7 Open-Source AI Agent Frameworks Worth Testing in 2026

The AI agent landscape moved fast in early 2026. The big names (LangGraph, CrewAI, AutoGen) are still worth knowing (I covered those in our AI agent frameworks comparison), but a new wave of open-source projects is where the interesting experiments are happening right now.

This post walks through seven of them. Some are full personal assistants you can run on a Raspberry Pi. Some auto-generate multi-agent workflows from a single prompt. One is a self-evolution engine that plugs into the agent you already use.

All seven are on GitHub. None of them require a managed platform. Here's what each one is, and when it actually fits.

Why these seven?

They don't all do the same thing. I split them by what they are best at:

Three are in the "personal assistant that lives in your messaging apps" lane (OpenClaw, ZeroClaw, NanoClaw, Hermes). Two are about workflows that evolve themselves (EvoAgentX, Evolver). Agent Zero is the general-purpose wildcard.

1. Agent Zero: An organic, general-purpose agent

Agent Zero pitches itself as a personal, organic agentic framework that grows and learns with you. It is deliberately not a predefined framework. Behavior is defined almost entirely by the system prompt in prompts/default/agent.system.md, and the agent uses your operating system (code, terminal, files) as its toolbox rather than a fixed tool registry.

What stands out

• Multi-agent cooperation. Any agent can spawn a subordinate to handle a subtask, keeping each agent's context clean.
• Persistent memory. Remembers solutions, code snippets, and facts across sessions.
• Skills (SKILL.md). Supports the open SKILL.md standard, so skills work across Claude Code, Cursor, Goose, Codex CLI, and Copilot.
• Dockerized. Ships with TTS/STT, web UI, and a browser agent plugin.
• Transparent by design. No hidden rails. Every prompt and message template lives in the prompts/ folder.

When to pick it

When you want a sandboxed agent that can write its own tools on the fly and you're comfortable driving it with strong prompts instead of a fixed workflow.

2. OpenClaw: The lobster multi-channel assistant

OpenClaw is the 363k-star elder of this cohort and the reference point most of the "personal assistant" forks below migrated from. It's a Node-based local-first Gateway that connects your agent to most major messaging surfaces and ships macOS/iOS/Android companion apps with voice wake and a Live Canvas.

What stands out

• Multi-channel inbox. WhatsApp, Telegram, Slack, Discord, Google Chat, Signal, iMessage (BlueBubbles), Microsoft Teams, Matrix, LINE, WeChat, QQ, and many more.
• Multi-agent routing. Route inbound channels/accounts to isolated agents with their own workspaces and sessions.
• Live Canvas + Voice Wake. Agent-driven visual workspace and wake-word voice on macOS/iOS, continuous voice on Android.
• Skills registry (ClawHub). Bundled, managed, and workspace skills with a central registry.
• MIT licensed with a huge contributor base and near-daily releases.

When to pick it

When you want the most mature multi-channel assistant with companion apps, and you don't mind a larger codebase in exchange for a much bigger feature surface.

3. Hermes Agent: A self-improving agent with a learning loop

Hermes Agent is Nous Research's entry and the only one in this list with a built-in learning loop as its headline feature. It creates skills from experience, improves them during use, searches its own past conversations, and builds a deepening model of who you are across sessions.

What stands out

• Closed learning loop. Agent-curated memory with nudges, autonomous skill creation, FTS5 session search with LLM summarization, and Honcho dialectic user modeling.
• Lives where you do. Telegram, Discord, Slack, WhatsApp, Signal, Email, CLI. One gateway process for all of them.
• Model-agnostic. Nous Portal, OpenRouter (200+ models), NVIDIA NIM (Nemotron), Xiaomi MiMo, GLM, Kimi/Moonshot, MiniMax, Hugging Face, OpenAI, or your own endpoint.
• Six terminal backends. Local, Docker, SSH, Daytona, Singularity, Modal. Daytona and Modal offer serverless hibernation so idle cost is near zero.
• Scheduled automations + parallel subagents. Cron jobs, delegated isolated subagents, RPC-style tool calls.
• Research-ready. Batch trajectory generation and Atropos RL environments for training tool-calling models.
• Migrates from OpenClaw via hermes claw migrate (SOUL.md, memory, skills, keys, workspace instructions).

When to pick it

When you want an assistant that actually gets better the more you use it, and you care about running it on cheap remote infra instead of your laptop.

For a step-by-step setup, see the Hermes Agent self-host guide on Hetzner. It covers provisioning, Telegram wiring, and running the gateway under systemd.

4. ZeroClaw: A Rust personal assistant for $10 hardware

ZeroClaw is a single-binary Rust rewrite of the "personal AI assistant that lives in your messaging apps" idea. It runs in a few-megabyte memory envelope (under 5MB RAM per the project's benchmarks) and boots in under 10ms. The readme claims it uses 99% less memory than OpenClaw and works on hardware as cheap as $10.

What stands out

• Multi-channel inbox. WhatsApp, Telegram, Slack, Discord, Signal, iMessage, Matrix, Email, Bluesky, Reddit, LinkedIn, and more. 20+ channels total.
• Hardware peripherals. First-class support for ESP32, STM32, Arduino, and Raspberry Pi via a Peripheral trait.
• Autonomous agent swarms (Hands). Background agents that run on cron and grow smarter over time.
• Swappable everything. Providers, channels, tools, memory, tunnels are all traits you can replace.
• Dual MIT / Apache-2.0 license. Friendly for both research and commercial use.

When to pick it

When you want an always-on, low-resource assistant you can actually deploy on an SBC or edge device without a full Node.js or Python runtime.

5. NanoClaw: Container-isolated Claude in your messaging apps

NanoClaw is a deliberate minimalist take on the OpenClaw pattern. Its premise: a full-featured assistant with shared-memory access to your whole system is scary, and allowlists aren't real isolation. NanoClaw runs each agent in its own Docker container and keeps the codebase small enough that you (or Claude Code) can read the entire thing.

What stands out

• True container isolation. Agents can only see what's explicitly mounted. Credentials route through OneCLI's Agent Vault, so raw API keys never touch the container.
• Skills over features. Channel adapters (Telegram, Discord, Slack, WhatsApp, Teams, iMessage, Matrix, Google Chat, etc.) are installed on demand via /add-<channel> skills on a channels branch.
• AI-native setup. When install fails, nanoclaw.sh hands off to Claude Code to diagnose and resume.
• Customization = code changes. No config sprawl. Tell Claude Code what to change and it patches your fork.
• Built on Anthropic's official Claude Agent SDK by default, with /add-codex, /add-opencode, and /add-ollama-provider for other models.

When to pick it

When you like OpenClaw's shape but want real OS-level sandboxing, a tiny codebase you can actually audit, and an AI-native install story.

6. Evolver: Auditable self-evolution for any agent

Evolver is the odd one out: it's not an agent. It's a GEP (Gene Evolution Protocol) engine that plugs into the agent runtime you already use (Cursor, Claude Code, OpenClaw, or via MCP) and turns ad-hoc prompt tweaks into auditable, reusable evolution assets called Genes, Capsules, and EvolutionEvents.

What stands out

• Auto-log analysis. Scans memory and history files for errors and patterns, then emits repair-focused directives.
• Mutation + personality evolution. Each run is gated by an explicit Mutation object and an evolvable PersonalityState.
• Strategy presets. EVOLVE_STRATEGY=balanced|innovate|harden|repair-only switches the intent mix.
• Skill Store. evolver fetch --skill <id> pulls reusable skills from the EvoMap network.
• Review mode. evolver --review for human-in-the-loop evolution instead of autopilot.
• Protocol-bound changes. Designed for teams that need deterministic, audit-friendly modifications to their agent stack.

One caveat: the maintainers announced that future releases will transition from fully open source to source-available. All already-published MIT and GPL-3.0 versions remain usable, and the npm package and repo still work. Plan accordingly if permissive licensing matters to you.

When to pick it

When you're maintaining agent prompts at scale and need an audit trail, or when your environment requires deterministic and protocol-bound changes instead of free-form edits.

7. EvoAgentX: Auto-built, self-evolving workflows

EvoAgentX takes a different angle: you describe a goal in natural language and it generates the multi-agent workflow for you, then keeps improving it with self-evolving algorithms driven by your dataset. It's the most research-flavored project on this list (there's a paper and a survey to go with it), but the API is surprisingly clean.

What stands out

• Workflow autoconstruction. From a single prompt, WorkFlowGenerator builds a structured multi-agent graph.
• Self-evolution engine. TextGrad, AFlow, and other optimizers improve prompts and structure over time.
• HITL checkpoints. Insert a human reviewer anywhere in the workflow with an interceptor agent.
• Short + long-term memory plus a rich built-in tool library (code interpreters, search, filesystem, MongoDB/Postgres/FAISS, browser automation, image tools, MCP).
• Works with OpenAI, Claude, Qwen, DeepSeek, Kimi, local LLMs through LiteLLM / SiliconFlow / OpenRouter.

When to pick it

When your project needs to improve the workflow, not just run it. For example, when you have an eval set and want the system to optimize against it, similar in spirit to MAKER's decomposed agentic processes.

Minimal example:

from evoagentx.workflow import WorkFlowGenerator, WorkFlow
from evoagentx.agents import AgentManager
 
goal = "Generate html code for the Tetris game"
workflow_graph = WorkFlowGenerator(llm=llm).generate_workflow(goal)
 
agent_manager = AgentManager()
agent_manager.add_agents_from_workflow(workflow_graph, llm_config=openai_config)
 
workflow = WorkFlow(graph=workflow_graph, agent_manager=agent_manager, llm=llm)
print(workflow.execute())

How to pick one

A quick shortcut if you're scanning:

• Want a general-purpose agent that writes its own tools? Agent Zero
• Want the most mature multi-channel assistant with voice + canvas? OpenClaw
• Want an agent with a native learning loop across sessions? Hermes Agent
• Want it to run on a Raspberry Pi or ESP32? ZeroClaw
• Want real OS-level sandboxing and a codebase you can read end-to-end? NanoClaw
• Want auditable evolution bolted onto the agent you already use? Evolver
• Want the workflow itself to improve over time? EvoAgentX

There's real overlap. OpenClaw, ZeroClaw, NanoClaw, and Hermes are all in the "personal assistant across your messaging apps" lane, and a lot of users migrate between them. If you're just getting started, I'd pick based on the language stack you're most comfortable hacking on: Rust for ZeroClaw, TypeScript for OpenClaw or NanoClaw, Python for Hermes or Agent Zero.

Whichever one you pick, once an agent starts acting on its own (sending messages, calling paid APIs, touching customer data), you need to know which agent did what and on whose authority. Kakunin↗ (opens in a new tab) covers that layer: it issues each agent an X.509 certificate tied to its operator and permitted actions, then keeps an audit trail of what the agent does.

For a look at how the more traditional orchestration frameworks compare, see the CrewAI / LangGraph / AutoGen / LangChain / Swarm breakdown. And if you're designing for reliability on very long workflows, the MAKER paper on million-step zero-error agents is worth the read.

Write better prompts for your agents

All seven of these frameworks live or die by the prompts you give them. I use promptbuilder.cc↗ (opens in a new tab) to draft mine.

• Multi-model: ChatGPT, Claude, Gemini, Grok, Perplexity, DeepSeek, Llama, Mistral, Cohere
• Model-specific generators that rewrite the same idea for each model's strengths
• Free, no extension, no signup to try
• Prompt engineering reference hub built in

Good prompts make good agents. Try promptbuilder.cc →↗ (opens in a new tab)

Want to keep up as these frameworks and models evolve? Prompt Insights↗ (opens in a new tab) covers AI news and prompt-engineering updates daily.

Create social content for your AI launch

The carousel and social images for this post were built with instaposts.ai. AI carousels for Instagram and LinkedIn in seconds. Branded, on-brand, no Canva.

Launching an AI tool? Here is a list of AI directories you can submit to, plus a quick basic SEO guide for getting the post-launch stuff right.