The Best OpenClaw Alternatives 2026 – from NanoClaw to NullClaw

21. Februar 2026Updated: June 4, 202622 min readDeep Dive

TL;DR: „OpenClaw is powerful but no longer the only option. NanoClaw wins on security, Nanobot on simplicity, memU on memory – and Microsoft Scout (announced) brings the first serious enterprise gateway. For local-AI-first stacks, pair them with the announced NVIDIA RTX Spark."

— Till Freitag

Why Alternatives? OpenClaw Has 200,000+ Stars After All

OpenClaw is the dominant open-source AI agent. Autonomous actions, 50+ messaging integrations, a massive plugin ecosystem – on paper, hardly any tool can compete.

But: 430,000+ lines of code also mean 430,000 lines of potential attack surface. Security researchers at Palo Alto Networks have called OpenClaw a "security nightmare." There have been cases where the agent independently made purchases or spammed contacts. On top of that, Anthropic killed third-party tool coverage under Claude subscriptions – bills jump from $20 to $500 fast.

Not everyone needs a "God Mode" agent. By mid-2026, the market has sorted itself: NanoClaw, Nanobot, and memU are established, Microsoft Scout has been announced as the first hyperscaler gateway, and the announced NVIDIA RTX Spark should make the entire stack runnable locally once available.

Our Top 12 at a Glance

Tool	Focus	Stars	Architecture	Standout Feature
NanoClaw	Security-first	8,400+	Single Process	Container isolation, WhatsApp
Nanobot	Ultra-lightweight	29,100+	4K lines Python	99% smaller than OpenClaw
memU	Long-term memory	8,200+	Knowledge Graph	Proactive agent
OpenCode	Coding agent	13,400+	Go CLI	Open source, multi-LLM
NullClaw	Edge & Minimal	3,100+	Zig Single Binary	678 KB, 22+ LLM providers
ZeroClaw	Rust performance	31,500+	Minimal Runtime	NullClaw successor in Rust
OpenFang	Agent OS	17,800+	Agent Operating System	7 autonomous "Hands", 38 tools
Moltworker	Serverless	–	Cloudflare Workers	No local access needed
SuperAGI	Multi-Agent	17,200+	Framework	Multiple agents in parallel
Anything LLM	LLM Hub	34,000+	Self-hosted	Multi-LLM, RAG, Plugins
Claude Code	Development	–	CLI/IDE	Coding focus, Anthropic
Microsoft Scout	Enterprise gateway	–	Managed (Azure)	OpenClaw-compatible, M365-native

💡 Hardware layer (announced): Once NVIDIA RTX Spark is available, virtually every open-source candidate on this list should run locally at 1,700 tokens/s – local-AI-first would then no longer be a niche topic but a real strategic option.

🧭 Deep dives per layer: Coding-Agent Layer · Multi-Agent Layer · Self-Hosted & Privacy Layer · Enterprise Gateway Layer (with decision flowchart)

1. NanoClaw – The Security Champion

Best for: Teams that need container isolation and WhatsApp control

NanoClaw is the radical answer to OpenClaw's security problems. Instead of 430,000 lines of code: 5 files, one process. Instead of unrestricted host access: Linux containers with filesystem isolation.

What Makes NanoClaw Special

Container Isolation: Agents run in Docker or Apple containers – even if the agent goes rogue, only the sandbox is affected
Native WhatsApp: Each WhatsApp group gets an isolated context with its own memory files
Raspberry Pi Support: Runs on a Pi 4 with 4GB RAM
Agent Swarms: Coordinate multiple Claude instances for complex tasks

Setup

git clone https://github.com/gavrielc/nanoclaw.git
cd nanoclaw
claude   # → /setup

Limitations

NanoClaw is Claude-only – no multi-LLM support. The plugin ecosystem is minimal. If you need enterprise integrations with Jira or Salesforce, look elsewhere.

License: MIT | GitHub | Our Analysis

2. Nanobot – 99% Less Code, Same Core Function

Best for: Developers who want to understand the entire codebase in an afternoon

Nanobot from Hong Kong (HKU) delivers OpenClaw core features in 4,000 lines of Python – with an impressive 26,800+ GitHub stars. The entire codebase can be read in a few hours – with OpenClaw, you'd need months.

Features

Persistent Memory: Conversations are saved across sessions
Web Search: Integrated web search for current information
Background Agents: Sub-agents for parallel tasks
Telegram & WhatsApp: Control via chat apps
MCP-based: Standardized tool integration

When to Choose Nanobot?

Nanobot is the perfect learning project. Want to understand how AI agents work? Fork Nanobot and build your feature. The minimal codebase makes it the ideal starting point for custom agents.

Limitations

Only 2 messaging platforms, no plugin marketplace, no GUI. Too bare-bones for enterprises – but that's exactly the point.

License: Open Source | GitHub Stars: 26,800+ | GitHub

3. memU – The Agent That Remembers Everything

Best for: Users who want a personal assistant that learns over time

Most agents forget everything when you close the session. memU doesn't. It builds a local knowledge graph of your preferences, projects, and habits – and gets smarter over time.

What Makes memU Special

Hierarchical Knowledge Graph: Not just flat memory files, but networked knowledge structures with RAG
Proactive Actions: memU acts based on context and behavior – without explicit commands
Token Optimization: Context is compressed before the API call, saving costs
Local-first: Everything stays on your device

Use Case

"You have the quarterly review tomorrow – should I summarize the latest performance data?"

memU recognizes recurring patterns and proactively offers help – like an assistant who knows you better than you know yourself.

Limitations

memU is more secretary than coder. For raw execution (writing code, bash commands, API calls), OpenClaw is stronger. memU excels at understanding and anticipating, not executing.

GitHub Stars: 6,900+ | GitHub

4. OpenCode – The Open-Source Coding Agent

Best for: Developers who want a free, fully open-source alternative to Claude Code

OpenCode is an AI coding agent written in Go for the terminal – with 11,100+ GitHub stars and an MIT license. Unlike Claude Code, OpenCode is fully open source and supports multiple LLM providers.

What Makes OpenCode Special

Multi-LLM: OpenAI, Anthropic, Google Gemini, local models – you choose your backend
Terminal-native: Elegant TUI (Terminal UI) with syntax highlighting and diff views
Multi-File Editing: Understands project structures and edits multiple files simultaneously
LSP Integration: Language Server Protocol for precise code analysis
Session Management: Conversations are saved and can be resumed

Setup

go install github.com/opencode-ai/opencode@latest
opencode

When to Choose OpenCode Over Claude Code?

When you don't want an Anthropic subscription, need a multi-LLM setup, or value full open-source transparency. OpenCode is the "freedom" pick among coding agents.

Limitations

No IDE plugin (terminal only), no PR workflow automation like Claude Code. The community is smaller, the ecosystem younger. For raw coding power, Claude Code still leads – but OpenCode is catching up fast.

License: MIT | GitHub

5. NullClaw – The Minimalist Among Agents

Best for: Edge deployments and environments with minimal resources

NullClaw takes minimalism to the extreme: An AI agent written in Zig that compiles to a single 678 KB binary. No runtime needed – runs even on $5 ARM hardware.

What Makes NullClaw Special

Smallest Footprint: 678 KB single binary – no Node.js, no Python, no dependencies
22+ LLM Providers: OpenAI, Anthropic, Mistral, Ollama, and many more
17 Messaging Channels: From Slack to Telegram to Discord
Zero-Dependency: Statically compiled, runs on virtually any hardware
Edge-ready: Ideal for IoT, Raspberry Pi, embedded systems

Setup

# Download pre-built binary
curl -sSL https://nullclaw.dev/install.sh | bash
nullclaw --llm ollama --model llama3

When to Choose NullClaw?

When you need an agent on resource-constrained hardware – edge devices, IoT gateways, old servers. Or when you fundamentally don't want runtime overhead. NullClaw is the "bare metal" pick.

Limitations

Young community (2,600+ stars), less documentation than established alternatives. Zig as a programming language is niche – writing custom plugins requires Zig expertise.

License: MIT | GitHub Stars: 2,600+ | GitHub

6. Moltworker – OpenClaw in the Cloud, Without Risk

Best for: Users who want OpenClaw power but don't want to install anything locally

Moltworker is Cloudflare's official adaptation of OpenClaw for Cloudflare Workers. The agent runs serverless in a sandbox – no access to your local system, no security risk.

Advantages

Serverless: No server management, no local installation
Sandboxed: The agent can only operate within the Cloudflare environment
Persistent State: State management via Cloudflare infrastructure
Global Edge: Runs globally distributed with low latency

Limitations

No access to local files or shell commands. If you need an agent that works with your filesystem, Moltworker isn't the right choice. Ideal for cloud-based assistance without installation overhead.

License: Open Source | GitHub

7. SuperAGI – The Multi-Agent Framework

Best for: Developers who want to orchestrate multiple specialized agents

SuperAGI isn't a finished product – it's a framework. You build your own agents with it – with custom logic, dedicated memory, and specific tools.

Features

Multi-Agent: Multiple agents work in parallel on different tasks
Long-term Memory: Built-in storage for context across sessions
Plugin System: Extensible with community plugins
Self-hosted: Full control over data and infrastructure
15,000+ GitHub Stars: Large, active community

When to Choose SuperAGI?

When you need a system where Agent A monitors the inbox, Agent B updates CRM data, and Agent C creates the weekly report – then SuperAGI is your framework.

Limitations

Steeper learning curve than finished products. You need to configure agents, define reasoning logic, and build integrations yourself. Not for non-developers.

License: Open Source | GitHub

8. Anything LLM – The Swiss Army Knife

Best for: Builders who want a self-hosted LLM hub with full transparency

Anything LLM isn't an agent in the traditional sense – it's a platform for working with LLMs. You upload documents, connect APIs, switch between models, and have full control over every prompt.

Features

Multi-LLM: OpenAI, Anthropic, local models – all through one interface
RAG: Load documents and chat about them (PDF, CSV, etc.)
Self-hosted: Runs on your server, your data stays with you
Plugin System: Extensible with web search, code execution, etc.
30,000+ GitHub Stars

Limitations

Anything LLM doesn't automate proactively. You need to initiate every interaction manually. It's a thinking tool, not an acting tool. Ideal for experimenting, not for automating.

License: Open Source | GitHub

9. Claude Code – The Coding Specialist

Best for: Developers who want a secure, focused code assistant

Claude Code is Anthropic's official CLI tool for developers. Not a general agent – but a pair programmer that understands your entire codebase.

Features

Multi-File Refactoring: Understands connections across file boundaries
PR Workflows: Generates code, tests, and pull requests from issues
Sandboxed: Suggests changes but executes nothing without confirmation
IDE Integration: Terminal, VS Code, JetBrains

Limitations

Coding only. No emails, no calendar, no WhatsApp. If you're looking for a personal assistant, Claude Code isn't the answer. But for software development, it's one of our favorite tools.

Price: From ~$20/month (Claude Pro) | Website

10. ZeroClaw – NullClaw's Big Brother in Rust

Best for: Teams that want Rust performance with a strong community

ZeroClaw is the spiritual successor to NullClaw – written in Rust instead of Zig, with a community nearly 10x larger. The agent compiles to a single binary with a 99% smaller footprint than OpenClaw, but offers a significantly more mature feature set than NullClaw.

What Makes ZeroClaw Special

Rust Performance: Blazing fast, memory-safe, no garbage collection
Modular Architecture: Plugins in Rust or via FFI
Self-hosted: Full control over data and infrastructure
Edge-ready: Small binary, runs on resource-constrained hardware
26,800+ GitHub Stars: Strong, growing community

Setup

curl -sSL https://zeroclaw.dev/install.sh | bash
zeroclaw --llm ollama --model llama3

Limitations

Rust has a steep learning curve. The plugin ecosystem is younger than OpenClaw's. For teams without Rust experience, extending the agent is harder.

License: Apache 2.0 | GitHub Stars: 26,800+ | GitHub

11. OpenFang – The Agent Operating System

Best for: Teams that want a complete agent operating system rather than a framework

OpenFang goes a step further than all other alternatives: it positions itself not as an agent framework, but as an Agent Operating System. Also written in Rust, it offers 7 autonomous "Hands" – specialized modules for scheduling, knowledge graphs, dashboards, and more.

What Makes OpenFang Special

7 Autonomous "Hands": Scheduling, knowledge graphs, dashboard, monitoring, and more – all built-in
38 Tools: Comprehensive tool collection out of the box
40 Messaging Channels: From Slack to Discord to Telegram
26+ LLM Providers: Broad model support
1,700+ Tests: Production-grade test coverage
14,200+ GitHub Stars

Setup

cargo install openfang
# or via Docker
docker run -d openfang/openfang:latest

Limitations

OpenFang is complex – the learning curve is steeper than lightweight alternatives. Still in v0.1.0, meaning breaking changes are possible. No edge support – use ZeroClaw or NullClaw for that.

License: Apache 2.0 | GitHub Stars: 14,200+ | GitHub

12. Microsoft Scout – The Announced Enterprise Gateway from Redmond

Best for: Enterprises and mid-market organizations with an existing M365/Azure stack that want OpenClaw-style functionality without self-hosting – once available.

Microsoft Scout has been announced as the first serious hyperscaler entry into the agent gateway market. Scout is expected to speak the OpenClaw protocol and run as a managed service in Azure – including Entra ID auth, Purview audit trails, and Copilot integration. Not yet publicly available.

What makes Scout special

OpenClaw-compatible: Existing skills and gateway configs can be migrated – no lock-in break
Native enterprise auth: Entra ID, Conditional Access, Purview logs out of the box
Deep M365 integration: Teams, Outlook, SharePoint as first-class surfaces
Managed runtime: No containers, no patching, no pager duty

Limitations

Scout is not open source, tied to Azure, and not yet available. For strict GDPR use cases we recommend self-hosting + Privacy Router or the local-AI-first stack on the announced NVIDIA RTX Spark.

License: Proprietary (managed service) | Our analysis

June 2026 Additions – 10 More Alternatives You Should Know

The market keeps expanding. These ten projects are strong enough in mid-2026 to belong in any serious evaluation – grouped by layer.

Coding Agent Layer

13. OpenHands (formerly OpenDevin)

Best for: Teams wanting an autonomous software engineer with sandbox execution and a large community.

OpenHands is the de-facto reference for open-source coding agents. Docker sandbox, browser use, multi-LLM, very active roadmap. If you want "Devin, but self-hosted," this is where you land.

License: MIT | GitHub Stars: 65,000+ | GitHub

14. Aider

Best for: CLI purists who want pair-programming directly in the terminal with clean Git integration.

Aider auto-commits every change, understands large codebases via repo map, and works with Claude, GPT, Gemini, and locally via Ollama. Minimal setup overhead, maximum control.

License: Apache 2.0 | GitHub Stars: 38,000+ | GitHub

15. Devika

Best for: Learning and experimenting with planning/reasoning architectures.

The first open-source Devin implementation. Custom UI, web browsing, multi-step planning. Less production-ready than OpenHands, but didactically valuable.

License: MIT | GitHub Stars: 19,500+ | GitHub

16. SWE-agent (Princeton)

Best for: Benchmark-driven teams who need state-of-the-art coding performance.

SWE-agent defines "agent-computer interfaces" and regularly tops the SWE-bench rankings. Academically well-documented, easier to audit than monolithic frameworks.

License: MIT | GitHub Stars: 17,000+ | GitHub

17. Continue.dev

Best for: Developers who want an open-source Copilot directly in VS Code or JetBrains – with full model choice (including local via Ollama).

Configurable via YAML, tab completion, chat, edits. The most honest open-source replacement for GitHub Copilot.

License: Apache 2.0 | GitHub Stars: 28,000+ | GitHub

Multi-Agent & Framework Layer

18. AG2 (formerly AutoGen)

Best for: Multi-agent conversation patterns where several specialized agents collaborate.

Forked from Microsoft's AutoGen and independently maintained. Proven in research and production setups, solid tool integration, strong event model.

License: Apache 2.0 | GitHub Stars: 4,600+ (fork) / 35,000+ (original) | GitHub

19. LangGraph

Best for: Production stacks that want to build agents as explicit state machines instead of "prompt lottery".

Graph-based orchestration from LangChain. Persistent state stores, human-in-the-loop, checkpointing. In mid-2026, the de-facto standard for serious agent backends.

License: MIT | GitHub

20. AWS Strands

Best for: AWS shops that want to plug agents directly into Bedrock, Lambda, and IAM.

Strands is Amazon's official agent stack: model-agnostic but deeply integrated into the AWS ecosystem. The natural choice if your compliance story is built on AWS.

License: Apache 2.0 | GitHub

Self-Hosted & Privacy Layer

21. Ontheia

Best for: Teams that want to self-host an MCP-native, multi-provider agent with a genuine "GDPR by Design" architecture.

TypeScript, Docker, AGPL-3.0. Speaks Anthropic, OpenAI, Gemini, Grok, and Ollama out of the box – ideal as a complement to the Privacy Router and the announced RTX Spark.

License: AGPL-3.0 | GitHub

22. dreb

Best for: Early adopters who want a lean, provider-agnostic coding harness with fast release cycles.

TypeScript fork of pi-mono, very active iteration (v2.21+ within weeks). Small enough to understand the code in a single session.

License: MIT | GitHub

Deep Dive per Layer – Which Use Cases, Which Tools?

Coding Agent Layer – When the Agent Should Write, Run and Debug Code

Tool	Purpose	Setup Effort	Privacy / Hosting	Typical Workflows
OpenHands	Autonomous software engineer with sandbox	Docker + API key (~15 min)	Self-hosted possible, cloud APIs optional	Issue → PR, test automation, bug fixing
Aider	Terminal pair-programming with Git integration	`pip install aider-chat` (~2 min)	Local-first, own API keys	Refactoring, code review, commit assistant
Devika	Learning & experimenting with planning architectures	Docker Compose (~10 min)	Self-hosted, UI included	Planning demos, architecture studies
SWE-agent	SWE-bench performance & research	Docker + Python (~10 min)	Self-hosted, academically audited	Benchmarking, paper reproduction
Continue.dev	IDE copilot with model freedom	VS Code extension (~1 min)	Local/Ollama-ready, no cloud required	Autocomplete, inline chat, local models
dreb	Provider-agnostic coding harness	npm / TypeScript (~5 min)	Self-hosted, minimal footprint	Multi-provider tests, fast iteration
Claude Code	Premium coding with Anthropic quality	CLI install (~2 min)	Cloud (Anthropic), no local option	Enterprise refactoring, PR workflows
OpenCode	Open-source coding agent in terminal	`go install` (~5 min)	Self-hosted, multi-LLM	TUI-based editing, LSP integration

This is software engineering territory: open PRs, get tests green, ship migrations, fix bugs. The agent needs repo access, a sandbox, and solid tool-use capabilities.

Typical use cases:

Ship a greenfield feature: Issue → PR with tests, no babysitting. → OpenHands (sandbox + browser) or Devika (for transparent planning steps).
Refactors & bugfixes in existing repos: Repo map needed, fast iteration. → Aider (terminal, auto-commit) or Claude Code (premium quality).
Inline autocomplete & chat in the IDE: Daily pair-programming. → Continue.dev (with Ollama for local) or Claude Code.
SWE-bench benchmarking & research: Reproducible, auditable runs. → SWE-agent.
Multi-provider experiments without lock-in: Switch models fast. → dreb or OpenCode.

⚡ Quick-Select: Coding Agent Layer

Criterion	Recommended Tool	Why
Fastest start	Continue.dev	VS Code extension in 1 minute, no infrastructure
Highest privacy control	OpenCode	Go binary, multi-LLM, fully open source, no cloud required
Best overall package	Aider	`pip install` in 2 minutes, local-first, Git integration, works with Ollama

Till Freitag recommendation: OpenHands as workhorse + Aider for CLI speed + Continue.dev in the IDE. For sensitive repos, use local models (e.g., via Ollama) or wait for the announced RTX Spark.

Multi-Agent & Framework Layer – When One Agent Isn't Enough

Tool	Purpose	Setup Effort	Privacy / Hosting	Typical Workflows
AG2	Multi-agent conversations & research	`pip install ag2` (~5 min)	Self-hosted, no cloud required	Research pipelines, report generation
LangGraph	Production agents as state machines	`pip install langgraph` (~5 min)	Self-hosted, persistent stores	Human-in-the-loop, checkpoints, retry
AWS Strands	AWS-native agent backend	AWS CLI + IAM (~30 min)	AWS cloud (Bedrock, Lambda)	Enterprise compliance, audit trails
SuperAGI	Framework for parallel specialist agents	Docker Compose (~20 min)	Self-hosted, full control	Inbox monitoring, CRM updates, reports
Anything LLM	Self-hosted multi-LLM hub	Docker or desktop (~10 min)	Self-hosted, local data	RAG document chat, knowledge base
OpenFang	Complete agent operating system	`cargo install` or Docker (~15 min)	Self-hosted, Rust stack	Scheduling, knowledge graphs, monitoring

As soon as tasks require specialist roles (researcher, coder, reviewer, QA) or explicit steps, a framework pays off. What matters here is state management, observability, and production hardening – not demo polish.

Typical use cases:

Research pipelines & report generation: Multiple agents collaborate over days. → AG2 (conversation patterns) or SuperAGI (pre-built business templates).
Production agents with human-in-the-loop: Approval steps, checkpoints, retry logic. → LangGraph (state machines, persistent stores).
AWS-native backend with a compliance story: Bedrock, Lambda, IAM, audit trails. → AWS Strands.
Multi-LLM hub with RAG for internal teams: Knowledge base + chat UI + plugins. → Anything LLM.
Full agent OS with custom "hands": Custom tool suite, multiple autonomous sub-agents. → OpenFang.

⚡ Quick-Select: Multi-Agent & Framework Layer

Criterion	Recommended Tool	Why
Fastest start	AG2	`pip install ag2` in 5 minutes, no cloud required
Highest privacy control	Anything LLM	Self-hosted, local data, Docker or desktop in 10 minutes
Best overall package	LangGraph	`pip install langgraph` in 5 minutes, self-hosted, persistent stores, production-ready

Till Freitag recommendation: LangGraph as orchestrator layer + AG2 for multi-agent patterns + Anything LLM as internal knowledge interface. Strands if AWS is already set.

Self-Hosted & Privacy Layer – When No Data Is Allowed to Leave the Building

Tool	Purpose	Setup Effort	Privacy / Hosting	Typical Workflows
Ontheia	MCP-native agent, GDPR by design	Docker Compose (~10 min)	Self-hosted, AGPL-3.0	Contract analysis, sensitive documents
NanoClaw	Container-isolated WhatsApp bots	`git clone` + `claude` (~5 min)	Self-hosted, Linux containers	Customer communication, isolated agent swarms
ZeroClaw	Rust performance on edge hardware	`curl \| bash` (~2 min)	Self-hosted, single binary	IoT gateways, Raspberry Pi, embedded
NullClaw	Minimalist agent for edge	`curl \| bash` (~2 min)	Self-hosted, 678 KB binary	ARM hardware, resource-constrained environments
NVIDIA RTX Spark (announced)	Local inference with cloud performance (once available)	Hardware + Ollama (~30 min)	On-premise, no data leakage	122B models locally, GDPR-compliant

Regulated industries (healthcare, finance, public sector) and anyone serious about GDPR need architectures where no token hits US cloud APIs. This layer combines self-hosting, sandboxing, and local models.

Typical use cases:

Analyze sensitive documents without the cloud: Contracts, patient data, financial reports. → Ontheia (MCP-native, GDPR by design) + local models via Ollama.
WhatsApp/messaging bot without data leakage: Customer comms with container isolation. → NanoClaw.
Edge/IoT deployment on weak hardware: Raspberry Pi, industrial gateways. → ZeroClaw or NullClaw.
GDPR-compliant routing between models: Sensitive prompts local, the rest in the cloud. → Privacy Router + self-hosting guide.
Local inference with cloud performance (once available): 122B models on a mini PC. → NVIDIA RTX Spark / DGX Spark as a hardware layer (announced).

⚡ Quick-Select: Self-Hosted & Privacy Layer

Criterion	Recommended Tool	Why
Fastest start	ZeroClaw	`curl \| bash` in 2 minutes, single binary, no Docker needed
Highest privacy control	Ontheia	GDPR by design, AGPL-3.0, MCP-native, Docker Compose in 10 minutes
Best overall package	NanoClaw	Container isolation, native WhatsApp, `git clone` + `claude` in 5 minutes

Till Freitag recommendation: Ontheia or NanoClaw as runtime + Privacy Router as routing layer. Once available: RTX Spark as hardware layer. That gives you a full local-AI-first stack without hyperscaler dependency.

Enterprise Gateway Layer – When Enterprises Need a Gateway Today

Microsoft Scout is announced but not yet available. If you need a productive enterprise gateway in front of your agents in mid-2026 – with central auth, quotas, audit logs, cost tracking, and model routing – you reach for the following building blocks. They are all live and deployable today and can be migrated later, once Scout goes GA.

Tool	Purpose	Setup Effort	Privacy / Hosting	Typical Workflows
LiteLLM Proxy	OpenAI-compatible multi-provider gateway (100+ LLMs)	Docker / `pip install litellm` (~10 min)	Self-hosted, EU hosting possible	Central API keys, per-team quotas, spend tracking, fallback routing
Portkey AI Gateway	Governance layer with guardrails & observability	Docker / cloud (~15 min)	Self-hosted (OSS) or EU cloud	Prompt versioning, PII redaction, caching, A/B tests
Cloudflare AI Gateway	Managed edge gateway with caching & analytics	DNS entry (~5 min)	Managed (Cloudflare edge, EU PoPs)	Rate limiting, logs, cost caps, multi-provider failover
Kong AI Gateway	Classic API gateway with AI plugins	Helm / Docker (~30 min)	Self-hosted or Kong Konnect (EU)	mTLS, OAuth/OIDC, audit trails, plugin ecosystem
AWS Strands / Bedrock AgentCore	AWS-native agent & gateway backend	AWS CLI + IAM (~30 min)	AWS cloud (Frankfurt region)	IAM-based skill grants, CloudTrail audit, Bedrock models
Self-hosted OpenClaw + Privacy Router	DIY enterprise gateway with full control	Docker Compose + Ollama (~30 min)	On-premise, no data leakage	Sensitivity-aware model routing, custom auth, guide

An enterprise gateway has three jobs: control access (auth, quotas), make cost & risk visible (logs, spend tracking, PII redaction), and route models (provider failover, cost/privacy-aware routing). Scout will deliver these integrated later – today, you compose them from the building blocks above.

Typical use cases:

Central API keys & per-team spend tracking: One key per provider, quotas and budgets per department. → LiteLLM Proxy as multi-provider front door.
PII redaction & prompt governance: Sensitive data must not leave the prompt, every change versioned. → Portkey AI Gateway (with guardrails) + LiteLLM behind it.
Edge gateway with caching for high volumes: Marketing use cases with identical prompts, cost caps. → Cloudflare AI Gateway in front of LiteLLM.
Enterprise auth & mTLS for regulated sectors: OAuth/OIDC, audit trails, no vendor lock-in. → Kong AI Gateway (open source or Konnect EU).
AWS-only stack with a compliance story: Bedrock + IAM + CloudTrail as a closed loop. → AWS Strands / Bedrock AgentCore.
Maximum GDPR strictness, no hyperscalers: Self-hosting + Privacy Router + local models. → self-hosting guide + Privacy Router.

⚡ Quick-Select: Enterprise Gateway Layer (available today)

Criterion	Recommended Tool	Why
Fastest start	Cloudflare AI Gateway	DNS entry in 5 minutes, instant logs & cost caps
Highest privacy control	Self-hosted OpenClaw + Privacy Router	Fully on-premise, sensitivity-aware model routing, guide
Best overall package	LiteLLM Proxy (+ optional Portkey)	OpenAI-compatible, 100+ providers, quotas, spend tracking, Docker in 10 minutes

Till Freitag recommendation: LiteLLM Proxy as multi-provider front door + Portkey as governance layer + Privacy Router for GDPR-critical paths. AWS-only shops take Strands / Bedrock AgentCore. Once Microsoft Scout is GA, this stack can be migrated with manageable effort – skills and MCP configs stay the same.

The Decision Matrix

You need...	Choose...
Maximum security, WhatsApp	NanoClaw
To understand how agents work	Nanobot / Devika
An assistant that learns over time	memU
Cloud agent without installation	Moltworker
Edge deployment, minimal resources	NullClaw
Rust performance, strong community	ZeroClaw
Complete agent operating system	OpenFang
Multiple specialized agents	SuperAGI / AG2
Agents as state machines (production)	LangGraph
Flexible LLM experimentation	Anything LLM
Coding agent with sandbox (open source)	OpenHands
Coding in the terminal (CLI-first)	Aider
Coding agent for SWE-bench performance	SWE-agent
Open-source Copilot in the IDE	Continue.dev
Coding support (premium)	Claude Code
AWS-native agent backend	AWS Strands
MCP-native & GDPR by Design self-hosted	Ontheia
Lean, fast-iterating coding harness	dreb
Enterprise gateway with M365 integration	Microsoft Scout
Local-AI-first on your own hardware (announced)	OpenClaw/NanoClaw + RTX Spark
Everything at once (with risk)	OpenClaw

Our Recommendation

For most of our clients, we recommend a combination:

NanoClaw or Microsoft Scout (once available) as personal or enterprise agent
Claude Code for software development
monday.com + Make for structured business automation
Optional (announced): NVIDIA RTX Spark as a hardware layer for local-AI-first

Why? Because no single AI agent solves all problems. The future belongs to the orchestrated interplay of specialized tools – not the one agent that does everything.