Agent Selma 👩🏻

A “Toy” Implementation of OpenClaw with Python

🌟 Overview

Agent Selma is a simplified, educational reimplementation of the OpenClaw project.

The primary goal is to deconstruct and understand the underlying architecture of autonomous agents by rebuilding them from scratch using Python. By stripping away complexity, Selma serves as a clean baseline for learning how agentic components interact.

[!WARNING] This is a toy project for learning purposes only. Selma is not hardened for production use. There are no security audits, no SLA guarantees, and no guarantees of stability or correctness. Do not use this in any production or commercial environment. Use at your own risk.

🛠 Tech Stack

🚀 Installation

Prerequisites: uv must be installed. If you don’t have it yet:

pip install uv

git clone https://github.com/YOUR_USERNAME/agent-selma.git
cd agent-selma
uv sync

Ollama (for local models) — download and install from ollama.com, then pull a model:

ollama pull llama3.2        # recommended default (~2 GB)
ollama pull qwen2.5:14b     # more capable, requires ~9 GB RAM

Ollama runs as a background service automatically after installation. Verify it is running:

ollama list

For the browser tool (Playwright), install Chromium once:

uv run playwright install chromium

⚙️ Configuration

Run the setup script once to create the .selma/ directory, generate a default selma.json, and deploy skills and templates into the workspace:

uv run setup.py

The script is safe to re-run — it never overwrites existing files.

The generated .selma/selma.json contains all available options with their defaults. The most important setting is the model:

"model": {
    "model": "ollama/llama3.1",
    ...
}

Use the format "provider/model-name", e.g.:

• "ollama/llama3.1" — local Ollama instance (default)
• "openai/gpt-4o" — OpenAI API (requires OPENAI_API_KEY in .env)
• "anthropic/claude-sonnet-4-6" — Anthropic API (requires ANTHROPIC_API_KEY in .env)

If the Telegram channel is enabled, add the bot token to a .env file in the project root:

TELEGRAM_TOKEN=...

The WebChat channel needs no additional credentials.

▶️ Running Selma

Start the gateway (FastAPI) and the dashboard (Streamlit) together:

./start.sh          # macOS / Linux
start.bat           # Windows

Or start them individually:

uv run gateway.py          # REST API on http://localhost:8000
uv run streamlit run dashboard.py

To restart only the gateway (e.g. after a code change):

./restart_gateway.sh    # macOS / Linux
restart_gateway.bat     # Windows

🖥 Dashboard

The Streamlit dashboard (dashboard.py) is the primary web interface for chatting with Selma.

Start it together with the gateway via ./start.sh, or standalone:

uv run streamlit run dashboard.py

The dashboard opens automatically in the browser at http://localhost:8501.

Features:

• Streaming chat — responses appear word-by-word as Selma generates them; active tool calls are shown inline while the agent works.
• Session continuity — each browser tab gets its own user ID, so multiple sessions run independently against the same gateway.
• Settings dialog — click ⚙️ in the sidebar to view or live-edit .selma/selma.json directly from the UI, with JSON validation before saving.

🔌 Channels / Adapters

WebChat

When channels.webchat.enabled is true, the gateway exposes a REST endpoint that any web client can use to chat with Selma.

Telegram

Set channels.telegram.enabled to true and provide TELEGRAM_TOKEN in .env. The bot listens to direct messages and group mentions (@BotName).

🧩 Skills

Skills extend Selma’s behaviour without touching core code. Each skill lives in its own folder:

skills/
  /
    SKILL.md        ← definition (YAML frontmatter + Markdown instructions)

Minimal SKILL.md:

---
name: my-skill
description: "One sentence describing when to trigger this skill."
user-invocable: true
---

# My Skill

## When to use
...

## Steps
1. ...

Built-in skills: summarize, web-research, blogwatcher, healthcheck.

🔭 Tracing & Monitoring

Selma uses OpenTelemetry to emit spans for every agent run and records all LLM calls via the OpenInference instrumentation. The traces are collected and visualised by Arize Phoenix — an open-source LLM observability tool that runs entirely locally.

Installation — Phoenix is already listed as a dependency and is installed automatically by uv sync. No separate account or cloud service is needed.

Starting the collector:

Phoenix is started automatically by start.sh / start.bat together with the gateway and dashboard. The UI is available at http://localhost:6006, logs go to phoenix.log.

To start Phoenix standalone (e.g. for debugging without the full stack):

./start_otel.sh     # macOS / Linux
start_otel.bat      # Windows

In the Phoenix UI you can inspect:

• every agent run as a trace with individual spans
• LLM input/output messages and token counts
• tool calls and their results
• Python log records attached to their span

Tracing is opt-in: if Phoenix is not running, Selma operates normally with a no-op tracer — nothing breaks.

🧪 Test Scripts

Scripts marked pytest are discovered and run by pytest. The others are standalone scripts that must be run directly — they require a running Ollama instance and a configured .selma/selma.json.

Run the pytest suite:

uv run pytest

Run a standalone script:

uv run test_agent.py

🎮 Related Projects

Agent Selma Game

agentselma-game — Play an Agent in an OpenClaw-like Architecture.

A browser-based game (self-contained HTML5, no build step) inspired by the C64 classic Elevator Action: you play Agent Selma, collect context documents from behind locked doors across eight floors, query a chat model and deliver the answers to the drones on the roof. Along the way the mechanics teach the same concepts this repository implements — context window management, debugging, channel routing and prompt-injection defense.

📅 Status

Version 1.0 — feature-complete first release (June 2026).

⚠️ Disclaimer

[!IMPORTANT] No Contributions Yet: At this stage, I am not accepting Pull Requests or changes. I am focusing on the initial build to establish the core learning path.

Communication: Feel free to reach out with questions or thoughts! However, please understand that due to time constraints, I may not be able to respond to every message personally.