answers from the attached context and the reply says that tools are off and why.
• A turn with tools can make up to 24 model calls. The last call runs without tools and asks the model to hand off what is done and what is left. The reply then ends with Stopped at the step limit and a Continue button, which sends "Continue" as a new message in the same conversation.
• Effort in the prompt bar sends reasoning_effort for High and Max; Default sends nothing. Thinking: Off in the connection settings sends thinking: {"type": "disabled"}. Both controls are hidden for the Ollama preset.
• Reasoning that the server returns as reasoning_content shows in a "Thinking" block above the answer.

There are no plans to give the built-in backend a shell, file or web tools. If you need a command line, files outside the vault or web search, use Claude Code, Codex or Antigravity with Computer access.

Driving memrynote from an external CLI agent

You can also run a CLI agent in your own terminal and point it at your vault. memrynote supports two setups.

Read through the MCP server

memrynote starts a local MCP server the first time you open Agent Chat, or the Agents or Connect tab of the AI Assistant settings, after a vault opens. Copy its URL and bearer token from Settings → AI Assistant → Agent MCP, then add them to your agent as an HTTP MCP server. For example, with Claude Code:

claude mcp add --transport http memrynote http://127.0.0.1:<port>/mcp \
  --header "Authorization: Bearer <token>"

With Codex, in ~/.codex/config.toml:

[mcp_servers.memrynote]
url = "http://127.0.0.1:<port>/mcp"
bearer_token_env_var = "MEMRYNOTE_MCP_TOKEN"

Then set MEMRYNOTE_MCP_TOKEN to the token in the shell that runs codex.

Limits of this setup:

• External clients are read-only. The write tools are listed, but a call fails with PERMISSION_DENIED, because writes need a turn that memrynote itself is running.
• The port and token change each time memrynote starts or switches vault, and when you rotate the token. Update your agent's configuration after that.
• vault_get_current_note returns null, because an external client has no memrynote window.

Write through the memrynote command

For changes, have the agent run the memrynote command line with --json, for example memrynote --json notes update <id> --append "...". The command writes to the vault directly. It does not go through Agent Chat's approval prompts, so give it only to an agent you trust with your vault.

Use Agent Chat instead

When you want both writes with approval and a shell, run Claude Code, Codex or Antigravity as the Agent Chat backend with Computer access. memrynote starts the CLI, hands it a write capability for that turn only, and shows each change for approval.