This page is a flat list of infrastructure for running agents, not for writing prompts. Sandboxes come first, with Daytona and a write-up on self-built microVM isolation. Runtime and orchestration entries follow, including Vercel's eve, Multica, Google AX, Zeron, and Pi with its durable harness. A cluster of very small assistants appears in the middle, from nullclaw and picoclaw to zeroclaw, each trading features for a tiny footprint. The rest covers context engines like Zep, local inference servers, computer-use drivers, channel SDKs, and model releases such as PersonaPlex and Ternary Bonsai. Many entries carry a detail line about licensing, versions, or caveats, and some link to a fuller note. Use the list to compare options at a given layer.
- https://github.com/daytonaio/daytona — Secure, elastic sandbox infrastructure for running AI-generated code and agent workflows with sub-90ms spin-up.
- https://alfonsocampodonico.com/posts/building-your-own-sandboxes — Article (2026-10-18) on self-building agent/untrusted-code sandboxes: Firecracker and Cloud Hypervisor microVMs behind one API, each VMM in a rootless namespace jail, default-deny per-VM nftables network that blocks cloud metadata, and OCI images converted to one erofs per layer. No public repo. See Building Your Own Sandboxes.
- https://github.com/vercel/eve — eve: Vercel's filesystem-first TypeScript framework for durable backend agents (Workflow SDK sessions, per-session sandbox, channels, evals, Agent Runs tracing); public beta, Apache-2.0. See Eve.
- https://github.com/nuphos/nuphos — Nuphos (Zeabur, Apache-2.0, early access): team workspace that runs Claude Code, Codex, Grok Build, Antigravity and OpenCode over ACP on local, remote or cloud runtimes, with shared sessions, skills, memory and DevOps GUIs; self-hostable. See Nuphos.
- https://github.com/getzep/zep — End-to-end context engineering platform; assembles relationship-aware agent context via Graph RAG with <200ms latency.
- https://github.com/eclaire-labs/eclaire — Self-hosted, privacy-first AI assistant that organizes notes, docs, photos, bookmarks, and tasks locally.
- https://github.com/lyogavin/airllm — Memory-optimized LLM inference library; runs 70B+ models on 4–8GB GPUs without quantization.
- https://github.com/knowns-dev/knowns — The memory layer for AI-native development; gives AI persistent understanding of software projects via tasks, docs, code intelligence, and Claude Code skills.
- https://github.com/nextlevelbuilder/goclaw — Multi-tenant AI agent platform in Go with 20+ LLM providers, 7 messaging channels, and agent team orchestration.
- https://github.com/nullclaw/nullclaw — Ultra-tiny Zig-based autonomous AI assistant; 678KB binary, ~1MB RAM, <2ms boot, 50+ providers, 19 channels.
- https://github.com/sipeed/picoclaw — Ultra-lightweight Go AI assistant for $10 hardware; <10MB RAM, 30+ providers, MCP support, runs on Android & embedded Linux.
- https://github.com/paperclipai/paperclip — Open-source orchestration platform for running AI agent "companies" with org charts, budgets, and governance.
- https://github.com/multica-ai/multica — Self-hostable workspace coordinating 26+ agent CLIs (Claude Code, Codex, Cursor, Copilot, etc.) with humans on a shared project board instead of scattered terminal sessions; Go/Next.js/PostgreSQL stack with Electron/Expo clients, full execution transparency (tool-call/command logs), Git-integrated PR review gates (GitHub/GitLab/Gitea/Forgejo), a "Skills" system for turning solved problems into reusable playbooks, and chat-triggered or scheduled "autopilot" tasks with role-based access control. 51k+ stars; Apache-2.0 base with added commercial-embedding/branding restrictions.
- https://github.com/tduccuong/dmh_ai — Self-hosted private AI chat app with companion memory, built-in web search, multi-user support, and rich media attachments.
- https://github.com/zeroclaw-labs/zeroclaw — Personal AI assistant in Rust; <5MB RAM, <10ms boot, multi-channel inbox, web dashboard, and hardware peripheral support.
- https://huggingface.co/nvidia/personaplex-7b-v1 — 7B real-time speech-to-speech conversational model with voice and role control; supports interruptions, overlap, and rapid turn-taking.
- https://huggingface.co/Zyphra/ZUNA — 380M-parameter EEG foundation model for denoising, reconstructing missing channels, and upsampling scalp EEG signals.
- https://github.com/jundot/omlx — LLM inference server for Apple Silicon with continuous batching and SSD caching, managed from the macOS menu bar.
- https://github.com/HumanAIGC-Engineering/OpenAvatarChat — Modular interactive digital human conversation system; supports MiniCPM-o, Qwen-Omni, LAM, LiteAvatar, and MuseTalk avatars.
- https://github.com/msitarzewski/agency-agents — Library of 144+ specialized AI agent personalities with multi-tool support (Claude Code, Cursor, Copilot, etc.).
- https://github.com/MontrealAI/AGI-Alpha-Agent-v0 — Meta-agentic AGI research framework with autonomous agent swarms and blockchain-based economic simulations.
- https://github.com/karpathy/autoresearch — Autonomous AI research on single-GPU nanochat training; agent iterates on train.py overnight.
- https://github.com/davebcn87/pi-autoresearch — Skill package for the pi coding agent, inspired by karpathy/autoresearch above, that runs a domain-agnostic autonomous optimization loop (edit → benchmark → keep/revert → repeat) on any measurable code metric (build time, test speed, bundle size, training metrics); survives restarts via
.auto/prompt.mdand.auto/log.jsonl, uses MAD-based confidence scoring after 3+ runs to separate real improvements from measurement noise, and supports.auto/checks.shbackpressure plusbefore.sh/after.shhooks, with a dashboard widget and fullscreen live view. 8k+ stars, MIT. - https://github.com/awslabs/agentcore-samples — Amazon Bedrock AgentCore samples; deploy and operate AI agents securely at scale (framework- and model-agnostic) with runtime, gateway, identity, memory, tools, and observability examples.
- https://github.com/zeronsh/zeron — Native local-first control plane (Rust, gpui) for Claude Code, Codex, Cursor, Devin, Grok, Hermes, Pi, and Antigravity. Sessions stay on the device unless you sign in; synced devices can drive each other's workspaces (gitignored files included when shown;
.gitexcluded). See Zeron. - https://github.com/earendil-works/pi — Self-extensible AI agent toolkit (Pi); unified multi-provider LLM API, agent runtime with tool calling, TUI library, and an interactive coding agent CLI. With Pi 1.0, Earendil also shipped experimental Pi Durable (
@earendil-works/pi-durable): a crash-survivable harness for long-running multi-conversation agent apps (memory/JSONL/SQLite, extensions/tools/hooks/tasks, compaction, documents) that does not replace the coding agent — announcement: https://earendil.com/posts/pi-durable/ - https://github.com/bytedance/deer-flow — DeerFlow (Deep Exploration and Efficient Research Flow); open-source long-horizon super agent harness that researches, codes, and creates using sandboxes, memory, tools, skills, subagents, and a message gateway for tasks taking minutes to hours.
- https://github.com/NousResearch/hermes-agent — Agent with a built-in learning loop; creates skills from experience, improves them during use, searches its own past conversations, and builds a deepening model of the user across sessions.
- https://github.com/oobabooga/textgen — Open-source desktop app for local LLMs (formerly text-generation-webui); text, vision, and tool-calling with an OpenAI/Anthropic-compatible API, multiple inference backends, and 100% private operation.
- https://github.com/triggerdotdev/trigger.dev — Build and deploy fully-managed AI agents and workflows; open-source background jobs platform with a task runner, queues, and dashboard.
- https://github.com/google/agents-cli — CLI and skills that turn any coding assistant into an expert at creating, evaluating, and deploying AI agents on Google Cloud, built around Google's Agent Development Kit (ADK).
- https://github.com/google/ax — Google's pre-stable agent orchestration runtime (Apache-2.0, Go). Declare
Task,Workspace, andModel; AX sandboxes each task as an actor on Agent Substrate, materializes git/MCP/skills, and can suspend and resume the/workspacevolume. The product site also names a Gateway primitive thatmaindoes not define yet. Not the accessibility tree. See Google AX. - https://developers.cloudflare.com/web-search/ — Cloudflare Web Search API (open beta, 2 October 2026). One call through AI Gateway to Ceramic.ai, Exa, or Linkup; same result shape; list price on AI Gateway credits or a stored provider key. See Cloudflare Web Search API.
- https://anymd.cc/ — anymd (Digitop.ai, MIT). Prefix a public URL with
anymd.cc/for Markdown. Signed-in reads land in a private library (BM25, FTS5, bge-m3, hybrid RRF). Remote MCP atanymd.cc/mcp. Search is free. Jev here is a TypeSafe tie-breaker, not Decision 2.0. See Anymd. - https://github.com/Kaelio/ktx — ktx (npm
@kaelio/ktx0.16.0, Apache-2.0). Local context layer for data agents: warehouse plus docs become a wiki and a semantic layer, served over CLI and MCP. Read-only connections, per the README. Node>=22. Not the unscoped npm packagektx, and not the Khronos KTX texture format. See ktx. - https://github.com/morluto/rea — REA (npm
rea-agents6.8.0, MIT). Local CLI and MCP. Static JavaScript and .NET need no native engine. Deep native analysis uses Hopper, Ghidra, IDA, or Kuna. Checked 2 November 2026: 420,599 stars, site rea.tools,maincommit93b8b82db7after the 25 October npm release. See REA. - https://github.com/trycua/cua — Open-source computer-use agent infrastructure ("computer-use 2.0"); cross-OS drivers and fleets plus benchmarks for training, evaluation, and data generation.
- https://github.com/CopilotKit/channels-sdk — Open-source SDK that brings any AG-UI-compatible agent (CopilotKit's own, LangGraph, CrewAI, Mastra, Pydantic AI, Google ADK) into Slack, Microsoft Teams, and Discord with native, interactive UI (Block Kit/Adaptive Cards), tool calls, file handling, and human-approval gates.
- https://github.com/CopilotKit/OpenDuet — Early self-hosted template for a proactive agent on a vendored OpenBot tree (MIT, TypeScript). Mail triage and the morning brief live in
openbot/server/src/openduet. Threads and memory need a CopilotKit Intelligence key or the server will not start. See OpenDuet. - https://github.com/andrewyng/openworker. Local desktop AI coworker (MIT, Python server, Tauri shell). Checked 3 November 2026. Main package and Tauri config say 0.3.4. Newest GitHub release is v0.3.1. PyPI openworker 0.0.1 is a placeholder. Not installed. See OpenWorker.
- https://github.com/browseros-ai/BrowserOS. AGPL-3.0 Chromium monorepo with BrowserOS neo for agents and BrowserOS for people. Checked 3 November 2026. Newest GitHub release is BrowserOS neo v0.52.0. Neither app was installed. See BrowserOS.
- https://github.com/milind-soni/OpenMausBot. Local-first chat app for a team of coding-agent bots. Apache-2.0 except the enterprise directory. Checked 3 November 2026. Main is 0.1.139. Release and npm latest are 0.1.138. Not installed. See OpenMausBot.
- https://github.com/skawld/skawld-sdk — Skawld Agent SDK (
@skawld/agent-sdk): open-source (MIT) TypeScript framework that embeds a full agent loop — tools, multi-turn sessions, permission modes, streaming events, and subagents — into any Node.js 18+ / Bun 1.1+ app with one import (ESM-only). Supports Anthropic Claude plus OpenAI Chat Completions and Responses APIs, built-in tools and MCP, and SQLite-backed session persistence (in-memory store for tests); similar in spirit to the Claude Agent SDK but provider-agnostic. Docs: https://skawld.com/docs - https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf — Ternary-quantized 27B model (derived from Qwen3.8-27B) compressed to ~5.95GB while claiming 98.2% of FP16 intelligence retained; GGUF with hybrid-attention CUDA/Metal/CPU kernels for llama.cpp, 262K context, aimed at on-device deployment on laptops/single GPUs.
See also: Agent Frameworks · Multi Agent Systems · Agents Overview · OpenDuet · OpenWorker · BrowserOS · OpenMausBot