Eliezer

A tiny, self-modifying AI agent. ~6K lines of TypeScript — small enough to fit in a context window, powerful enough to be useful.

Eliezer runs as a persistent background process. It talks to you through a chat interface, remembers conversations across sessions, works through task lists while you're away, and can edit its own source code. Bring your own LLM key — works with Anthropic, OpenAI, Kimi, Grok, DeepSeek, or any OpenAI-compatible provider.

What it does

• Runs shell commands, reads/writes/edits files
• Searches the web, reads webpages (headless Chrome)
• Downloads files with automatic security vetting
• Manages scheduled jobs and persistent task lists
• Compresses old conversations to stay within context limits
• Explores codebases using a sub-agent
• Modifies its own code, restarts, and picks up where it left off

Why it exists

Most agent frameworks are massive. Eliezer is the opposite — no abstractions, no plugins, no config DSLs. Just a main loop, some tools, and a database. The entire codebase fits in a context window many times over, so the agent can fully understand itself when it self-modifies. You can read it in an afternoon, then make it do whatever you want.

Quick start

git clone https://github.com/Eliezer-app/eliezer.git
cd eliezer
cp .env.example .env    # add your LLM API key
make dev                # builds, starts SearXNG + Eliezer

Eliezer listens on port 3200. Pair it with ClawChat, a companion chat PWA with push notifications, interactive widgets, and mobile support.

make logs    # follow logs
make status  # memory stats and health
make shell   # shell into the container
make stop    # stop everything

Configuration

All env vars in .env. No silent fallbacks — missing vars crash on startup.

Variable	Description
LLM_PROVIDER	anthropic or openai
LLM_API_KEY	Your API key
LLM_BASE_URL	API endpoint
LLM_MODEL	Model name
CONTEXT_WINDOW	Max tokens for the model
SEARCH_URL	SearXNG instance URL
CHAT_URL	Chat server URL
USER_TZ	Your timezone

Use COMPACTION_LLM_* to run compaction on a cheaper model.

Prompts

prompts/
  system.md       — agent identity and rules
  user.md         — user persona and preferences
  memory.md       — persistent scratchpad (agent-writable)
  compaction.md   — how to summarize conversations
  widgets.md      — widget/app building guidelines

Copied from prompts-default/ on first run, gitignored. Yours to edit.

Hacking

make test    # typecheck + integration tests (Docker)
make clean   # tear down everything

eliezer.mts      — main loop, tool dispatch
llm.mts          — LLM providers (Anthropic, OpenAI-compatible)
tools.mts        — exec, read, write, edit, wget, schedule, etc.
memory.mts       — conversation storage and context assembly
compaction.mts   — memory compaction engine
agent-tool.mts   — sub-agent framework
server.mts       — HTTP API

Adding a tool is ~30 lines: extend ToolBase, add it to the array in eliezer.mts.

Links

www.eliezer.app

License

MIT — Victor Dramba