26 October 2026

Local AI agent with multiple harnesses and models

Cindy is an open source AI agent that runs on your own computer and uses the files and apps you are already logged into. Its first harnesses are Claude Code and Codex.

• Models and harnesses can be mixed and switched mid-task, while the workspace, memory, skills, and tools stay the same.
• A single task can be planned, run in parallel, and reviewed by agents with different harness and model combinations.
• A free client under the Apache-2.0 license. It can use an already-paid Claude Code or Codex Coding Plan, your own API key, or a local model.

Design skill for coding agents so pages do not look AI-made

Hallmark is a skill for Claude Code, Codex, Cursor, and OpenCode. It picks a theme from a catalog of 48 themes, builds the page, then checks it in a real browser.

• There are three tiers of quality gates: Floor must not fail (contrast, focus, phone layout), Reflex can be overridden with a reason, and Finish does not block a release.
• /hallmark redesign rebuilds a page and preserves its content and wiring. /hallmark audit only produces a report.
• Made by Together AI. For photos, 755 openly licensed photos are available.

A locally running ElevenLabs alternative

VoiceStudio is an open source app for voice cloning, voice design, video dubbing, dictation, transcription, and audiobook creation in 646 languages.

• The default engine is OmniVoice from k2-fsa. Other engines can be chosen from a catalog.
• The local workflow runs on your own hardware. Remote services are optional, and usage analytics require consent.
• There is a local API and MCP for agents, reusable recipes, and a watch folder.

Design guidance for AI coding agents

Impeccable contains 1 skill, 24 commands, live iteration in the browser, and 62 deterministic detector rules for AI-made frontend design.

• /impeccable init inspects the project and writes PRODUCT.md with product facts: audience, goals, context, constraints, and voice.
• Commands include polish, audit, critique, distill, animate, bolder, and quieter.
• The deterministic rules run through the CLI and browser extension without an LLM and without an API key.

Small bipedal duck robot that moves with RL

Microduck is about 25 cm tall and weighs 800 g. This repo is its brain: several daemons on a Rockchip RK3566 that run a 50 Hz control loop for fifteen servos.

• The neural policy is trained in the microduck_rl repo with MuJoCo and PPO, then exported to ONNX to be loaded by this repo.
• It can walk, roll on wheels, pick up objects with its beak, and get up on its own after falling.
• robotctl handles steering, configuration, sound, wifi, updates, and logs. duckctl controls the robot over Bluetooth from a laptop.

Financial trading framework with multiple LLM agents

TradingAgents is a multi-agent LLM framework for financial trading analysis. Release v0.7.0 (October 2026) adds backtesting and support for ChatGPT Plus or Pro plans.

• Analysts run in parallel. Managers and analysts can use different model providers for each model tier.
• The CLI can run without prompts using flags such as --ticker and --date. Reports are saved as a single HTML page.
• Backtests only see data published as of the analysis date. Fundamentals are pulled from SEC EDGAR.

Open source Zi Wei Dou Shu chart engine

This project opens up the Zi Wei Dou Shu charting algorithm based on Ni Haixia's Tianji system, complete with the Si Hua system, a pattern knowledge base, and classical text data.

• The v4.0 interpretation dataset contains 674,976 charts with 13 dimensions, 8,774,688 text segments in total. It is 56.6 GB after extraction and 161 MB compressed.
• License: code is MIT, data is CC BY 4.0. The interpretation sentence library is not open source.
• The interface in this repo is a demo.

Agentic IDE from JetBrains built on the pi coding agent

ThinkRail wraps the pi coding agent with an IDE interface, separate workspaces, and project knowledge through living specs and skills.

• Each workspace is a git worktree with its own branch, so several agents can work in parallel without touching the main checkout.
• A graph of SPEC.md files is stored next to the code. Agents read it and update it through the spec_* tools.
• It includes a Monaco editor, file tree, terminal, and a pull request flow through gh.

Self-hosted PaaS as an alternative to Vercel and Heroku

Coolify manages servers, applications, and databases on your own server over just an SSH connection. It can be a VPS, bare-metal, or a Raspberry Pi.

• Deploy from GitHub, GitLab, Bitbucket, or Gitea with Nixpacks, Railpack, Dockerfile, Docker Compose, or a Docker image.
• There are more than 360 one-click services, pull request previews, rollback, custom domains, and automatic HTTPS.
• If Coolify is stopped, running resources stay up and can still be managed.

Multi-phase security audit skill for coding agents

This skill runs a security audit in six phases using isolated agents and produces machine-readable findings. It became the seed of Cloudflare's vulnerability discovery harness.

• Each candidate finding is handed to a fresh verifier that tries to refute it. Results are split into confirmed, needs_validation, and rejected.
• Findings are written to findings.json and validated against report-schema.json. REPORT.md is derived from the verified records.