🛠 ECC (Agent Harness OS) MIT
https://github.com/affaan-m/ECC ↗67 agents 281 skills 94 commands hooks + memory AgentShield security
plan→test→implement→review→verify→remember→improve
install
npx ecc-install --profile minimal --target claudeWorks across Claude Code · Codex · Cursor · OpenCode · Gemini · Zed · Copilot · Qwen. 281 reusable skills also listed on Agent skills.
Match the tool to your team and use case rather than chasing star counts. Practical decision paths for SMB/mid-market as of June 2026:
CODE FRAMEWORKS (you have engineers).
- Python, want fine-grained control over multi-step/stateful workflows: LangGraph. Most production-proven, but has a learning curve.
- Python, want type-safe, clean, maintainable code with validated outputs: Pydantic AI. Best developer experience for structured agents.
- Python, want a 'team of role-playing agents' framing fast: CrewAI. Easiest mental model for multi-agent.
- Python, want minimal and hackable / open-weights and local: Hugging Face smolagents (pair with a sandbox).
- TypeScript/Node/Next.js shop: Mastra (full framework) or OpenAI Agents SDK (TS) — don't force Python.
- Already standardized on a model provider: OpenAI Agents SDK (OpenAI), Anthropic Claude Agent SDK (Claude, also best for code/computer-use agents), Google ADK (Gemini/GCP, also Java/Go), Strands (AWS/Bedrock).
- .NET / C# / Azure shop: Microsoft Semantic Kernel — but target its successor, Microsoft Agent Framework (MAF), for new builds.
- Document/RAG-heavy product: LlamaIndex Workflows (best retrieval, parsing, extraction stack).
- Long-term memory / personalization is the core differentiator: Letta.
LOW-CODE / VISUAL (limited or no deep engineering capacity — often the right call for SMBs).
- All-in-one production LLM app platform (RAG + agents + deploy): Dify. Best 'ship without a big eng team' option; note its license has minor restrictions.
- Automation-first, connect AI into existing SaaS (CRM, email, Slack): n8n. Fair-code license (source-available, commercial limits) — check terms.
- Fully permissive, self-owned visual builder: Flowise (Apache-2.0, no usage strings).
CODING AGENTS (automate software development).
- Per-developer AI pair programmer in the IDE with reviewable diffs: Cline (most installs) or Roo Code (Cline fork with extra modes/orchestration). For a local terminal/desktop alternative: Goose (Block, Rust, MCP-extensible).
- Terminal-native coding agent on a specific model: Claude Code CLI (Anthropic — strongest on agentic SWE-bench), OpenAI Codex CLI (open-source, ChatGPT subscription or API).
- Diff-driven repo editing from the terminal/IDE with explicit edit-format control: Aider — pioneered the 'edit format' paradigm and powers the Aider Polyglot leaderboard.
- Open VS Code/JetBrains assistant with first-class custom-model support: Continue.dev (works with any local or remote LLM).
- Autonomous 'do the whole ticket' agent you self-host: OpenHands. Heavier to run but powerful.
- Transparent, minimal, research-grade issue-fixer: SWE-agent / mini-swe-agent (for technically strong teams).
AVOID FOR NEW PROJECTS: Microsoft AutoGen is in maintenance mode (frozen async stack, breaking infra changes flagged for late 2026). Choose Microsoft Agent Framework if you're in the Microsoft camp, or AG2 only if you're already committed to that community fork.
GENERAL SMB GUIDANCE: Favor permissive licenses (MIT/Apache-2.0) and self-hostable cores to avoid lock-in and per-seat costs; nearly all the code frameworks here are free OSS where you pay only model-token costs. Don't adopt a heavy graph framework for a single simple chatbot — start with a lightweight SDK (OpenAI Agents SDK, Pydantic AI, Strands) or a low-code platform (Dify/Flowise) and graduate to LangGraph/CrewAI only when workflows genuinely become multi-step. Be deliberate about model-provider gravity: provider-native SDKs are the smoothest path when you've already picked a model, while LangGraph, CrewAI, Pydantic AI, Mastra and smolagents keep you model-agnostic. Budget for safety/observability (sandboxing for code-execution agents, human-in-the-loop review for autonomous coding agents) regardless of framework.
| Harness | Type | License | Stars | Best for | SMB fit |
|---|---|---|---|---|---|
affaan-m · ecc.tools · Shell + TypeScript + Python | Agent harness OS | MIT | 235.1k | Turns any coding agent (Claude Code, Codex, Cursor, OpenCode, Gemini, Zed, Copilot, Qwen) into a coordinated engineering system — 67 agents, 281 skills, 94 commands, hooks + memory, and AgentShield security scanning. Workflow: plan → test → implement → review → verify → remember → improve. | Free MIT install (npx ecc-install); great for standardizing how a small team’s coding agents plan, test and review. Pro (hosted GitHub App) covers private-repo automation. |
n8n GmbH · TypeScript/Node.js (used via visual UI) | low-code | Sustainable Use License (fair-code; source-available, commercial restrictions) | 194.0k | Visual workflow automation with 400+ integrations now augmented by a first-class AI Agent node, so you can chain LLM calls into real business automation (CRM, email, Slack, DBs). Best for connecting agents to the rest of your SaaS stack. | Excellent SMB fit. Self-host free or use n8n Cloud; integrates AI into existing operational workflows that SMBs already care about. Ideal when automation + light agentic AI beats building a bespoke agent app. Check license terms for commercial redistribution. |
LangGenius · Python (backend), TypeScript (frontend) — used via visual UI | low-code | Dify Open Source License (Apache-2.0 with minor restrictions) | 129.0k | All-in-one visual platform combining agentic workflow builder, RAG pipelines, model management, prompt engineering, observability and one-click app deployment. Best for shipping production LLM apps fast with minimal code. | Excellent SMB fit. Lets non-deep-engineering teams build RAG chatbots, internal assistants and workflows visually, self-host for data control, or use Dify Cloud. One of the best 'get to production without a big eng team' options. |
Anthropic · TypeScript (terminal CLI; VS Code & JetBrains integrations) | coding-agent | Source-available (commercial — usage governed by Claude usage policies) | 115.0k | Frontier coding-agent behaviour out of the box on Claude models — terminal-native, codebase-aware, runs commands, edits files with reviewable diffs, supports MCP tools, GitHub @claude integration, plan-mode, subagents and hooks. Strongest published SWE-bench Verified/Pro and Terminal-Bench 2.1 scores in the open ecosystem. | Excellent SMB fit for teams already on Claude — install, point at a repo, ship. Less ideal for shops that need provider portability or fully OSS tooling. Pairs with the Claude Agent SDK when you want to embed the same engine into your own product. |
ByteDance · Python + TypeScript | SuperAgent harness | MIT | 73.9k | Long-horizon autonomous work (research, coding, content). LangGraph/LangChain-based multi-agent runtime with a filesystem sandbox, persistent cross-session memory, on-demand skill modules, MCP tools, and native IM integration (Slack/Telegram/Lark/WeChat). | Strong for self-hosted research/ops automation; free and MIT, but expect real infra to run it well. |
All Hands AI · Python | coding-agent | MIT | 70.0k | Autonomous AI software engineer that executes complex coding tasks end-to-end (reads code, edits files, runs commands, browses) and collaborates with developers. Model-agnostic. Best open-source 'do the whole ticket' coding agent. | Good fit for SMB software teams wanting to automate GitHub-issue-style dev work without per-seat SaaS fees. MIT-licensed and self-hostable, or use the OpenHands cloud. Requires engineering comfort to operate safely. |
Cline · TypeScript (VS Code/JetBrains extension, SDK, CLI) | coding-agent | Apache-2.0 | 62.0k | In-IDE autonomous coding agent that reads your codebase, creates/edits files with reviewable diffs and checkpoints, runs terminal commands and drives a real browser, asking approval at each step. Best for human-supervised pair-programming inside VS Code/JetBrains. | Excellent SMB fit for developer productivity. Free, open-source (Apache-2.0), pay only for model tokens, no per-seat SaaS lock-in. Easiest 'give every developer an AI coding agent' option for small teams. |
Microsoft (AutoGen) / ag2ai community (AG2 fork) · Python, .NET | multi-agent | MIT (AutoGen) / Apache-2.0 (AG2) | 55.0k | Conversational/event-driven multi-agent research and prototyping; AG2 preserves the classic GroupChat style for teams already invested in it. Strong for experimentation and agent-to-agent conversation patterns. | Weak for new SMB projects due to maintenance-mode status and fork confusion. Only pick if you already run AutoGen; otherwise choose Microsoft Agent Framework (if .NET/Azure) or CrewAI/LangGraph. |
CrewAI Inc. · Python | multi-agent | MIT | 52.0k | Role-based multi-agent 'crews' (e.g. researcher + writer + reviewer) with built-in process types (sequential/hierarchical). Fastest path to a believable multi-agent demo with readable, role-oriented code. | Strong SMB fit. Approachable Python API, fast to prototype, MIT-licensed and free to self-host. Enterprise AMP add-on is optional. Good when a small team wants 'team of agents' framing without low-level plumbing. |
FlowiseAI (acquired by Workday) · TypeScript/Node.js (used via visual UI) | low-code | Apache-2.0 | 51.0k | Drag-and-drop builder with three modes — Agentflow (multi-agent), Chatflow (single-agent RAG) and classic LLM chains. Apache-2.0 with no usage restrictions, so you fully own your self-hosted deployment. Best free, unrestricted visual builder. | Strong SMB fit. Truly permissive Apache-2.0 license (no commercial strings unlike n8n), self-hostable, good for prototyping RAG bots and simple multi-agent flows. Best when you want a no-restrictions visual tool you fully control. |
LlamaIndex (run-llama) · Python, TypeScript | graph / event-driven | MIT | 50.0k | Event-driven orchestration (steps emit/consume typed events; branch, loop, parallelize, persist, recover) layered on LlamaIndex's best-in-class RAG, document parsing (LlamaParse) and structured extraction. Best when your agent is fundamentally about documents and retrieval. | Strong SMB fit for document-heavy use cases (knowledge bases, contract/PDF processing, RAG chatbots). Free OSS core, with optional managed LlamaCloud for parsing at scale. Less ideal for non-RAG agent work. |
Block (Square / Cash App) · Rust | coding-agent | Apache-2.0 | 47.0k | Local-first, extensible general-purpose agent (desktop app, CLI, API) that installs, executes, edits and tests code with any LLM (15+ providers) and connects to 70+ MCP extensions. Best for a fast, native, vendor-neutral local agent. | Strong SMB fit for dev teams wanting a free, local, MCP-extensible agent with no per-seat cost and no lock-in (Apache-2.0, LF-governed). Runs on macOS/Linux/Windows. Good Cline alternative for terminal/desktop-centric workflows. |
OpenAI · Rust (CLI; VS Code & Cursor extensions) | coding-agent | Apache-2.0 | 40.0k | Lightweight, OSS terminal coding agent from OpenAI — runs in your shell, reads and edits files, executes commands in a configurable sandbox, supports MCP tools and screenshot/image inputs. Best fit for teams already on ChatGPT or OpenAI API who want a first-party coding agent without vendor lock-in on the harness. | Strong SMB fit, especially for ChatGPT-subscribed teams (no extra API spend on the included quota). Apache-2.0 and self-hostable. A clean default for a small team that wants OpenAI quality with an OSS, terminal-first harness. |
Aider-AI · Python (CLI; works alongside any editor) | coding-agent | Apache-2.0 | 38.0k | Terminal-native pair-programming with strict, model-agnostic edit formats (diff / whole-file / architect-editor). Maps your repo, edits multiple files, runs your tests/linters, and auto-commits each change with a sensible message. Best for teams that want full control over how the model touches code. | Excellent SMB fit. Free, Apache-2.0, pay only for model tokens, works with every major provider (Claude, GPT, Gemini, DeepSeek, Qwen, local). The benchmark-grade reference for 'does this model actually edit code reliably' — pairs naturally with the Aider Polyglot leaderboard. |
LangChain · Python, JavaScript/TypeScript | graph | MIT | 34.0k | Stateful, long-running, controllable agent workflows where you need explicit control over branching, loops, checkpointing, human-in-the-loop and durable execution. The default 'serious' framework when you outgrow simple chains. | Good for SMBs with at least one engineer who will own the agent. Open-source core is free and self-hostable; managed Platform costs money but is optional. Overkill for a single simple chatbot, ideal once workflows get multi-step. |
Continue · TypeScript (VS Code & JetBrains extensions; agent CLI) | coding-agent | Apache-2.0 | 30.0k | Customizable open IDE assistant that connects to ANY model — frontier APIs, local Ollama/LM Studio, your own endpoints — with first-class config for tools, MCP servers, context providers and per-team rules. Best when model choice and rule customization are non-negotiable. | Excellent SMB fit for teams that want a free, vendor-neutral IDE assistant they fully control. Apache-2.0, BYO keys (or local model), good fit for compliance-sensitive shops that need to keep code on-prem or on a specific model provider. |
Microsoft · C#/.NET, Python, Java | SDK | MIT | 28.0k | Enterprise-grade orchestration for .NET (and Python/Java) shops: plugins/skills, planners, session state, type safety, middleware, telemetry, plus A2A protocol support. Best choice for Microsoft/Azure-centric organizations, especially C#. | Good SMB fit specifically for .NET/Azure shops (rare among these frameworks in offering first-class C#). MIT-licensed and free. For new projects, start on Microsoft Agent Framework rather than legacy Semantic Kernel directly. |
OpenAI · Python, TypeScript | SDK | MIT | 26.0k | Lightweight multi-agent workflows with handoffs, guardrails and built-in tracing. Cleanest option if you're already standardized on OpenAI models, though it's provider-agnostic and supports 100+ LLMs via Chat Completions. | Excellent SMB fit if OpenAI is already your model provider. Minimal, readable API, free and open-source, fast onboarding. Provider-agnostic enough to avoid hard lock-in. Great default for a small team building 1-3 agents. |
Hugging Face · Python | SDK / code-agent | Apache-2.0 | 26.0k | Minimal, ~1,000-line library for agents that 'think in code' (CodeAgent writes/executes Python) plus a classic ToolCallingAgent. Model-agnostic (any HF Hub, OpenAI, Anthropic, Ollama, local via LiteLLM). Best for simplicity and code-first reasoning. | Strong SMB fit for technical teams wanting something tiny and hackable with no lock-in. Pair with a sandbox provider for safety. Great for local/open-weights deployments where avoiding cloud LLM bills matters. |
Mastra (ex-Gatsby team) · TypeScript | SDK / graph | Apache-2.0 (Elastic License v2 for some components) | 22.0k | Modern TypeScript-native framework for agents and AI apps: agents, workflows (graph), RAG, memory, evals and observability in one cohesive package. Best agent framework for JavaScript/TypeScript and Node/Next.js teams. | Excellent SMB fit for web/full-stack teams already in TypeScript who don't want to context-switch to Python. Free OSS core, integrates with Vercel/Next.js. The natural pick alongside OpenAI Agents SDK (TS) for JS shops. |
Roo Code Inc. · TypeScript (VS Code/JetBrains extension) | coding-agent | Apache-2.0 | 22.0k | In-IDE coding agent with built-in role modes (Code / Architect / Ask / Debug / Orchestrator) and a custom-mode system that lets you script specialized agents per file pattern. Best when you want Cline's reviewable-diff UX plus more out-of-the-box workflow control. | Strong SMB fit. Free, Apache-2.0, pay only for model tokens. The right pick over vanilla Cline when your team wants orchestrated multi-mode workflows or wants to ship internal 'specialist' coding agents (test-writer, migration agent, reviewer). |
Princeton & Stanford (academic) · Python | coding-agent | MIT | 20.0k | Taking a GitHub issue and automatically fixing it with your chosen LM via a well-designed agent-computer interface; also used for offensive cybersecurity and competitive coding. mini-swe-agent is a radically simple ~100-line variant scoring >74% on SWE-bench Verified. | Niche SMB fit. Best for technically strong teams or researchers who want a transparent, hackable, MIT-licensed coding agent and bench-style automation. mini-swe-agent is a great minimal starting point; less suited to non-engineers. |
Pydantic (Samuel Colvin's team) · Python | SDK | MIT | 17.0k | Type-safe agents with validated, structured outputs from the ground up. Excellent IDE/type-checker ergonomics, moving errors from runtime to write-time. Model-agnostic across nearly every major provider. Ideal when output correctness and DX matter. | Excellent SMB fit for Python shops that value clean, typed, maintainable code. Low ceremony, free OSS, broad model support avoids lock-in. Great default for structured-data and tool-using agents. |
Letta (UC Berkeley Sky Lab founders) · Python | SDK / stateful-agent platform | Apache-2.0 | 17.0k | Stateful agents with advanced, tiered long-term memory (core/recall/archival) that learn and self-improve over time. Best when persistent identity and memory across sessions are the core requirement (personal assistants, long-lived support agents). | Niche but valuable SMB fit when memory/personalization is the differentiator (e.g. a customer assistant that remembers each user). Apache-2.0 and self-hostable. Overkill if you only need stateless task agents. |
Google · Python, Java, Go (plus TypeScript/Kotlin) | graph / SDK | Apache-2.0 | 16.0k | Code-first, multi-language agent building with a graph execution engine (routing, fan-out/fan-in, loops, retries, HITL, nested workflows) and structured A2A agent-to-agent delegation. Best if you're on Google Cloud / Gemini or need Java/Go. | Good SMB fit if you're a Google Cloud / Gemini shop or need non-Python (Java/Go) agents. Apache-2.0 and free to self-host. Less compelling for teams with no Google footprint. |
Anthropic · Python, TypeScript | coding-agent / SDK | MIT | 8.0k | Building agents that act like Claude Code: built-in file read/edit, command execution, codebase awareness, GitHub @claude integration, structured JSON outputs, programmatic subagents. Best-in-class for coding and computer-use style agents on Claude models. | Very good SMB fit for teams using Claude, especially for dev-tooling and code-automation agents. Batteries-included tools mean less plumbing. Watch the Agent SDK credit billing model for cost predictability. |
AWS · Python, TypeScript | SDK | Apache-2.0 | 6.1k | Model-driven agents in a few lines of code — give a prompt plus tools and let the LLM plan and chain tool calls autonomously, with strong built-in observability. Best for AWS/Bedrock-centric teams wanting a clean, any-model, any-cloud harness. | Good SMB fit, especially on AWS/Bedrock, but useful any-cloud. Apache-2.0 and free; the model-driven, low-boilerplate design is approachable for small teams. Verify maturity for mission-critical use given its newer status. |
Agentica (UC Berkeley) + Together AI · Python (open-weights agent; Qwen3-32B base + rLLM/R2E-Gym scaffold) | coding-agent | Open-weights (permissive research release) | — | A fully open, RL-trained coding agent — weights, training recipe and agent scaffold released — reaching ~59% on SWE-bench Verified on a 32B open base. Best when you want a self-hostable, no-API-cost SWE agent and full transparency into how it was trained, rather than a closed frontier agent. | Niche but valuable for technically strong, cost- or sovereignty-sensitive teams that want an open coding agent they can host and fine-tune. Note: DeepSWE is an AGENT/model (its eval set is SWE-bench Verified), not a benchmark — model cards that list it as a row are citing the agent. |
At a glance
ECC (Agent Harness OS)
Agent harness OS⚠ Broad surface (67 agents / 281 skills) is a lot to adopt at once; strongest on Claude Code, adapter depth varies by harness.
DeerFlow
SuperAgent harness⚠ Heavier runtime (Docker sandbox, memory layer); "SuperAgent" scope is more than simple workflows need.
LangGraph
graph⚠ Steeper learning curve; graph/state mental model is verbose for simple tasks; tied to the broader LangChain ecosystem which some teams find heavyweight; best features (Platform, persistence, observability) push you toward paid LangSmith/LangGraph Platform.
CrewAI
multi-agent⚠ Higher-level abstraction means less fine-grained control than LangGraph; can be hard to debug when agents loop or stall; orchestration is opinionated; production reliability historically weaker than the graph frameworks.
Microsoft AutoGen / AG2
multi-agent⚠ AutoGen is no longer Microsoft's recommended path for new production builds (steer to Microsoft Agent Framework); frozen async stack with breaking infra changes flagged for late 2026. Fragmentation across AutoGen v0.7, MAF, and AG2 creates real adoption confusion. No native concept of a 'process' without extra code.
OpenAI Agents SDK
SDK⚠ Thinner on durable state/persistence than LangGraph; tracing/observability nudges you toward the OpenAI platform; ecosystem younger than LangChain's; advanced features (sandbox, long-horizon) tie best into OpenAI infra.