Find the open-source solution that fits.

Project overview

Atom is a self-hosted platform for running business workflows from text or voice requests. Specialized agents plan and execute work across connected services, while visual workflows, memory, and governance levels help operators control how tasks progress. Workflow data, agent state, and memory remain on infrastructure controlled by the operator. Prompts still go to a configured cloud model when bring-your-own API keys are used, unless local models such as Ollama or Llama.cpp are selected. Deployment requires operating the backend, frontend, database, and model configuration.

Project overview

Hive turns an objective into a graph-based runtime that coordinates specialized agents in parallel while preserving role memory, checkpoints, and execution state. It adds human-in-the-loop controls, observability, cost limits, browser and compute tools, and LiteLLM-compatible hosted or local models; teams building only simple one-off chains may not need this heavier harness.

Project overview

Use SkillPack to combine skills from GitHub, URLs, or local folders with prompts, policies, personas, and optional scheduled jobs. It produces a lightweight archive that teammates can unpack and launch into a local browser interface, with optional Slack or Telegram access. The runtime keeps team data on the host machine but requires Node.js 22.19.0 or newer and an LLM API key.

Project overview

DeepChat is a local-first desktop client for running cloud and local models alongside agent tools. It supports installable Skills, MCP services, and ACP-compatible agents as first-class session options, while its Tape and Trace views preserve context, tool calls, requests, and results for inspection or recovery. Users can also control sessions through messaging integrations such as Telegram, Feishu/Lark, Discord, and others, making it suited to longer-running agent workflows.

Project overview

CoWork OS is a GUI-first desktop application with a companion `cowork` CLI for running AI-assisted work in one local workspace. It brings coding, email, research, browser tasks, documents, spreadsheets, presentations, agent sessions, and automations into a shared operator console instead of separate apps. Long-running tasks can be organized through session timelines, boards, teams, plugins, and multiple model providers, while normal local CLI use shares the desktop configuration. It is aimed at users who want a broad, locally governed environment for complex AI workflows.

Project overview

OpenSail lets teams build reusable agents, applications, and automations inside workspaces connected to systems such as Slack, email, issue trackers, repositories, internal APIs, and MCP servers. Jobs can start from schedules, buttons, webhooks, messages, or app events, with approvals, budgets, permissions, logs, and run history added as needed. Its desktop and web interfaces can run projects through local processes, Docker Compose, or Kubernetes for teams that want control over both agent behavior and deployment.

Project overview

Memoh gives each agent an isolated workspace with its own filesystem, desktop, browser, network, tools, and long-term memory. You can run its built-in agent with your own API keys or host existing Claude Code and Codex agents, then reach them through Telegram, Discord, Lark, WeChat, the web interface, and other channels. Agents can keep running scheduled and browser-based work on a server after your laptop is closed. The full platform is currently self-hosted; the zero-setup Memoh Cloud offering is still on a waitlist.

Project overview

AgenticX combines a Python SDK, the agx CLI, a Studio server, and the Machi desktop app into one platform for building and operating multi-agent systems. It supports focused automation assistants, graph-based workflows, and collaborating agent teams, with multiple model providers, MCP tools, hierarchical memory, managed skill lifecycles, sandboxed execution, and session management. Developers can embed the core ReAct agent primitive or manage runs through the studio and server interfaces.

Project overview

Devo is a model-neutral coding agent with both a desktop application and a terminal TUI/CLI backed by the same local Rust runtime. It connects to OpenAI-compatible or Anthropic-compatible APIs, DeepSeek, Qwen, Kimi, private gateways, and local endpoints, while adding semantic code search, MCP servers, reusable skills, and multi-agent work. It is designed for private and intranet environments as well as normal local use. Current macOS desktop builds are unsigned and require manually removing the quarantine attribute after installation.

Project overview

SkVM profiles a model-and-agent-runtime combination, compiles a skill against those capabilities, and then evaluates and improves it through precompilation and runtime-optimization workflows. It is aimed at teams that need the same skill to behave across heterogeneous agents rather than tuning prompts by hand for each runtime. Provider configuration is required, and the bundled benchmark suite lives in a separate data submodule.

Project overview

SwarmClaw is a self-hosted runtime and dashboard for operating individual agents, swarms, and orchestrators. It provides durable memory, schedules and heartbeats, delegation, background jobs, MCP tools, and connections to many model APIs and coding-agent CLIs. It is aimed at long-running workflows that need observation and operator control, rather than an agent library embedded in another application.

Project overview

Skrun is a self-hostable, multi-model agent runtime that turns SKILL.md, AGENTS.md, or another instruction format into a POST /run endpoint. Its CLI initializes, versions, verifies, and invokes agents, while the runtime adds streaming events, asynchronous webhooks, tool calling, persistent state, an operator dashboard, and generated API documentation. Users bring their own model credentials, and the current self-hosted setup requires Node 20 or later with SQLite or Postgres.

Project overview

NanoClaw hosts personal AI agents in separate Linux containers and connects them to channels such as Telegram, Discord, and WhatsApp, with per-agent workspaces, memory, and scheduled tasks. Channels and alternative model providers are added on demand, while customization is primarily code-based; setup requires Docker, Node, pnpm, and Claude Code.

Project overview

Set up OpenClaw, Hermes, or LangChain Deep Agents inside NVIDIA OpenShell with a reference stack that combines guided onboarding, routed inference, network policy, and sandbox lifecycle management behind one CLI. The project focuses on safer long-running agent operation and controlled egress. It is currently described as alpha, with community support provided on a best-effort basis.

Project overview

nanobot is a lightweight personal agent runtime that works through a browser UI, terminal, or chat services including Telegram, Discord, WeChat, Slack, and email. It supports tools, web access, MCP, memory, scheduled automation, model routing, a Python SDK, and an OpenAI-compatible API; installation requires Python 3.11 or newer and a configured model provider.

Project overview

IronClaw runs a personal AI assistant through a CLI, web interface, and messaging channels while keeping its state under your control. It supports multiple model providers, MCP connections, scheduled routines, and extensible tools, with WASM sandboxing and credential isolation for untrusted operations. Building the full system from source requires a Rust and PostgreSQL environment.

Project overview

PicoClaw packages a personal AI assistant as a portable Go binary for x86, ARM, RISC-V, MIPS, and other architectures, with a reported core memory footprint below 10 MB. It can connect to many model providers and chat channels, manage MCP server configurations, schedule jobs, and spawn asynchronous sub-agents. Most hosted model choices still require their corresponding credentials.

Project overview

Galley runs Claude Code, Codex, or GLM tasks in managed Git worktrees, records commands and diffs, and asks a supervisor to accept, retry, escalate, or stop each attempt. Repository profiles define checks, scope, secrets policy, and optional pull-request handoff, while evidence remains on disk for later inspection. Task YAML is trusted execution input, so Galley should be used only with task authors and repositories you trust.