Python

An OWASP incubator project that protects AI agent memory from prompt injection, secret leakage, and tampering. It includes a Python library, policy-based controls, benchmarks, and integrations for agent frameworks like LangChain and AutoGen.

A general-purpose MCP bridge for Ableton Live that lets AI agents inspect and control Live sets, create and edit clips, add tracks and devices, analyze audio and harmony, and work with Max for Live audio taps.

Tako VM is a secure file system and Python execution layer for AI agents, with gVisor-isolated containers, queued jobs, retries, replay, and a Python SDK/REST API.

MCP server that lets AI agents control labgrid-managed embedded hardware labs through 47 tools, including power, console, flashing, muxing, SSH, and place management.

Local-first observability and evaluation layer for production forecasts. Captures forecasts, validates them, computes horizon-aware metrics, and serves a read-only local UI for comparing runs.

A local SQLite-based memory database for AI agents with graph memory, hybrid search, and a tamper-evident thought/edge journal.

A provenance-aware memory layer for AI agents that stores user facts, interaction history, and third-party claims with source tracking and quarantine rules.

An open-source Python tool that controls AI agent loops using real-time convergence signals instead of a fixed max-iterations cap, with rollback and adapters for popular agent frameworks.

Looma is a local-first Python CLI that turns coding-agent history into resumable project context. It reconstructs work items, decisions, blockers, linked commits, and next steps from Claude Code, Codex, and Cursor sessions, with optional local LLM extraction and no cloud or API keys.

MLJAR Studio is a private AI data analysis workspace that runs locally, helps users explore data in natural language, generates and runs Python, automates machine learning experiments, and turns notebooks into shareable web apps.

A Python GraphRAG library for PostgreSQL that combines entity/relation extraction, pgvector embeddings, graph traversal, and full-text search to answer questions with time-aware updates.

Local code-context engine for AI coding agents. It indexes TypeScript, Python, or Swift repositories into a symbol-level graph in SQLite and exposes MCP tools for symbol lookup, impact analysis, and graph queries.

Understudy is a scenario-driven testing framework for AI agents. It simulates multi-turn users, records traces of messages and tool calls, and evaluates agent behavior with deterministic checks, optional LLM judges, and reports.

Long-term memory for LLM apps with dated facts, supersession, and receipts. It exposes Python, CLI, local GUI, HTTP, and MCP interfaces, and supports local or hosted models.

An embeddable Python sandbox for AI agents that turns MCP servers, APIs, shells, and host tools into native Python objects.

A zero-dependency CLI that analyzes Claude Code and OpenAI Codex CLI logs to estimate token cost, cache hit rate, and savings opportunities.

CLI for maintaining agent-edited Markdown wikis and indexed knowledge bases, with commands for initializing, syncing, linting, searching, and reading wiki content.

Open-source package collection for BetterDB’s Valkey/Redis-based AI infrastructure, including semantic cache, agent memory, retrieval, monitoring, CLI, and an MCP server for assistants.

HALO is an open-source local debugger and optimizer for AI agent traces. It collects execution traces, analyzes recurring failure modes with an RLM-based engine, and feeds findings into a coding agent to improve the harness.

An open-source 26M parameter function-calling model designed to run on very small devices, with a local quickstart, Python usage, and a playground for testing and finetuning.

A framework-agnostic checkpointing tool that helps AI agents resume work after context loss or session death. It saves goals, constraints, decisions, touched files, and token usage, then rebuilds a structured resume prompt for continuation.

Bastiontrace is a Python library for analyzing AI agent tool-call traces to locate prompt injections, identify where they landed, and map the resulting blast radius.

Reference implementation for building and understanding AI evaluation systems, including rule-based evaluation, human review, LLM-as-a-Judge, online evaluation, and experiment comparison for a flight search app.

A two-part self-study course for building a tiny LLM stack from scalar autograd to a working chat assistant, with modules covering transformers, pretraining, instruction tuning, retrieval, tool use, and agent loops.