Loach v1.5.0 works in your folders, follows LOACHFILE.md, remembers across chats and runs local MCP servers

[ztcs-software]

v1.5.0 is about letting the model do real work, and keeping you in charge while it does. Hand a chat a folder and the model can read, search and edit the files in it, with every change shown to you before it lands. A LOACHFILE.md tells it how things work in that folder, memory can now follow you from chat to chat, and MCP servers can run as local programs - with each tool call waiting for your say-so.

👉 Download Loach v1.5.0
or update directly from the app in Settings → Updates.

Work in a folder

Click the + next to the message box, choose Add directory and pick a folder. From then on the model can list, find, read and search the files in it, and write, edit, move or delete them - the way a coding assistant works, minus the shell. The message box reads Working in <folder> for the rest of the chat; click the name to open it in your file manager. Nothing outside the folder is reachable: absolute paths and .. are refused, and every link is checked again before it's used.

Before anything on disk changes, Loach asks - and shows you the change itself, not the raw tool call: the contents of a write, a line-by-line diff for an edit, from → to for a move, a red warning for a deletion. Answer Allow once, Allow <tool> for this chat or Deny; anything inside .git always asks, because a hook there runs commands the next time you use git. Files keep their line endings and their encoding, so a CRLF or UTF-16 file isn't quietly converted. One thing to keep in mind: with a cloud provider, the files the model reads are sent to it like any other message.

Instructions in a LOACHFILE.md

Drop a LOACHFILE.md in the root of that folder and its contents are added to every turn - the same idea as a CLAUDE.md or AGENTS.md. Put in what you'd tell a new collaborator: how to run the tests, what not to touch, the house style. It's read fresh every time, so an edit applies on your next message. Keep it short - local models follow brief instruction files far better. A folder you just cloned can say anything, so the model is told the file is the folder's text and not yours: it can't grant permissions, and your own instructions win when the two disagree. There's a sample to start from.

Global memories

Until now, memory lived inside Spaces. Turn on Global memories in Settings → Features (or in setup's Pick your defaults step) and Loach notes lasting facts about you from chats outside a Space too, then brings them into every chat, Spaces included. It's off by default, and worth a thought before you flip it: global facts go with every chat whatever provider it uses, so one learned talking to a local model also reaches a cloud one. Private Chat never reads or writes memory.

Memory that keeps up

Memory used to only ever add, so an old fact stayed right next to the one that replaced it. Now Space and global memory can update a fact that changed or drop one that no longer holds. Every change shows a toast with an Undo, and facts you added yourself are left alone. The memory editor also gained search, shows when each fact was saved and links back to the chat it came from.

Local MCP servers

Settings → MCP used to talk to servers over HTTP only. Most published MCP servers ship as a program you run, so Loach can now start them itself: give it a command (npx, uvx, node or a full path), its arguments and any environment variables. Running a program on your computer is a bigger deal than calling a URL, so before a new or changed server starts, a native dialog shows exactly what will run, argument by argument, with Cancel as the default button. Once approved, it starts with Loach without asking again. On macOS and Linux the server gets the same PATH your terminal sees, so tools installed through Homebrew or nvm are found. Environment variables often hold keys, so they're left out of exports.

Approve MCP tool calls

An MCP tool call now pauses the reply and shows you the server, the tool and its arguments. Allow once runs it this time, Always allow <tool> stops asking for that tool on that server, and Deny tells the model you declined so it carries on without it. This is on by default, which means servers you already had will start asking too. For a server you trust, switch Ask before each tool call off - its row then shows an Auto-approve badge, so the exception stays visible.

Other changes

• A Space's Files tab is now called Sources, so it isn't mixed up with the folder a chat works in.
• Forking a chat keeps its folder; exports never carry one, since a path means nothing on another machine.
• The context usage popover lists LOACHFILE.md on its own row.

Fixes

• Compacting a chat in a Space with its own instructions no longer loses the summary, and a compacted chat keeps following your global Custom instructions.
• Files dropped onto Private Chat no longer land in the chat behind it, and Private Chat now sends the settings its panel shows.
• /remember no longer saves a fact in a Space whose memory is switched off.
• Models pulled as hf.co/user/repo:tag can now be saved as new models from the Models editor.
• Running out of memory while loading a model now says so, and an API key withheld over plain http:// no longer looks like a wrong key.

👉 View the full changelog in GitHub Releases

Notes on previous minor releases

v1.4.1 was a round of fixes on top of v1.4.0: easier-to-read context menus, reply stats that no longer vanish when streaming stops, a setup that suggests a model that fits your free disk space, no stale OpenAI models after you remove your API key, and in-chat find whose highlights match its counter.