ARCHIVED 2026-09-05. Superseded by the Jarvis daemon in the dottie monorepo: https://github.com/jcdavis131/dottie/tree/main/apps/jarvisd (see docs/JARVIS_HARNESS_PLAN.md there for the rationale). No development happens here. Agents: do not install, run, or take tasks from this repo; work in dottie.

Agent Lasso

A local-first web app for configuring LLM agents, chatting with them, and scoring them on exam-style benchmarks. FastAPI backend, LangChain/LangGraph agents, Jinja + Tailwind single-page UI, SQLite persistence.

Status: experimental, not maintained since mid-2025. (The UI and some code refer to the app by its working name, "Silver Lasso".)

What it does

• Assemble agents from configurable LLM backends (OpenAI, Anthropic, Mistral, Groq) and a set of built-in tools: file operations, basic data analysis, calculator, current time, DuckDuckGo/SearXNG search, arXiv search, optional weather, and an interactive canvas (tools.py). The file_operations tool is sandboxed to a workspace/ directory (override with AGENT_LASSO_FILE_SANDBOX) so an agent steered by untrusted web content can't read or write files elsewhere on disk.
• Chat with agents over SSE streaming and keep conversations, agent configs, and scores in a local SQLite database (agent_log.db).
• Run YAML-defined benchmark exams (benchmark_exams/ — MMLU math, ARC, HellaSwag, Winogrande, and others) against agents via /api/benchmark/*, with a sortable leaderboard; an ExamBuilder agent can scaffold new exam files.
• Optional GraphRAG layer (graphrag_engine.py): Neo4j 5.x with vector and graph indices, or a zero-dependency in-memory JSON fallback, using Sentence-Transformers embeddings.

Quick start

python -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -r requirements.txt
uvicorn main:app --reload
# open http://127.0.0.1:8000

API keys are optional. Any KEY=value pairs in a git-ignored secrets.txt in the project root are loaded at runtime:

OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=...
NEO4J_PASSWORD=...

Configuration

File	Purpose
secrets.txt	API keys and passwords (git-ignored)
config.py	Model providers, embeddings, GraphRAG and tool settings
database.py	SQLite schema and helpers
vercel.json	Vercel Functions deployment config

Most settings can also be overridden with environment variables.

License

MIT