Framework board · hand-maintained
/
Last checked 16 Sep 2026
10 protocols & frameworks indexed
Protocols & frameworks · maintained by hand
AI Framework Leaderboard
What holds it together. The wire format everything is converging on, and the libraries worth reaching for when a plain loop stops being enough. There is no composite score here: the right framework depends on how much structure the job needs, so this board compares type, licence, languages and focus. Tap any column to re-sort.
⚙ How to read this board
Order is editorial. Open any row for its full spec sheet.
Not a score. Judgement calls on when each framework earns its place over a plain loop.
| # | Framework / protocol | Type | License | Languages | Focus | Best for |
|---|---|---|---|---|---|---|
| 1 | MCP 2026-07-28protocol open specCurrent spec. Stateless core, per-request capability negotiation, cacheable list results, and a formal extensions framework. |
Open specification | Open | Any (wire format) | Tool + context wire format | A standard tool interface across agents |
| 2 | MCP extensionsextensions opt-in, negotiatedTasks for long-running work with durable handles, Apps for inline UI, Skills over MCP for structured agent instructions. |
Opt-in, negotiated | Open | Any | Tasks, Apps, Skills over MCP | Long-running work, inline UI, structured instructions |
| 3 | LangGraphorchestration LangChainExplicit state machines. Worth it when the control flow is genuinely a graph, not a loop. |
State-machine library | Open source (LangChain) | Python, JS | Explicit graph control flow | When the control flow is really a graph |
| 4 | Pydantic AItyped PydanticTyped outputs and typed tools for people who already think in models. |
Library | Open source (Pydantic) | Python | Typed tools and outputs | Teams that already think in models |
| 5 | DSPyoptimizer Stanford NLPCompiles prompts against a metric instead of asking you to write them. |
Library | Open source (Stanford NLP) | Python | Compile prompts against a metric | Optimising instead of hand-writing prompts |
| 6 | Agent Zeroagentic open sourceGeneral-purpose agentic framework that treats the whole computer as its tool and writes its own code to finish a task. Nothing hard-coded; you shape it as you go. |
General-purpose | Open source | Python | Computer-as-tool, self-written code | Fully customisable autonomous agents |
| 7 | PaperclipAIautonomous open sourceAutonomous multi-agent framework for long-horizon goals — a planner, worker agents and a budget it can't overrun. Ambitious; keep the guardrails on. |
Multi-agent | Open source | Python | Long-horizon goals under a budget | Ambitious autonomy (with guardrails on) |
| 8 | OpenAI Agents SDKorchestration OpenAILightweight multi-agent orchestration: agents, handoffs, guardrails and sessions in a small, sturdy API. |
Library | Open source (OpenAI) | Python, JS | Agents, handoffs, guardrails, sessions | Lightweight multi-agent apps |
| 9 | Google ADKorchestration GoogleAgent Development Kit — a code-first Python/Java framework for multi-agent systems, model-agnostic and built to deploy. |
Framework | Open source (Google) | Python, Java | Code-first multi-agent, model-agnostic | Deployment-ready multi-agent systems |
| 10 | CrewAImulti-agent open sourceRole-based multi-agent crews; quick to stand up when the work splits cleanly into roles. |
Framework | Open source | Python | Role-based crews | Work that splits cleanly into roles |
Editorial index · hand-maintained, not a feed
Hand-maintained, not a feed. Checked 16 Sep 2026 against MCP specification. The MCP rows come from the specification; the framework rows are judgement, not vendor copy.