public inbox for kaagum@systemreboot.net
help / color / mirror / Atom feed
* (no subject)
From: Enrico Schwass @ 1970-01-01 0:00 UTC (permalink / raw)
To: kaagum
Hello,
very nice project. Almost all other coding agents besides aider are node based and not reproducible. Thank you for starting this project.
I noticed one thing when using agent-shell and kaagum talking to vllm. vllm returns usage besides other attributes via prompt_tokens_details and if this is ‘null’ kaagum crashes.
I will try to fix it locally, but maybe it is worth a change upstream
Have a nice day
Enno
^ permalink raw reply
* (no subject)
From: Arun Isaac @ 1970-01-01 0:00 UTC (permalink / raw)
To: guile-user; +Cc: kaagum
Hi,
https://klaus.systemreboot.net/kaagum/
I would like to introduce kaagum, a tiny AI agent written in Guile with
minimal dependencies. It features a security-focused design with
sandboxing via Guix containers and the principle of least privilege for
both filesystem and network access. kaagum works with any
OpenAI-compatible API, including local models—no data ever needs to
leave your machine.
kaagum's core event loop is an implementation of The Elm Architecture
(TEA)—it is a purely functional state machine cleanly separating state
transitions from side-effects.
kaagum implements the Agent Client Protocol (ACP) and works with any
compatible client. Thus it composes well not only with Emacs but with
any editor or environment of your choice.
I realize AI is a contentious and polarizing topic in the community at
the moment. My thoughts on the matter are in the README.
If you have feedback to offer, or are interested in hacking on this
project, please write to the public inbox at kaagum@systemreboot.net
https://lists.systemreboot.net/kaagum/
Regards,
Arun
^ permalink raw reply
* (no subject)
From: Arun Isaac @ 1970-01-01 0:00 UTC (permalink / raw)
To: kaagum
Hello,
kaagum is a tiny security-focused AI agent written in Guile. Here are
some quick development highlights from last month.
Two big quality-of-life improvements have landed:
kaagum now enables automatic prompt caching with a TTL of 1 hour. On
repeated prompts that share a common prefix, which is typical of a
kaagum session, this can cut costs by up to 90% on the Anthropic API!
kaagum now reports the token usage, context length and session cost to
your ACP client. Keep an eye on your costs.
Internally, kaagum uses TEA (The Elm Architecture) to separate
side-effects from a purely functional core state update function. This
core has now been rewritten using the state monad, simplifying it
further.
https://klaus.systemreboot.net/kaagum/commit/03a0edf0e2f0df29afdded81824c7aa75745fa24/
Cheers!
Arun
^ permalink raw reply
* (no subject)
From: Arun Isaac @ 1970-01-01 0:00 UTC (permalink / raw)
To: kaagum
Hi everyone,
I am pleased to introduce kaagum 🐦⬛, a tiny AI agent written in Guile
with minimal dependencies. It began as a weekend project because I had
gotten tired of bloat and vulnerabilities in AI software. kaagum brings
care and craft, and thoughtfully designed security (through capabilities
and Guix containers to isolate tool calls) to AI agents. kaagum
implements the Agent Client Protocol (ACP) and works with any compatible
client. Thus it composes well not only with Emacs but with any editor or
environment of your choice.
I am concerned about the environmental and social impacts of using AI,
but at the same time am tempted by some of its possibilities. Please use
your own wisdom and judgment. This project is an experiment with AI. It
is not an endorsement of AI. Please see the README for more.
I'd love to collaborate with you. If you're interested in hacking on
this together, please look at the Brainstorming section of the README.
Let's use this public inbox for discussion.
Cheers!
Arun
^ permalink raw reply
* (no subject)
From: Arun Isaac @ 1970-01-01 0:00 UTC (permalink / raw)
To: kaagum
This is a test email.
^ permalink raw reply
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox