The resource shelf

The official site: docs, SDKs, and quickstarts for the agent↔tools standard. Start with the server quickstart.

The current spec: the stateless revision. Read its changelog first if you maintain a server built on the older protocol.

The official extensions that sit outside the core: MCP Apps (interactive UI), Tasks (long-running calls), Skills, and authorization profiles.

Docs for the mcp 2.x package, including the migration guide from 1.x (FastMCP is now MCPServer).

Docs for the split v2 packages, @modelcontextprotocol/server and @modelcontextprotocol/client.

The spec (v1.0) for cross-vendor agent delegation: task lifecycle, discovery, and signed Agent Cards. Hosted by the Agentic AI Foundation.

The normative v1.0 text, including the Agent Card fields and the JSON-RPC, gRPC and HTTP+JSON bindings.

Event-based protocol (1.0) for streaming agent state into frontends: reasoning steps, tool results, interactive components.

Source, SDKs, and release notes for AG-UI, including the 0.x to 1.0 migration notes.

The Linux Foundation home for open agent standards: MCP, A2A, AGENTS.md, goose and agentgateway.

The open convention for a repo-level instructions file that most coding agents read.

The open spec for Agent Skills: folders of instructions and scripts that agents load on demand, portable across tools.

The vendor-neutral telemetry schema for model calls, agent spans, and MCP calls, now in its own repo. Still marked Development, so pin the version you adopt.

The contract language behind every tool definition and structured output you will ever write.

Claude API reference, models, tool use, and the Agent SDK. (The old docs.anthropic.com address redirects here.)

Install, skills, hooks, subagents, plugins, and MCP setup for Claude Code.

API docs for GPT models; the OpenAI-compatible protocol is the de-facto local-serving interface too.

Agents, handoffs, guardrails, sessions, sandboxed agents, and MCP support, in Python (JS version too).

Gemini API docs, model list, and agent tooling.

Open-weights family: downloads, licenses, and docs (now served from Meta's developer site).

Strong open-weights and hosted models out of Europe; good tool-calling small models.

Alibaba's open-weights family — consistently strong at agentic/tool-use benchmarks per size class.

CPU/GPU inference anywhere, first-class ARM64. llama-server speaks the OpenAI and Anthropic APIs and can host stdio MCP servers itself (experimental).

The GPU-throughput serving standard for open-weights in production.

The developer-convenience layer for local models — one command to run.

Open-source LLM gateway: one API for 100+ providers, keys, budgets, fallbacks. Pin it with hashes; two 2026 PyPI releases were compromised.

Open-source LLM tracing, evals, and prompt management; self-hostable. Version 4 changed the data model and breaks older SDKs, so upgrade the SDK with the server.

Open-source tracing + evaluation, OTel-native.

Tracing and evaluation from the Weights & Biases stable.

Config-driven eval runner — a good step up when your hand-rolled harness grows.

Home of the OWASP Top 10 for LLM applications and related guidance on securing generative AI.

The peer-reviewed agent threat taxonomy (ASI01–ASI10, published December 2025): goal hijack, tool misuse, identity abuse and more. Map your threat model to it.

The US reference framework for AI risk — voluntary, widely cited in enterprise governance programs.

Regulation (EU) 2024/1689. Prohibitions applied from Feb 2025 and GPAI obligations from Aug 2025; most transparency duties (such as telling people they are dealing with an AI) applied from 2 Aug 2026. The Digital Omnibus, Regulation (EU) 2026/1744, moved the high-risk dates: 2 Dec 2027 for Annex III systems and 2 Aug 2028 for Annex I.

The April 2026 interagency guidance that replaced SR 11-7. It explicitly leaves generative and agentic AI out of scope, which is why agent governance needs its own controls.

The paper behind the think→act→observe loop every agent runs.

Models teaching themselves when to call tools — tool use becomes mainstream.

Agents improving via verbal self-feedback — the seed of self-correction loops.

The Smallville simulation — memory, reflection, and planning working together.

The benchmark that made coding agents measurable on real GitHub issues.