mudler/LocalAI
LocalAI is an open-source AI engine that enables running various AI models (LLMs, vision, voice, image, video) on any hardware, including CPU-only systems, with a composable and extensible architec...
Awesome AI Agents › LLM Framework Tools
universal-tool-calling-protocol/code-mode
Code-Mode is a library for calling tools by executing code instead of emitting one structured tool call at a time. Its premise is that language models write code more reliably than they orchestrate long sequences of function calls, so rather than exposing hundreds of tool definitions to the model, it exposes a single tool that runs TypeScript with the whole registered toolkit in scope. A client is created, tool manuals are registered from a configuration, and a code block is passed to callToolChain, which returns the result together with captured logs. Because several dependent operations live in one block, a workflow that would take many round trips becomes a single execution; the README cites a third-party Python benchmark reporting iteration reductions from three to one for simple cases and sixteen to one for complex ones, and links to work from Apple, Cloudflare and Anthropic arguing for the same approach. Supporting features include progressive tool discovery, where a search call surfaces only the handful of relevant tools instead of the full catalog, auto-generated TypeScript interfaces and namespaces for every registered tool so the model sees typed signatures, and runtime introspection for adaptive workflows. Execution runs inside Node.js VM isolates with configurable timeouts, and tools are reachable only through registered UTCP or MCP servers. Four call template types are supported: MCP servers, REST APIs with automatic discovery, local JSON or YAML files, and command line tools. The package installs from npm, and a companion utcp CLI is recommended for shell-capable coding agents because it self-configures by writing a config file and can drive discovery, tool chains and interactive OAuth logins from the terminal; an MCP server is provided for MCP-only clients such as Claude Desktop.
https://github.com/universal-tool-calling-protocol/code-mode
Also in LLM Framework Tools
langfuse/langfuse
Langfuse is an open source LLM engineering platform that enables collaborative development, monitoring, evaluation, and debugging of AI applications with comprehensive observability, prompt management, and integration features.
567-labs/instructor
Instructor is a popular Python library that simplifies working with structured outputs from large language models by providing validation, retries, and streaming support across multiple LLM providers.
TEN-framework/ten-framework
TEN Framework is a real-time, distributed, cloud-edge collaborative multimodal AI agent framework supporting multiple programming languages for building high-performance AI applications.
google/adk-samples
A collection of sample agents built with the Agent Development Kit (ADK) to accelerate development of conversational and multi-agent workflows, designed for customization and integration with Google Cloud services.
truefoundry/cognita
Cognita is an open-source, modular, and production-ready RAG framework by TrueFoundry that simplifies building, customizing, and deploying scalable Retrieval Augmented Generation applications with a no-code UI and extensive integration features.
lemony-ai/cascadeflow
In-process Python and TypeScript library that cascades model calls inside the agent loop, routing each query or tool call between small and flagship models by quality, cost, latency and budget.
shroominic/codeinterpreter-api
An open-source API that implements the ChatGPT Code Interpreter for sandboxed Python code execution, enabling data analysis, visualization, and interactive AI-driven computational tasks with local execution and scalable production deployment.