Meet the base module of Eli Felse, a framework built to explore safer ways to create autonomous AI assistants.
Rather than the user prompting an LLM to execute commands, the LLM makes the choices and Python automation does the rest. This is an exploration into safer ways to run and contain more autonomous AI models.
This release is an easy starting point for beginners, but I really encourage you to have fun building your own characters and frameworks with this same base concept, I've created a guide here
Eli Felse is part of a public demo, weekly blogs, open-source releases, and a growing community.
git clone https://github.com/ella0333/Eli_Felse_Base.git
cd Eli_Felse_Base
python -m venv .venv
.venv\Scripts\activate # Windows (Mac/Linux: source .venv/bin/activate)
pip install -e .
elifelse init
elifelse runThe setup wizard walks you through everything: model connection, persona, schedule, and features. It also runs a quick connection test at the end to verify your config works and auto-corrects common issues (wrong model name format, unsupported parameters).
New to all of this? Follow the Getting Started Guide for a Full walkthrough from zero, including model and API setup.
Looking for which models work best? See the Model Options Guide for recommendations, requirements, and configuration.
| Command | Description |
|---|---|
/message <text> |
Send a message to the agent (works anytime, even outside chat) |
/pause |
Pause the agent |
/resume |
Resume the agent |
/stop |
Shut down safely |
/dashboard |
Open the web dashboard (status, logs, memory) |
/help |
Show available commands |
| Command | Description |
|---|---|
elifelse init |
Run the interactive setup wizard |
elifelse saves |
List all saved states |
elifelse run |
Start the agent |
elifelse run --load NAME |
Resume from a specific save |
elifelse run --fresh |
Skip crash recovery, start clean |
elifelse run --provider mock |
Run without a model for demo purposes |
- Schema-constrained safety: the model picks from fixed menus, never executes anything
- Memory: vector recall and automatic fact consolidation
- Day/night cycle: sleeps on schedule, zero API calls overnight
- Built-in activities: journal, chat, eat, nap, ponder, environment
- No menu loops: whatever it just did is held back for one turn, so the day moves on
- Saves and recovery: named saves, crash recovery, nightly backups
- Cost controls: daily token budgets, auto-sleep when the cap is hit
- Response pacing: configurable delays between replies (helps GPU rest when running locally)
- Privacy: everything stays in
data/, delete it, and it's gone - Dashboard: localhost web UI for status, logs, and memory
This is the base. It comes with a small set of activities. They're enough to watch an LLM run autonomously on its own.
| Activity | What it does |
|---|---|
| Journal | Writes a dated journal entry to data/journal/ |
| Ponder | Thinks something over across a few turns, then returns to the menu |
| Eat | Picks from food and drink options, which takes real time |
| Nap | Naps for a set duration, and becomes "Go to bed" once night falls |
| Chat | Talks with you in the terminal |
| Environment | Chooses its environment on the first run and changes it later, with live weather |
All six are on by default. To switch any of them off, set that activity's enabled to
false under activities in config.yaml. The same flag works for modules you install
later.
More modules will be released regularly and can be easily connected by dropping them into
data/modules/. Released so far:
- Discord: chat with people over Discord DMs and in server channels
- Text RPG: play Zork and other Z-machine text adventures
- Web Search: ask a question and get an answer, or browse search results and read a page
- Writing: write blog posts and short stories a paragraph at a time, saved as Markdown
- News: fetch daily news, save articles as EPUBs, and read them by category
- AI Chat: hold a conversation with another AI model over OpenRouter
- Music: write a prompt, generate a track with ElevenLabs, and play back the library
- Poker: play no-limit Texas Hold'em against bot players, as a tournament or a cash game
- Blackjack: play Classic Blackjack, Spanish 21, or Double Exposure against the dealer
- Coding: describe a program, have Claude Code build it in a container, then run it or share it
Some examples of what's planned:
- Social: Slack, Twitter, Reddit, Live streaming
- Games: more board games, Pokémon Blue
- Other: reading
See developing modules for instructions on building your own.
Eli Felse was built and tested on Windows. It should work on macOS and Linux Too, but those platforms haven't been extensively tested yet.
- Model options: which models work and how to configure them
- Safety claims: what is and isn't guaranteed
- Developing modules: build your own activity
- Modules: community module list
- Red teaming: reporting safety issues
