Recently updated open-source projects

Agent Deck provides one terminal view for coding-agent sessions across many projects, making it easy to see which are running, waiting for input, or finished. Groups, search, session forking, and Git worktrees help organize parallel work, while the conductor can surface sessions that need attention remotely.

SWE-bench evaluates language models on real software issues collected from GitHub by giving each model a codebase and an issue, then testing the generated patch in a reproducible Docker environment. It is useful for comparing coding agents on actual repair work, but local evaluation is resource intensive: the project recommends an x86_64 machine with at least 120 GB of free storage and 16 GB of RAM.

CardRadar reads your credit-card reward rules from your Notion database and chooses among the cards you actually own. For a purchase, it ranks real cash as min(amount × reward rate, monthly cap), so a high headline percentage cannot hide a low monthly ceiling. Card profiles, spending habits, and local IDs stay on your machine; this is a wallet-specific Claude skill, not a broad card catalog or payment app.

Claude Agent SDK for Python exposes an async query function and ClaudeSDKClient for bidirectional, interactive conversations with Claude Code. Custom tools and hooks can run as in-process MCP servers, while external MCP servers can be used alongside them. The package bundles the Claude Code CLI by default, requires Python 3.10 or newer, and allows a custom CLI path.

Agentlas OS and its Hephaestus engine compile plain-language requests into portable agent or team packages with roles, tools, memory boundaries, permissions, routing, and verification contracts. Users can borrow public Hub specialists or store owned packages in a private Agent Cloud, while the chosen host executes them. Credentials, local files, and machine permissions stay local.

This starter pack combines four standalone 2D and 3D game projects, a bonus isometric-sprite workflow, eight game-development agent skills, and beginner tutorials. Builders can begin with Phaser platformer or tactics examples, Three.js arena and mini-game templates, or the asset workflow instead of starting from an empty folder. The repository root must be opened for the shared skills to be visible to the agent, and some starters—such as the Oak Woods project—require external art assets.

MCP Registry exposes a searchable list of MCP servers to clients and provides an API, publisher tooling, and a locally runnable service for maintainers. Publishing verifies namespace ownership through GitHub OAuth, GitHub OIDC, DNS, or HTTP challenges. The documented service remains in preview, with the v0.1 API temporarily frozen for integration validation rather than declared generally available.

Open Claude Cowork combines a desktop work assistant with a messaging-based personal assistant. The desktop app supports persistent chats, streamed responses, visible tool activity, and custom skills, while the companion bot connects through WhatsApp, Telegram, Signal, or iMessage for reminders, memory, and browser tasks. Both use Composio integrations and require Anthropic and Composio API keys.

Vibe Kanban has announced that it is sunsetting. Its existing workflow organizes work as prioritized kanban issues and creates agent workspaces with a branch, terminal, and development server. The interface supports diff review, inline comments, application previews, and switching among coding agents such as Claude Code, Codex, and Gemini CLI.

OMK is a provider-neutral control plane for Codex, Claude Code, OpenCode, and local coding agents. It routes work into scoped DAG lanes, verifies and measures execution, exposes MCP and context-budget state, and records artifacts that can be replayed; the monorepo also ships an interactive CLI, agent runtime, multi-provider LLM API, and operator TUI. OMK extensions and MCP servers support workflow integrations such as Slack or chat automation.

feature-plan moves a rough feature idea through three ordered stages: shape the problem, stress-test it against the project’s vocabulary and context documents, then produce a plan spanning product, design, engineering, and developer experience. Its defining constraint is keeping the domain-language check between ideation and architecture review.

hcom links coding agents running in separate terminals so they can send messages, report status, and spawn one another. It supports several popular CLIs through a single Rust binary with no background service, letting you split work, collect results, or ask one agent to check another’s changes.

Letta offers a CLI and TypeScript Agent SDK for running agents with advanced memory, either fully locally, in its cloud, or against a self-hosted App Server. This repository mainly retains the legacy Letta Server behind the V1 API and SDKs; active development has moved to the Letta Agent repository, and new projects are encouraged to use the newer Agent SDK.

kimiflow runs an explicitly invoked feature or bug-fix loop through clarification, diagnosis, planning, implementation, verification, review, and commit approval. Tested scripts enforce blocker resolution, red-and-green evidence, durable state, and other gates instead of relying on the model’s self-report. It can also maintain a local codebase map and learning memory, with optional Obsidian integration; jq is required, and the workflow stops for explicit approval before committing.

Dexto is an agent harness for defining models, tools, and behavior in YAML rather than wiring every workflow by hand. It includes sessions, persistent memory, tool orchestration, multimodal input, and observability, with a ready-made coding agent as an example. Agents can run locally, in the cloud, or through CLI, web, REST, Discord, and Telegram interfaces.

Patterns.dev Skills packages 58 development patterns for JavaScript, React, and Vue as independently installable agent skills. Users can add an entire technology stack or select individual topics such as hooks, singletons, composables, AI interface patterns, streaming SSR, code splitting, or rendering strategies. The collection gives coding agents concise prose and code examples for implementation decisions. It deliberately omits the animations, diagrams, sandboxes, and other visual walkthroughs available on the full patterns.dev website.

Puppeteer Real Browser MCP Server starts and owns a local Chrome process for an MCP client, exposing navigation, content reading, element lookup, clicks, typing, waits, and scrolling. It defaults to a visible browser, supports headless mode, proxies, custom Chrome paths, and dedicated profiles, and closes only the process it launched. The workflow requires reading page content before interaction, allows one browser session, and does not provide a real CAPTCHA solver.

Metric Brand Auditor evaluates a brand through seven research dimensions and five scoring lenses. Parallel researchers gather public material, independent judge panels score the same synthesis, and a lead merges consensus, disagreement, leverage, and a 90-day action plan into Markdown and HTML reports. Re-running a brand can use an evolution mode that focuses on changed dimensions and preserves dated report versions.