Prime Agent is an MIT-licensed open-source coding and research agent from Prime Intellect, built around a Recursive Language Model (RLM) runtime and a persistent IPython control environment. The model works through programmatic Python: file/shell operations, data transforms, skills, context management, and recursive subagents are all callable from the durable kernel rather than exposed as a large fixed tool palette.
Local-first AI agent whose control loop and all state run on the machine. Drives browser, files, approved shell, documents, memory across sessions, scheduled tasks, and MCP tools. Default path is llama.cpp (their TurboQuant fork) so small quantized models stay useful for long multi-step work on consumer hardware.
Kubernetes-native platform for isolated, stateful singleton workloads used as AI agent runtimes and development environments. Sandbox, SandboxTemplate, SandboxClaim, and SandboxWarmPool CRDs provide a declarative API for long-running containers with stable identity, optional pers
Agentation is a visual feedback layer for agents working on web UIs. Humans click elements, annotate, optionally use layout mode (rearrange/resize), then copy structured markdown (or push via MCP/webhooks) so the agent gets element paths, CSS selectors, positions, and related context instead of vague prose.
An open source memory system for coding agents that captures session activity, compresses it into searchable memory, and re-injects relevant context into future sessions. The repo positions it as agent-agnostic infrastructure that can work across Claude Code, OpenClaw, Codex CLI, Cursor, Gemini CLI, and other MCP or HTTP-capable agents.
Sitegeist is an AI assistant that lives in a Chromium side panel, plus a bridge so coding agents can use the same browser. README framing: collaboration over “autonomy theater” — you guide, it executes. Automates repetitive web tasks, extraction, navigation, forms, research → docs/spreadsheets; BYO API key or subscription login (Anthropic, OpenAI/ChatGPT, GitHub Copilot, Gemini, etc.); claims local-only / no tracking.
Vellum Assistant is a personal AI assistant product (OSS + managed cloud) in the same broad class as Hermes / OpenClaw / NanoClaw: multi-channel agent, skills, memory, computer use, OAuth-connected tools, managed or self-hosted runtime. Entry capture was the OAuth Integrations key-concept doc: browser OAuth2 instead of pasting API keys; each integration is a bundled skill with tools (e.g.
Landing hero: Connect any agent to everything. Tagline: The gateway to connect your agent to everything. Page: Executor is an MCP gateway. Anything that speaks MCP, like Claude Code, Cursor, or Codex, points at one endpoint and reaches every tool you connect.
bb is an MIT-licensed, self-hosted agentic IDE that coordinates installed coding-agent CLIs in persistent, steerable threads. Its first-class control surfaces are a desktop app, web app, CLI, and HTTP API. The project’s central differentiator is that much of its UI/workflow surface is implemented as plugins, so an agent can extend the IDE itself rather than only working inside a fixed product.
Box is a managed sandbox product from Ascii (legal entity Dedale AI, Corp.). It sells persistent Linux VMs for coding agents and agent platforms: SSH, Docker inside the VM, a dedicated IPv4 per machine, snapshot/stop/resume, fork-from-snapshot, optional 60fps desktop stream, and a CLI/API/box prompt path for Claude Code or Codex.
fx is a Vercel Labs coding-agent harness and CLI written in Zig. Official copy positions it as a tiny, open, native agent originally used internally, now open sourced for research and embedding in larger systems. The durable object is the product at fx.sh / vercel-labs/fx, not the tweet.
gws is a single Google Workspace CLI. It dynamically reads Google's Discovery Service at runtime and builds its command surface from the available APIs instead of shipping a static command list. The README positions it as built for both humans and AI agents: every response is structured JSON, and the repo includes 40+ agent skills.
OpenRouter is a unified large language model inference router and model marketplace: one OpenAI-compatible HTTP application programming interface across many underlying inference providers, with automatic fallbacks and cost-aware routing. The marketplace object is models and inference providers (tokens in → tokens out), not installable agent tools ranked by dual human-and-agent reputation.
A mature AGPL-3.0 open-source agent-control surface: a local Paseo daemon runs and manages coding-agent CLIs on a user-controlled machine, while desktop, mobile, web, and CLI clients monitor/control those agents. It supports Claude Code, Codex, Copilot, OpenCode, Pi, and 30+ other providers/integrations according to its site.
tldraw offline is tldraw’s local, file-based desktop infinite canvas (no accounts / no server required for core use). Documents save as portable .tldraw files (canvas + media + optional embedded document scripts). The 2026-08-03 @tldraw post demos the agent-native angle: you can make your own tools on the canvas — e.g.
Treg is an agent-oriented tool registry and credential-injection proxy. Its hosted catalog claims roughly 2,600 endpoints across about 40 providers, searchable by task with request/response shape and per-call prices. A team can also register its own API credentials, OAuth connections, CLIs, and SKILL.md bundles; callers use one Treg token while upstream credentials remain server-side.
README tagline: "One MCP for reverse engineering across binaries, applications, and runtime behavior." The agent calls REA through MCP to inspect a local target and trace relevant code; REA returns findings with their evidence, and the agent uses them to explain the behavior or write and test an implementation.
MIT-licensed Node CLI + agent skill by Peter Yang. It opens a target HTML/Markdown file or localhost route in a local browser review UI. The reviewer can directly change text/basic formatting, resize images, delete elements, and attach comments to selected text or blocks.
Portable agent skill that edits drafts to remove 20+ AI-slop patterns (min effective edit + self-check via eval.md) or detects patterns with quoted lines — without claiming AI-authorship detectors.
BetterWright is materially faster once warm, slightly more compact, and has a strong agent-oriented design: persistent sessions, code snippets, diff snapshots, credential redaction/vaulting, network policy, evidence artifacts, and an optional contained agent loop.
GitHub description: An open source agent microharness featuring persistent agency and recursive LLMs. Of bash, by bash, for bash; it's shells all the way down. Homepage field: headlong.ai. README: complete harness with a core of less than 10K lines of Bash; defining feature is persistent agency — the agent keeps thinking between external interactions in a self-guided loop.
The computer underneath the agent. Instant full cloud desktops (GUI, browser, filesystem, terminal, apps) that AI agents can see and control via API — boot claimed under ~500ms. Always-on until stopped. Linux + Windows GA; Mac closed beta. People use it three common ways (docs): 1.
Agency is a local card-feed app from Browser Use: a coding-agent coordinator prepares suggestion cards; the human approves with one click (Send / Apply / Merge). Thesis (X + README): agents find useful work; you decide — “Never prompt again.
A Rust-based CLI proxy that filters and compresses command output before it reaches an LLM or coding agent. The repo positions RTK as a low-overhead way to reduce token consumption on common development commands like git status, git diff, test runners, grep/search, logs, and cloud/devops tooling.
