MCP servers

Describes Hatoba 0.8.6.

Settings → AI → MCP Servers: adding servers, Always ask, import and export, and each server’s state and tools. The panel’s MCP Tools and /mcp, MCP prompts as / commands and MCP resources under @ are in AI assistant panel, how MCP calls ask in AI tools, permission modes and approvals, and errors in Troubleshooting: AI assistant, search, skills and MCP.

MCP (Model Context Protocol) servers add tools to the assistant, and prompts and resources that you attach yourself: a prompt runs as a / command, and a resource is picked after @. Resource templates are not listed. Two kinds: Command (stdio: a program started on this device with the user’s privileges, which Hatoba talks to over standard input and output) and HTTP (a Streamable HTTP URL). A stdio server runs arbitrary code, so add only servers you trust. Tool descriptions reach the model as instructions.

Other labels: Import MCP Servers.

  • Add Server: give it a Name and choose the Type. For Command: the program (looked up on PATH like a shell does, so npx finds npx.cmd on Windows), the Arguments (one per row; if a whole command line is pasted into the command, Split Into Arguments separates them) and Environment variables. For HTTP: the URL (HTTPS or HTTP) and Headers such as Authorization. Environment and header values are stored encrypted, can’t be viewed after saving, and only the program or the server receives them.
  • Add from Template fills in Add MCP Server for GitHub, Linear, Context7 or Cloudflare (HTTP servers that take a token, built into the app): the name, the URL and the Authorization header, with a link under it to where the token is created. Paste only the token: an Authorization value without a space is saved with Bearer in front of it (for any HTTP server, also when it replaces a saved value). Always ask starts off. The saved server is an ordinary one.
  • Always ask makes this server’s tools ask before every call, even in Bypass mode. It syncs with the server.
  • Import reads pasted JSON or a JSON file in the mcpServers format of Claude Desktop, Claude Code and Cursor (or VS Code’s servers key), shows a Preview, and imports; values move into the vault. Export writes the same format with placeholders instead of the secret values.
  • The server list syncs. Whether a server is on, and any Always allow, stay on each device. A server added on another device arrives on if it is HTTP and off if it is a Command. Editing a server, here or on another device, keeps whether it is on and its Always allow, except that an HTTP server changed into a Command on another device turns off here, as a new Command server arrives.
  • A Command server starts when a conversation first needs its tools and stops when the app quits or the vault locks. Click a row to see its state, start, restart or stop it, read the error with the last lines of stderr, and see its tools with their hints (Read-only, Destructive, Idempotent, Open-world). Hints never change whether a call asks.

Other labels: Stopped, Starting…, Running, Failed, Off on this device, Start, Restart, Stop, Last lines of stderr.

  • In manual mode every MCP tool asks for approval. Always allow all tools (per server) or Always allow (per tool) lets calls run without asking on this device. OAuth sign-in for MCP servers is not supported.

Edit this page on GitHub