Recently updated open-source projects

LangGraph is a low-level orchestration framework for long-running, stateful workflows and agents. It supports durable execution that resumes after failures, persistent state, and human-in-the-loop control for tasks with branches, loops, or extended runtimes. Developers can use it independently or combine it with models, tools, and observability services from the broader LangChain ecosystem.

LangChain provides common interfaces and modular components for building agents and other LLM-powered applications. Developers can connect models, embeddings, vector stores, retrievers, and third-party tools while swapping providers without rebuilding the whole application. The core framework works on its own; LangGraph handles lower-level orchestration, while LangSmith products support debugging, evaluation, and deployment workflows.

Shadow signs each credit decision with Ed25519 and chains attestations with SHA-256 so a later verifier can detect silent changes to the verdict, reason codes, model manifest, or committed inputs and outputs. The verdict engine uses deterministic rules, while LLM personas provide advisory rationale that cannot alter the decision. The project sends no telemetry and explicitly remains pre-1.0 rather than a compliance product. It is a research layer for lenders building an auditable decision narrative.

OpenCode is an open-source coding agent for the terminal. Its built-in build agent handles development work, the plan agent is read-only for analysis and exploration, and a general subagent supports complex searches and multistep tasks. The project also offers a beta desktop app for macOS, Windows, and Linux, so users can choose a CLI or desktop workflow.

Anthropic Skills collects Claude skill examples, the Agent Skills specification, and a starter template for document, development, enterprise, communication, and creative tasks. The repository can be registered as a Claude Code plugin marketplace, and its example skills can also be used through Claude.ai or the API. A custom skill is a directory containing a SKILL.md file with YAML frontmatter and instructions. Most examples use Apache 2.0, while the document-creation skills are source-available rather than open source.

Codex CLI provides a local command-line entry point to OpenAI's coding agent. After installing the binary or package, run `codex` to begin working from a project terminal on macOS, Linux, or Windows. The repository focuses on the CLI while directing users to separate IDE, desktop, and web experiences when those interfaces are preferable.

Cline works as a VS Code extension, terminal assistant, or SDK, allowing an agent to create files, run shell commands, browse the web, and call tools with human approval. Project rules and skills carry across its clients, while the SDK can register custom tools, lifecycle hooks, plugins, or MCP servers. It therefore spans personal coding assistance as well as custom agents and scheduled or embedded automation.

MCP Inspector connects to servers over stdio, SSE, or Streamable HTTP and lets developers inspect tools, resources, prompts, and responses in a browser UI. Its CLI supports scripted checks and continuous integration, and server configurations can be exported for other clients. Because the local proxy can spawn processes and connect to specified servers, it should not be exposed to untrusted networks.

Point Track Coach at an audio bounce and optionally an Ableton project to inspect structure, tonal balance, stems, masking, rhythm, automation, and version-to-version changes. It produces an interactive local HTML report where charts, observations, and recommendations link back to audible moments. The full workflow is currently aimed at macOS and relies on tools such as ffmpeg, Python, librosa, and Demucs.

CrewAI combines two approaches to agent automation: Crews organize autonomous agents with roles, goals, tools, and tasks, while Flows provide explicit event-driven workflow control. Python developers can use these primitives for both collaborative and tightly managed processes, with optional tracing and centralized management. Example projects typically require API keys for the chosen model and any external tools.

Browser Use gives an AI agent direct browser interaction for navigation, clicking, typing, and form completion. The CLI is designed for one-off browser tasks driven by an existing agent such as Claude Code or Codex, while the Python library supports repeatable automation, parallel jobs, custom tools, and product integrations with a model provider of the developer's choice.

Web Asset Generator creates favicons, app icons, and social sharing images from a logo, text, or emoji, then produces common sizes and checks file and accessibility requirements. It can detect frameworks such as Next.js or Astro and help integrate the assets. The scripts require Python 3.6+ and Pillow; emoji rendering uses optional dependencies.

PraisonAI supports both single-agent and multi-agent workflows for research, code generation, content work, data tasks, and cross-channel automation. Developers can start with the Python SDK or CLI, then add a visual flow builder, chat UI, or messaging dashboard as needed. Memory, knowledge retrieval, planning, MCP, agent handoffs, and broad model support are built in, while some web-search and messaging features require additional API keys or platform tokens.

ECC provides a modular environment of specialized subagents, development skills, rules, commands, session state, and security hooks for planning, testing, review, debugging, and reusable knowledge. It is primarily distributed as a Claude Code plugin, with adapters and configuration for Codex, OpenCode, Cursor, and other assistants. Its selective installer lets advanced users add only the components needed for their workflow rather than adopting the entire collection.

Pydantic AI is a Python framework for building agents with typed instructions, dependencies, tool calls, and structured outputs. It supports many model providers, MCP integrations, streamed results, tool-call approval, durable execution, and graph-based workflows. This makes it a fit for developers who want agent behavior checked by Python types and evaluated over time.

Give the plugin an open-source repository and it can develop problem-led content angles, publish across Bluesky, Dev.to, Hashnode, and Mastodon, and prepare or submit listings to developer directories. A rule-based gate rejects common marketing filler before posting, while scheduled cycles rotate angles and collect engagement metrics. Platform credentials are required.

The OpenAI Agents SDK organizes agent workflows around tools, guardrails, handoffs, sessions, and tracing. It also includes sandbox agents for filesystem work and realtime agents for low-latency voice interactions, with support for Node.js, Deno, and Bun. Sandbox agents are marked beta, and browser voice sessions should use a server-created ephemeral client token rather than exposing a long-lived API key.

Langflow lets developers assemble models, retrieval, vector databases, and multi-agent steps on a visual canvas, while retaining Python access for custom components. Flows can be tested interactively, exported as JSON, deployed behind an API, or served as MCP tools for other applications. A bundled desktop edition is available for Windows and macOS, alongside Python and Docker installations.