mcp-better

What is 7/28?

7/28 is the speakable era name for the modern MCP release. Humans say 7/28. Machines negotiate the protocol version string 2026-07-28 on the wire.

That release is the largest update to MCP since launch: stateless remote servers, Discover lifecycle, list-cache stamps, and a path that doesn’t force every host through last year’s initialize cargo-cult. Building for 7/28 means your claims match what Discover and tools/list actually return — not a banner that says “modern” while the wire is still 2024-shaped.

Why this was built

When 7/28 landed, three paths showed up fast:

• A — Wrap legacy. Slap a date on an old server. Ship vibes.
• B — Wait. Let the ecosystem sort it out.
• C — Build real. Greenfield to the modern model, small enough to verify.

We took C. Lessons from AAIF work (Agentic AI Foundation — discipline, honest claims, evidence over slogans) kept the surface tiny: two tools, stamped list, Discover-first smoke, CI that fails if purity or smokes break.

Not a platform. Not a FAF product install. A textbook you can run so teams can learn 7/28 without inventing their own “good enough” server first.

What makes it BETTER?

The ladder is intentional:

NONE  →  GOOD  →  [BETTER]  →  BEST
                         ▲
                      this repo

BETTER means, concretely:

1. Protocol honesty — claim 7/28 / 2026-07-28 only for surfaces you implement and test.
2. Discover-compatible — prefer Discover (or Auto → 7/28), not only legacy initialize.
3. List cache stamps — tools/list returns positive ttlMs and cacheScope (static catalog → public). SDK defaults are often unstamped; we stamp on purpose.
4. Stable tool order — same process, same order across N list calls.
5. Transports in one era — stdio (default) and Streamable HTTP (--http) share the same 7/28 claim surface; HTTP is more road, not a new protocol era.
6. BETTER purity — no project.faf tax on this repo’s main. It’s a protocol textbook, not a FAF install funnel.

Who is it for?

• MCP builders learning 7/28 without forking a mega-server — copy the patterns (Discover, stamps, dual smoke).
• Host authors (Cursor, Claude Desktop, custom) who need a green stdio binary that negotiates modern protocol and returns an honest list.
• AAIF / ecosystem peers who want a small, Registry-listed cargo package with evidence bars (CI, smokes, SECURITY.md for HTTP).
• Teams teaching MCP — two tools (health, echo) is enough to demo lifecycle without drowning students in product scope.

If you need persistent project context, multi-format DNA, and agent memory — that’s BEST, not this crate. Start at faf.one/agents.

What’s in the box (0.1 → 0.2)

• v0.1 — 7/28 over stdio: Discover, ttlMs=60000 / cacheScope=public, tools health + echo, CI + stdio smoke.
• v0.2 — same era + Streamable HTTP (--http), Mcp-Method / Mcp-Name smoke, loopback Host guards, SECURITY.md threat posture (no auth/TLS — local demo only).
• Registry — package transport is stdio only by design (cargo install / host spawn). HTTP is opt-in local demo in the same binary.

What’s next

• Keep the era honest — more road inside 7/28, not rebranding every transport as a new “protocol.”
• Harder smokes where they earn their keep — host matrix notes stay docs-grade; CI stays the ship bar.
• BEST remains one hop up — persistent context is not smuggled into the textbook.
• Feedback welcome — issues and stars if the textbook earns them.

Try it

First cargo install compiles the ecosystem once (not 100+ of our tools — just Rust deps). One-time wait; then you’re done.

Release: v0.2.0 · crates: mcp-better · Registry: io.github.Wolfe-Jam/mcp-better

The numbers

• v0.2.0 — crates.io + MCP Registry + GitHub Release
• 2 tools — health, echo (stable order)
• 2 smokes — stdio Discover · Streamable HTTP (list + health + echo)
• Identity — io.github.Wolfe-Jam/mcp-better (cargo)
• Purity — no project.faf on main