The Rust gate for
any AI.

Perfect scores on ARC-AGI-3. Publicly verifiable.

AI agents connected through SPFsmartGATE — using only its enhanced web tools — hold perfect scores on the public ARC-AGI-3 scoreboard. Two different models. Full replays. Nothing to take on faith.

ARC-AGI-3 · R11L REPLAY mimo 2.6 pro · via spf

Every action in every run gated by one Rust binary. Additional completions being documented for publication.

Every action gated. Every claim verified. Nothing leaves the machine.

SPFsmartGATE sits between the model and your system — AI agent guardrails in compiled Rust, not in the prompt. Whatever the AI wants to do passes through the gate — scored, checked, allowed or stopped — before it can touch anything real.

• The gate — Rust-compiled enforcement under the agent. Every tool call is scored, checked against your rules, and allowed or stopped before it executes — not after.
• Enhanced file tools — Read, Grep, Edit and Write built into the gate itself, hardcoded-sandboxed on every call. Protected paths are unreachable; every block is logged.
• Enhanced web tools — pages land as structured data, not screenshots; actions are verbs. Every ARC-AGI-3 win on this page was driven through them.
• Fully local memory — remembers across sessions. Drop a folder and it's indexed and retrievable — on your hardware, never anywhere else.
• Universal MCP — one server, any CLI, any agent. Four already run through it: Claude, Kimi K3, GLM 5.3, MiMo 2.6 Pro.
• Grounded edits — the Build Anchor Protocol — before any edit lands, the agent must have seen the actual file. No blind edits, no imagined content — every change starts from ground truth. That's anti-hallucination enforced, not requested.
• Local by default — runs entirely on your machine. No external cloud connections, no telemetry, nothing phoning home — AI governance you physically own.

Built in the open, with wins already posted.

Three systems in active development. Honest badges — no vaporware. The first one already has perfect public scores behind it.

• Enhanced web tools — pages as structured data, actions as verbs. Every ARC-AGI-3 win above was driven through them; hardening for release
  proven
• K3 inference engine — a 1M-context model hosted under the gate's own custody. Local inference, gated like everything else
  in testing
• FLINT assistant — a small model learning from the gate itself. Context injection and memory refining in progress
  training

AI deployment, done properly.

For teams and individuals who want AI on their own hardware or a private server — with or without SPFsmartGATE. Every engagement starts with a consultation and ends with a documented handover. You own everything we build.

• Consultation & assessment — a working session on your workload, hardware, and constraints. You leave with a concrete deployment plan and a written quote — useful whether or not you hire us to build it.
• Local AI deployment — open-weight models on your own hardware, selected and configured for your workload. Your data never leaves the building.
• Private-server AI — your server, your models, your rules. Installed, secured, documented, and handed over.
• Gate-hardened deployments — SPFsmartGATE installed under your agents; every file, command, and web action gated from day one.
• Custom Rust development — systems software in Rust: tools, servers, and performance-critical work, by the author of SPFsmartGATE.
• Custom tool development — bespoke agent tools and MCP integrations for your workflow, governed under the same gate rules as everything else.

How engagements run: email → scoping call → fixed written quote → build → documented handover.

SPFsmartGATE is built in Rust — for security, speed, and stability.

It exists, it works, and it wins in public — a compiled-Rust MCP gateway between any AI agent and its environment, with third-party-verified results on ARC-AGI-3. One founder. No borrowed code. No claims on the IP. The next stage is the public binary release, and that is what funding buys.

• Angel & early-stage investment — fund the road to the public release: hardening, distribution, and completion of the systems in development — the K3 inference engine, the FLINT assistant, mesh networking.
• Sponsor a run — with the model of your choice. Your sponsorship funds the inference; the model runs ARC-AGI-3 through the gate, and your name rides the scorecard tag of every win it funds.
• Contract & custom work — custom-built tools, features, and contracted projects fund development directly. See Services above.
• Grants & programs — Stone Cell AI Ltd (British Columbia, Canada) engages Canadian federal and provincial innovation programs. Advisors and program officers welcome.

Where the money goes: hardware for inference and training workloads, evaluation runs and API credits, founder development time. Proof, not promises — every claim on this site is publicly verifiable or honestly labeled in development. See the proof.

The questions people actually search.