// category
Open-source models, local inference, and community-built AI dev tooling.
Open-source AI tooling — local model runners (Ollama, LM Studio), OSS coding agents, open-weights frontier models (Llama, DeepSeek, Qwen), and community projects pushing the AI dev stack forward without a vendor lock.
A developer released Hermes IDE version 1.1.5, an open-source IDE that runs Claude AI in isolated sessions per project tab, each with its own working directory, permission mode, model selection, and MCP plugin configuration. The tool is available at hermes-ide.com and on GitHub, and was built to ...
A developer released Neural Viz Lab, a browser-based tool for visualizing neural network internals including weight interactions, tensor flow through layers, and attention head mechanics. The tool runs without a GPU or Python environment and is available at a Cloudflare Pages URL.
Vercel's Chat SDK added a Messenger adapter, enabling support for messages, reactions, multimedia downloads, postback buttons, and automatic display name retrieval from user profiles. The adapter was based on a community contribution via GitHub PR #461.
A developer released ARIA (Autonomous Reasoning and Intelligent Agent), a free open-source CLI coding agent installable via `pip install aria-x`. It supports any Ollama model and includes AST pre-validation for Python, RAG-based Stack Overflow search, and automatic approach-switching after three ...
A developer released Veritas, an open-source Python library that tracks confidence levels for AI agent beliefs using a four-component vector covering estimated certainty, source fragility, staleness, and source diversity. The system applies temporal decay to evidence based on source type, with em...
Yugabyte launched Meko, an open source data infrastructure tool designed to manage state in multi-agent AI systems. The company cites research finding that 37% of multi-agent system failures stem from state inconsistencies rather than model reasoning errors.
Simon Willison released llm-gemini 0.31, updating the plugin to reflect that Google's Gemini 3.1 Flash-Lite model is now generally available, having exited preview status.
Hermes Agent includes a Kanban-style task board backed by SQLite that dispatches jobs to self-hosted LLMs via a daemon process. Without parallelism limits, the dispatcher can flood single-GPU or CPU-bound hosts; the current workaround is using `hermes kanban dispatch --max` to cap concurrent prom...
A developer built two open-source AI workflow tools — `project-complexity-mapper` and `project-complexity-action-planner` — based on TU Delft's TOE framework, which categorizes project complexity into Technical, Organizational, and External dimensions to guide management responses.
Addy Osmani, Engineering Director at Google Chrome, released "Agent Skills," a MIT-licensed collection of 20 Markdown-based engineering workflows for AI coding tools such as Claude Code and Cursor. The project, which has accumulated over 30,800 GitHub stars, encodes structured checklists to enfor...
A 2026 roundup identifies seven small language models suited for local coding use on consumer hardware, including OpenAI's open-weight gpt-oss-20b, a 21-billion-parameter MoE model with 3.6 billion active parameters and a 128k context window, and Alibaba's Qwen3-VL-32B-Instruct.
A 2026 roundup identifies seven small language models suited for local coding use on consumer hardware, including OpenAI's open-weight gpt-oss-20b, a 21-billion-parameter MoE model with 3.6 billion active parameters and a 128k context window, and Alibaba's Qwen3-VL-32B-Instruct.
A developer released Network-AI, an open-source coordination layer for multi-agent AI systems that uses atomic propose-validate-commit cycles to prevent concurrent state conflicts. The tool supports 14 frameworks including LangChain, AutoGen, and CrewAI, and includes per-agent token budget contro...
datasette-llm 0.1a7 adds a configuration mechanism for setting default options on specific LLM models, allowing users to define defaults such as model selection and temperature for enrichment operations within Datasette.
Simon Willison released llm-echo 0.5a0, a plugin for the LLM tool that provides a fake "echo" model for automated testing. The update adds a `-o thinking 1` option that simulates a reasoning block, compatible with LLM 0.32a0 and higher.
GitHub launched its sixth annual Maintainer Month, announcing new tools including granular pull request limits for unknown contributors and pull request archiving to remove spam. The releases follow GitHub data showing merged pull requests have nearly doubled year over year, with AI-generated con...
A developer and their partner released Apra Fleet, an open-source MCP server on GitHub that allows a single Claude Code session to dispatch tasks to agents running on multiple machines or in isolated local workspaces. The tool supports doer-reviewer loops within a single machine using separate fo...
Vercel open sourced "deepsec," a security scanning tool that uses AI coding agents (Claude and Codex) to identify vulnerabilities in codebases, running locally or scaling to 1,000+ concurrent Vercel Sandboxes. The tool reports a 10-20% false positive rate and follows a five-step pipeline: scan, i...
WinkTerm is an open-source AI terminal tool that writes commands directly into the shell input line within a shared PTY session, triggered by typing a `#`-prefixed comment. Built on Python, FastAPI, LangGraph, and Next.js, it is MIT licensed and available via Docker or as a desktop app.
A developer released Spine v1, an open-source tool that scans code repositories and generates architecture maps, prioritized reading orders, and a persistent context file at `.claude/REPO_CONTEXT.md` for use with Claude Code sessions.
A developer released Claude Code Cost Explorer, an open-source local dashboard that reads Claude Code's session files to show API costs broken down by individual turn, prompt, and tool call. It installs via pip and requires no API keys or external data uploads.
A developer built README Clew, a tool that audits public GitHub repositories for discrepancies between README documentation and actual code, using Claude Sonnet 4.5 to extract claims and five deterministic verifiers to check dependencies, scripts, environment variables, and file references. The t...
OpenSearch released versions 3.5 and 3.6 in February and April 2026, adding Better Binary Quantization (BBQ) that compresses high-dimensional vectors by 32x and the SEISMIC algorithm for sparse neural approximate nearest neighbor search. The project is positioning the open-source search engine to...
Simon Willison released version 0.32a0, an alpha build, of his open-source `llm` command-line tool for interacting with large language models.
Simon Willison released LLM 0.32a0, an alpha version of his Python library and CLI tool for accessing LLMs. The update refactors the core abstraction so model inputs can be represented as a sequence of messages and responses can include multiple typed parts, replacing the previous single-prompt, ...
Simon Willison released llm 0.32a1, a bug fix for the prior 0.32a0 release, correcting an issue where tool-calling conversations were not correctly restored from SQLite storage.
Warp, maker of a Rust-based agentic development environment, released its client as open source under the AGPL license, with OpenAI named as founding sponsor of the repository. The agent workflows powering the platform are built on GPT models, and the company cited faster community-driven develop...
Mininglamp Technology released Mano-P, an open-source GUI agent model under Apache 2.0 that operates graphical interfaces using only screenshots, without DOM parsing or system APIs. The model uses three-stage progressive training to enable AI agents to interact with software that lacks programmat...
Microsoft released VibeVoice, an open-source voice AI system, publishing the code on GitHub. The project appears to be a speech/voice model or toolkit made publicly available by Microsoft.
Microsoft released VibeVoice, an MIT-licensed speech-to-text model with built-in speaker diarization, on January 21, 2026. The 17.3GB model (available in a 5.71GB 4-bit version) handles up to one hour of audio and transcribed 60 minutes of speech in 8 minutes 45 seconds on an M5 Max MacBook Pro.
Attackers compromised elementary-data, an open source Python CLI tool with over 1 million monthly downloads, by exploiting a vulnerability in the developers' account workflow to push malicious version 0.23.3 to PyPI and Docker. The package collected warehouse credentials, cloud provider keys, API...
A four-person engineering team built and open-sourced a self-hosted web dashboard for managing Claude Code sessions across eight machines, replacing manual SSH workflows they said cost ~30 minutes daily. The MIT-licensed tool uses inotify file watching, PostgreSQL storage, and WebSocket streaming...
A developer released "Swing" (thestack-ai/swing-skills), an open-source set of six Claude Code skills designed to counter AI failure modes including confirmation bias and hallucination. The author reports agent-induced production bugs fell from 28% to 4% over 200 sessions across six weeks of pers...
OpenAI published Symphony, an open-source specification for orchestrating Codex agents that connects issue trackers to automated agent systems, aiming to reduce manual context switching in engineering workflows.
Four open-source tools for Claude-based development workflows include Superpowers (structured prompt orchestration), UI UX Pro Max Skill (UI/UX prompt sets for React/Tailwind), Awesome Claude Code (curated resource list), and Claude Mem (persistent memory across sessions).
A developer released Network-AI, an open-source coordination layer for multi-agent AI systems that addresses race conditions in shared state access. The tool uses a propose-validate-commit cycle and claims support for 14 frameworks including LangChain, AutoGen, CrewAI, and MCP.
Ali Khokhar released free-claude-code, a FastAPI-based local proxy that redirects Anthropic API requests from Claude Code to free backends such as NVIDIA NIM, OpenRouter, or Ollama by changing one environment variable. The project accumulated over 14,300 GitHub stars and 2,000 forks, topping GitH...
A network engineer released NetCopilot, a beta SSH, Telnet, and Serial terminal client that embeds AI assistance to help diagnose network issues such as BGP failures and routing misconfigurations. The tool supports connections to Cisco, Juniper, and Linux devices and analyzes command outputs in c...
A developer published an open-source personal finance agent system built on Claude, using two specialized agents routed by text prefixes and lazy-loaded from local JSON files. The project, available at github.com/jasonsieg22/claude-finance-agents under MIT license, claims an approximately 80% tok...
Simon Willison released version 0.31 of his open-source `llm` CLI tool, adding support for OpenAI's GPT-5.5 model, a verbosity level option for GPT-5+ models, and an image detail level parameter for image attachments.
A developer released SAMP (Simple Agent Message Protocol) and a reference implementation called "agent-message," enabling AI coding agents to pass messages across separate repository sessions using append-only JSONL log files with no servers or daemons. The system uses content-addressed message I...
Mistral AI launched Leanstral in March, an open-source code agent that uses formal verification via the Lean 4 programming language to mathematically prove code correctness. The model uses a Mixture-of-Experts architecture with 119 billion total parameters and 6.5 billion active parameters, relea...
Developer changmyoungkim released Doby, an open-source tool that uses pre-compiled indexing and a spec-first workflow to reduce token usage in Claude Code navigation from 2,000–5,000 tokens per query to approximately 100 tokens, a claimed 95% reduction.
OpenAI released GPT-5.5 in its Codex product and for paid ChatGPT subscribers, but withheld API access pending safety reviews. Developer Simon Willison built an open-source LLM plugin, llm-openai-via-codex, that accesses GPT-5.5 through the Codex CLI's authentication mechanism, which OpenAI has s...
AgentBox SDK is a new TypeScript library that provides a single API for running coding agents such as Claude Code across different sandbox providers, including local Docker. It launches agents as server processes communicating via WebSocket or HTTP, preserving approval flows and streaming events.
Simon Willison released llm-openai-via-codex 0.1a0, a plugin that reuses Codex CLI credentials to make OpenAI API calls via the LLM command-line tool, bypassing the need for separate API key configuration.
A developer built a Chrome extension called Claude Exporter that allows users to save Claude AI conversations to PDF, Word (.docx), Google Docs, and Notion, with preserved code block formatting. The tool also supports exporting partial conversations rather than entire sessions.
Simon Willison built a browser-based version of LlamaIndex's LiteParse PDF text extraction tool, available at simonw.github.io/liteparse/. The tool uses PDF.js and Tesseract.js to extract and spatially order text from PDFs entirely client-side, with no data leaving the user's machine.
OpenAI released an open-weight model called OpenAI Privacy Filter designed to detect and redact personally identifiable information (PII) in text. The company describes it as achieving state-of-the-art accuracy for PII detection tasks.
A developer found two bugs in Longhand, a local session-memory tool for Claude Code: a 2,526-line work session was never ingested due to a missing SessionEnd hook, and two other sessions were stored with NULL project IDs linked to the home directory instead of the correct project path.
A developer published an AI skill called "Pinocchio" that appends a confidence percentage to responses from Claude and GitHub Copilot in VS Code, intended to flag potential hallucinations. The source code is available on GitHub at luizwbr/pinocchio-skill.
ccusage-web is an open-source web dashboard for ccusage that displays Claude Code token costs broken down by day, month, session, and 5-hour billing block, with interactive charts and per-model breakdowns.
n8n is an open-source, self-hostable workflow automation platform that connects apps and APIs without extensive coding. DigiMateAI published a guide promoting its use for Qatar enterprises, citing local data residency compliance and fixed-cost pricing as advantages over cloud-based alternatives l...
The Eclipse Foundation announced the Open VSX Managed Registry, a foundation-operated managed service for the Open VSX extension registry, an open source alternative to Microsoft's VS Code Marketplace. Commercial adopters receive a 99.95% uptime SLA, support tiers, and service credits; users incl...
A developer released an unofficial Linux desktop client for Claude AI, built as a wrapper around Anthropic's official Claude CLI. The app stores chats locally in SQLite, runs file-access sessions inside a bubblewrap sandbox, and is currently in alpha.
OpenClaw's plugin ecosystem supports native gateway packages for channels, model providers, memory, voice, and web search, installable via CLI with validated manifests. The most mature plugins cover browser access, memory, and observability, while business-automation categories such as CRM and in...
Simon Willison released llm-anthropic 0.25, a plugin providing LLM access to Anthropic's Claude models. The update adds support for claude-opus-4.7 with a thinking_effort option, new thinking_display and thinking_adaptive boolean flags, and increased default max_tokens limits per model.
Alibaba's Qwen3.6-35B-A3B, run locally as a 20.9GB quantized model on a MacBook Pro M5, produced higher-quality SVG illustrations than Anthropic's Claude Opus 4.7 in informal tests conducted by Simon Willison on April 16, 2026.
Hugging Face released HoloTab, a browser-based AI agent designed for "computer use" tasks — allowing the agent to interact with web interfaces autonomously. The project is open-source and operates through the browser to automate computer interactions.
Google's Gemma 4 AI model can run natively on iPhones with full offline inference, eliminating the need for cloud connectivity to use the model.
A developer released TokenBar, a macOS menu bar application that monitors token usage for Claude, Cursor, and other AI tools in real-time. The app costs $5 as a one-time purchase and aims to help users track spending before receiving bills.
Enterprises are increasingly adopting self-hosted open-source image generation models combined with affordable APIs instead of relying solely on proprietary services like DALL-E and Midjourney, with per-image costs dropping from $0.02-$0.25 to $0.0005-$0.001 when self-hosted at scale.
A developer released a free, browser-based tool that removes visible and invisible watermarks from Google Gemini-generated images using local processing with no server uploads or account requirements.
Datasette released version 1.0a27, replacing Django-style CSRF tokens with modern browser headers and adding a RenameTableEvent for plugin compatibility. The alpha also includes new actor parameters for client methods, temporary disk database options, and improvements to the upsert API.
Rod Johnson, creator of the Spring Framework, launched Embabel, an Apache-licensed agentic AI framework for Java built on Spring Boot, at Microsoft's JDConf conference to address enterprise predictability challenges in large language model applications.
GitHub introduced Code Security Risk Assessment, a free tool that scans up to 20 repositories using CodeQL to identify vulnerabilities by severity and language, available to organization admins and security managers at no cost.
Datasette pull request #2689 replaces token-based CSRF protection with Sec-Fetch-Site header protection, removing the need for hidden CSRF token form inputs and simplifying the security implementation based on research by Filippo Valsorda and Go 1.25.
A developer released an MCP server for CronSignal that allows Claude and other AI tools to diagnose failed cron jobs, retrieve error logs, and manage monitors through terminal commands. The server exposes tools including list_checks, diagnose_check, get_check_output, pause_check, and resume_check.
Quick Codex is an open-source workflow layer for Codex CLI that stores task state in files to address context loss during multi-turn coding sessions. The tool provides two modes—qc-flow for exploratory work and qc-lock for strict execution—along with utilities for resuming, verifying, and checkin...
A developer launched SimplyLouie, a Claude API gateway offering $10/month flat-rate access with unlimited calls, as an alternative to Anthropic's $15 per million token pricing model.
The Servo browser engine was released as an embeddable Rust crate on crates.io. A CLI tool was built to demonstrate its ability to render web pages as screenshots, though compiling Servo itself to WebAssembly proved infeasible due to threading and dependency constraints.
Everything-claude-code is an open-source enhancement framework for Claude Code that includes 181 skills, 47 sub-agents, and 34 rules designed to improve productivity and code quality. The project, created by Affaan Mustafa, reportedly has over 150,000 GitHub stars and supports multiple AI coding ...
A developer published Sessioncraft, an open-source governance system for managing state and context drift across long-running AI-assisted projects using Claude, after identifying recurring problems across 180+ sessions including stale information and forgotten constraints.
A developer created cc-statistics, an open-source tool that aggregates AI coding costs from Claude Code, Gemini CLI, Codex, and Cursor into a unified view via CLI, web dashboard, and macOS menu-bar app.
Simon Willison released a web-based demo tool for SQLite's Query Result Formatter library, which allows users to test various rendering options for SQL result tables using WebAssembly.
Open Claude Proxy v3.5.0 allows multiple devices on a local network to share a single Claude subscription by routing requests through a central server; new clients can connect by running a single curl command instead of manually configuring API keys.
Claude-code-best-practice is an open-source reference guide by shanraisshan with over 1,000 GitHub stars that provides strategies for using Anthropic's Claude Code CLI tool, including guidance on configuration files, prompt engineering, and coordinating multi-step development tasks.
Google's Gemma 4 is an open-source AI model designed to run locally on personal devices without API costs or cloud dependencies. The model supports 256K context windows, multimodal inputs, and function calling, enabling uses such as local coding assistants, document analysis, and AI agents.
A developer created Routiform, an AI gateway tool, after encountering bugs in 9router and OmniRoute including Docker build failures, authentication issues, and cost tracking problems. Routiform addresses these issues with improved Docker support, fixed login flows after backups, and better OAuth ...
The PyTorch Foundation announced three new projects joining its portfolio at PyTorch Conference EU: Safetensors for secure model distribution, ExecuTorch for inference capabilities, and Helion. The additions expand the foundation's vendor-neutral infrastructure for the AI lifecycle.
Simon Willison created a tool that displays GitHub repository sizes by querying GitHub's API, which GitHub's web interface does not expose. Users can input a repository name to see its total size.
Blackdesk is an open-source market research terminal built in Go that consolidates stock quotes, charts, news, screeners, and AI analysis into a single keyboard-driven interface. The tool runs locally and supports multiple AI connectors including Codex and Claude Code for context-aware market ana...
Cole Medin released second-brain-skills, an open-source toolkit that extends Claude Code with skills for knowledge work tasks including brand guidelines, presentations, operations documentation, and video generation. The project uses progressive context disclosure to manage token usage and had 64...
A developer built kerf-cli, a local CLI tool that analyzes Claude Code's session logs stored in SQLite, enabling detailed cost tracking and optimization recommendations that Claude Code's native `/cost` command does not provide.
A developer released Octopal, an open-source app that lets users manage multiple Claude AI agents organized by project folder, with agents communicating via chat interface and an orchestrator routing tasks between them.
Simon Willison's research confirms that SQLite WAL mode functions correctly across separate Docker containers sharing the same volume, as containers on the same host share the same shared memory required for WAL coordination.
Google released the AI Edge Gallery app for iPhone, enabling users to run Gemma 4 language models locally on their device with image questioning, audio transcription, and interactive tool-calling demonstrations.
An open registry called TokRepo launched to centralize AI development assets like MCP configurations, Claude Code skills, and system prompts, enabling searchable discovery and one-command installation similar to npm for JavaScript.
A developer built PokeClaw, an Android app that runs a 2.3B-parameter AI model locally on the phone without Wi-Fi or API keys, using the device's accessibility APIs to read UI elements and execute tasks like sending messages entirely on-device.
A developer built Ustaad, an LLM-powered wiki system that automatically organizes and cross-references documents by having the LLM maintain a structured knowledge base at ingestion time rather than searching raw documents at query time.
A developer created CodeIsland, a tool that displays a Claude AI companion in a MacBook's notch to manage up to 8 simultaneous Claude Code sessions with smart notifications and a gamified buddy character that tracks usage statistics.
Quarkus-chat-ui is a web interface that allows multiple LLM instances to communicate with each other via HTTP MCP servers, with each instance exposing tools that others can call. The tool was developed as a real-world application of POJO-actor, a Java actor framework for managing concurrent commu...
A developer released CodeIsland, a free open-source macOS app that displays Claude AI coding agent status in the MacBook's notch, allowing users to monitor multiple sessions, track API usage, and view chat history without switching windows.
Simon Willison released scan-for-secrets 0.1, a Python tool that scans directories for API keys and other secrets, including detecting common encoding variations. The tool supports storing frequently-checked secrets in a configuration file.
scan-for-secrets 0.2 released with streaming results output, support for scanning multiple directories and individual files, new Python API functions, and a verbose mode option.
A developer released Apfel, a free tool on GitHub that provides access to AI capabilities already present on macOS systems. The project received 643 upvotes and 138 comments on Hacker News.
Rotato is an open-source Node.js proxy that automatically rotates between multiple API keys to bypass rate limits on services like OpenAI and Gemini. It requires no external dependencies and acts as a drop-in replacement for direct API calls.
Vercel Sandbox SDK 1.9.0 added the ability to set file permissions directly when writing files via a new `mode` property in the writeFiles API, eliminating the need for separate chmod operations.