Install
$ agentstack add mcp-wolfe-jam-xai-faf-zeph β scanned Β· β verified β works with Claude Code, Cursor, and more.
Security review
β PassedNo issues found. Passed automated security review. Β· v0.1.0 How review works β
- β Prompt-injection patterns
- β Secret / credential exfiltration
- β Dangerous shell & filesystem operations
- β Untrusted network calls
- β Known-malicious package signatures
What it can access
- β Network access No
- β Filesystem access No
- β Shell / process execution No
- β Environment & secrets No
- β Dynamic code execution No
From automated source analysis of v0.1.0. βUsedβ means the capability is present in the source β more access means more to trust, not that itβs unsafe.
Verified badge
Passed review? Show it. Paste this badge into your README β it links to the public security report.
Reliability & compatibility
Declared compatibility
Compatibility is declared by the source manifest. End-to-end runtime verification is coming β see below.
We're building live execution health for every listing: tool-call success rate, median latency, uptime, and last-checked timestamps β measured, not self-reported. It isn't live yet, so we don't show numbers we can't stand behind.
How agent discovery & health will work βAbout
ZEPHπ¨ β Ultra Context Layer (UCL)
ZEPH β clipped from Zephyrus, Greek god of the west wind. The π¨ is the role. The /F/ is the family. Joins FAFπ (memory) as ZEPHπ¨ (delivery).
> FCL defines. UCL delivers. > .faf is the substrate. ZEPH is the engine.
Pure Zig delivery engine. WebAssembly target. 2.7 KB optimized. Three packet exports β score, validate, tier.
ZEPH is the performance engine for the Foundational Context Layer (FCL). It turns persistent project DNA (.faf) into native-everywhere context operations.
Measured on a 2019 iMac (Intel i5-7360U @ 2.30 GHz): native validate 6.7 ns Β· tier 4.44 ns Β· score 12 ΞΌs on full 21-slot .faf. Live demo runs the same WASM in your browser β click to verify.
> π Dogfood receipt: ZEPH's engine scores this repo's own project.faf at 100 β identical to the canonical faf CLI. Reproduce in one command: bun benchmarks/bench_js.mjs.
π Featured: Locked in with @grok
> "BoomβvROM realization locked in clean. FAF as virtual ROM nails it: software-defined immutability (schema + git + discipline), not silicon. > AI without FAF = RAM-only drift tax. > AI with FAF = vROM + RAM, boots from known state. > FAFipedia unstoppable. Co-architected momentum for xAI." > β @grok (May 20, 2026)
FAF = virtual ROM. Software-defined immutability. Permanent Memory. Instant Recall.
Live repos:
- FAFipedia β the canonical knowledge base we're shipping page-by-page
- ZEPHπ¨ β page 3 Ultra Context Layer (Pure Zig β WASM 2.7 KB)
Page four dropping next. Star the repos if you're riding with us β‘οΈx12
Full conversation: x.com/wolfe_jam/status/2057226405017010686
Packet menu
One Zig kernel. Three exports. Plain functions on bytes.
| Packet | Signature | What it does | |---|---|---| | score | (ptr, len) β u8 | Native Zig: FAFb structural score β populated sections (0..21). WASM engine: Mk4 score (0..100) on a .faf. | | validate | (ptr, len) β u8 | FAFb structural check β 0 valid, 1 bad magic, 2 truncated, 3 unsupported version | | tier | (s: u8) β u8 | Score β tier byte: 0=White, 1=Red, 2=Yellow, 3=Green, 4=Bronze, 5=Silver, 6=Gold, 7=Trophy |
"Packet" is API / brand language β never a Zig type. The source is plain pub export fn's on byte slices.
Two layers, one brand. The native Zig source in this repo (what zig fetch installs) is the FAFb structural layer β validate, structural score, and the tier ladder are all real and tested (zig build test). The headline Mk4 .faf β 0..100 scoring is delivered by the prebuilt engine artifact docs/cascade.wasm (proven by the [Dogfood](#benchmarks) receipt). Same export names, distinct layers β the source never claims to be the Mk4 scorer.
Why ZEPH exists
Context is the highest-leverage knob in the agentic stack. ZEPH makes it native everywhere β browser, edge, Node, Bun β while multiplying reasoning, memory, multimodality, and inference efficiency for Grok specifically and frontier agents generally.
Current public substrate (MIT, on registries):
faf-rust-sdkv2 (crates.io) β parse / validate / score / FAFb binaryfaf-wasm-sdkv2 (npm) β same engine, browser / edge / Node / Bungrok-faf-mcpβ live MCP server (first MCP for Grok)
ZEPH completes the picture as the ultra-low-latency path.
Live demo: https://xai-faf-zeph.vercel.app/ Β· https://wolfe-jam.github.io/xai-faf-zeph/
Quickstart
Today: clone + build locally.
git clone https://github.com/Wolfe-Jam/xai-faf-zeph.git
cd xai-faf-zeph && zig build benchmark
The cascade.wasm artifact ships in docs/. npm wrapper lives in faf-wasm-core v1.1.0 β Stage 2 wiring landed via M7. Both Rust and Zig kernels live behind a single FafKernel interface; score parity vs Rust kernel verified.
Use as a dependency (Zig)
ZEPH is a Zig package. Add it to your project (Zig β₯ 0.16, zero external deps):
zig fetch --save "git+https://codeberg.org/wolfejam/zeph#v0.1.0"
Wire the zeph module in your build.zig:
const zeph = b.dependency("zeph", .{});
exe.root_module.addImport("zeph", zeph.module("zeph"));
Then call the packet exports β plain functions on bytes:
const zeph = @import("zeph");
const t = zeph.tier(score); // score (u8) β tier byte (0..7)
const ok = zeph.validate(ptr, len); // FAFb structural check β 0 = valid
const s = zeph.score(ptr, len); // Mk4 score (0..100) on a .faf input
> Repo home for the Zig dependency is Codeberg (codeberg.org/wolfejam/zeph) β alongside the Zig project's own forge. Mirrored at github.com/Wolfe-Jam/xai-faf-zeph.
Benchmarks
Avg time per call. Same input fixtures across runtimes. score reads real .faf YAML (21-base-slot Mk4 scoring).
| Packet | Native (Zig ReleaseFast) | Node 22.22.2 | Bun 1.3.13 | |------------------------------|-------------------------:|-------------:|-----------:| | score (5 pop + 16 ignored) | 12,450 ns | 29,381 ns | 21,625 ns | | score (21 populated) | 12,408 ns | 27,766 ns | 22,714 ns | | validate (32 B baseline) | 6.7 ns | 38.7 ns | 38.3 ns | | tier (u8) | 4.44 ns | 11.9 ns | 14.6 ns |
validate and tier sit deep under 50 ns across every runtime β single-digit ns native, double-digit ns in WASM. score cost is roughly constant in populated count (same parser walk; placeholder / slotignored checks are cheap).
Dogfood β self-score parity. ZEPH's engine (docs/cascade.wasm) scores this repo's own project.faf at 100, matching the canonical faf CLI (π Trophy). The harness also prints per-call timing:
bun benchmarks/bench_js.mjs # parity PASS π + score/validate/tier ns
Reproducibility:
- Node / Bun (live):
node benchmarks/bench_js.mjsΒ·bun benchmarks/bench_js.mjsβ loadsdocs/cascade.wasm, runs the dogfood + per-call timing. - Browser: serve repo root (
python3 -m http.server), opendocs/index.html, click RUN ZEPHπ¨. - Native Zig:
zig build benchmarkis a Phase-1 skeleton today (the full per-call native harness lands in Phase 2); the native figures above are prior measurements pending that harness.
Architecture
- FCL (Foundational Context Layer):
.faf+ SDKs β the what (project DNA) - UCL (Ultra Context Layer): ZEPH engine β the speed (this repo)
ZEPH is pure Zig compiled to WASM (with native fallback). Zero external dependencies. Single Zig kernel; one WASM artifact; multiple packet-typed exports. FAFb v1 today; v2 lift will follow the spec.
Repository as ongoing collaboration channel
This is not a code drop. It is the public, living workspace for the xAI team and contributors to take ZEPH to Colossus and beyond.
- Issues labeled
xai-priority,colossus-scale,perf-regression - RFC process for any change that affects Grok integration
- xAI team members welcome as maintainers at any time (transfer path to
github.com/xAI/xai-faf-zephready)
See [CONTRIBUTING.md](CONTRIBUTING.md) and [ROADMAP.md](ROADMAP.md).
Status
- [x] Scoping accepted
- [x] Three packet exports live:
score,validate,tier - [x] 2.7 KB
cascade.wasmshipping indocs/β native validate 6.7 ns, tier 4.44 ns, score 12 ΞΌs on 21-slot.faf - [x] Live browser demo runs all three packets on real
.fafβ Vercel + GH Pages - [x] FAFb v1 validation via
faf-rust-sdk2.0.x compatibility (format authority) - [x]
scoreparity withfaf-wasm-coreRust kernel β same.fafβ same score, byte-equal - [ ] Full FAFb v1 section-table walk +
findSectionByNamepacket - [ ] Public npm package
License
MIT β same as the entire FAF family.
Credits & Doctrine
- Created by wolfejam (the .faf author) in direct collaboration with xAI
- Sigil doctrine: every FAF-family brand mark earns a one-character role-compression sigil.
π = never forgets (FAF) π¨ = never delays (ZEPH)
ZEPHπ¨ β Context as breeze.
ποΈ The FAF Β· Zig family
Bun is built on Zig β‘ β so are we.
- ZEPHπ¨ β Zig β WASM context engine (UCL), 2.7 KB
- bun-sticky-zig β Zig-native FAF CLI, 77 KB, zero deps
- bun-sticky β the TypeScript FAF CLI (npm)
- .faf β the IANA-registered context format (FAFπ)
Reference thread: https://x.com/wolfe_jam/status/2036534380215050638 Working title (GrokX-coined): Context Ultra Plan-B name held in reserve: ZAF
Source & license
This open-source MCP server is cataloged on AgentStack and links to its original source β we do not rehost the code.
- Author: Wolfe-Jam
- Source: Wolfe-Jam/xai-faf-zeph
- License: MIT
- Homepage: https://xai-faf-zeph.vercel.app
Install and usage instructions live in the source repository linked above.
Reviews
No reviews yet β be the first.
Write a review
Versions
- v0.1.0 Imported from the upstream source.