"Allama is an AI-native project"
So I've been messing around with local LLMs for a while now, and one thing kept nagging me — whenever a model decides it needs to call a tool, most clients just... run it. No heads-up, no confirmation, nothing. You're basically trusting the model to not do something weird, and that always felt off to me.
So I built Allama to scratch that itch.
It's a lightweight Electron desktop chat client, and the whole point is giving you the final say before any MCP tool actually executes. That's really the core idea everything else is built around.
What's in it right now:
The permission gating is the main thing — every tool call surfaces as a prompt you have to approve before anything runs. You get a live status indicator showing whether something is Pending, succeeded, errored out, or was denied by you. No silent background executions.
Beyond that:
Works with any OpenAI-compatible backend — Ollama, LM Studio, LocalAI, Groq, OpenAI, you name it. Switching is painless.
Auto-discovers and lists MCP tools from any server you add via the sidebar.
Image and file attachments that inject straight into context.
Dark mode UI, multiple threads, auto-generated titles, all history stored locally — nothing leaves your machine.
Built with vanilla JS/CSS, no heavy frontend frameworks. Keeps the RAM free for what actually matters.
Uses DOMPurify and strict IPC patterns. Security wasn't an afterthought here.
Where things stand:
Still very early — just a few commits in, no prebuilt binaries yet. You'll need to clone the repo and run npm install && npm start to try it. The permission gating and visual feedback are working though, which is the part I most wanted to validate first.
Repo: https://github.com/onedayallday33-a11y/allama
Screenshots and a proper release are coming soon. I took cues from Dive, Witsy, Open WebUI, and Claude Desktop for the UI feel, but wanted to push harder on the user-control side of things.
This is my first public project, so I'm genuinely curious what you all think:
Is the permission gating something you'd actually find useful day-to-day, or is auto-execute fine for your workflow?
Which MCP tools or integrations would you want to see supported first?
Anything obviously broken or missing for a v0.1?
Brutal feedback welcome. PRs even more so. Thanks for reading.
Top comments (0)