ash #
ash is an ai chatbot for creatives.
this is quite the minimal demo, but right now it does two things well:
- talks to you with a really good personality & chats about your projects
- remembers what you're working on over the span of weeks via a memory wiki system
in the future, i hope to add:
- support for messaging platforms beyond telegram
- bash/web search/browser use tools
- composio support for pulling in context
- send_message revamp to send messages before tool usage + send multiple messages
- a "global wiki" shared by all ash users. this is the core of ash since it lets ash introduce you to people working on similar things
you can talk to ash on telegram, right now!
ash was built in bun, using vercel's AI SDK for a nice inference api, as well as the npm package typescript-telegram-bot-api for nice bindings to telegram's http api.
you can talk to the hosted ash instance using the link above - you'll still have your own memory wiki, and it won't be shared. but if you want to host ash yourself, you can do that!
how to deploy #
because ash is just bun, it's really easy to deploy. services like railway or fly.io should pick it up automatically if you connect a git repo. if you'd like to deploy on bare metal, though:
git clone https://tangled.org/hex4.xyz/ash && cd ash/ash-bot- make a telegram bot by messaging the @BotFather bot. copy the token it gives you
- make sure you've got some way to get inference (a /chat/completions endpoint). this could be HCAI, an OpenRouter account, or something else entirely! if you're not using openrouter or something openrouter-based, add CHAT_MODEL and LIBRARIAN_MODEL to your .env.
- add TELEGRAM_TOKEN, OPENROUTER_API_KEY, and OPENROUTER_BASE_URL (don't include the /chat/completions at the end!) to a .env file. bun will pick this up automatically.
bun run index.tsand you're off to the races!- write a systemd service or whatever idk
ai disclosure #
surprisingly, i used a (relatively :p) lower amount of ai on this project! i talked to a coding agent for help with designing the architecture, and used it to help with explaining things like sql. zed's code completions were also used.