Pacific Slate

What it is in one line

A self-hosted, multi-agent personal assistant written by one MBA student for himself: models (rented) and data (his own) are strictly separated, and all data stays on his own server.

Who built it

Builder field says badwx (HN handle). The site self-describes: a current MBA student, married with two children, no formal software background. The development method is "AI-native" — he designed the architecture himself (components, connections, budgets) and let coding AI agents write most of the code. The system has run daily since early 2026 through August 2026.

Read: this is not a product, it is a "single-user system architecture reference." Its value is not how many people use it, but that it proves a non-programmer can assemble a daily-use private AI system out of self-hosting, multiple models, and cost control.

What it actually does

• Autonomous information stream → pulls from email, calendar, GitHub, HN, Reddit, arXiv, RSS, weather, markets, earthquake data, and more on a schedule; scores items by relevance, importance, and novelty; produces briefings
• Multi-agent collaboration → 1 operator + 7 expert agents (coder, researcher, analyst, productivity, reviewer, evaluator), built on Google's ADK, each with least-privilege permissions
• Model-agnostic → capabilities exposed through an MCP gateway and OpenAI-compatible endpoints; swapping models is one config change and touches neither memory, data, tools, nor routing
• Four-layer memory → knowledge corpus, continuity graph, semantic memory, and hierarchical context; remembers preferences and past decisions
• Cost routing → a deterministic scorer (no model involved) rates prompt complexity 1–5 and, combined with budget usage, picks the model tier; there is a monthly budget cap
• Reliability engineering → health tracking and cooldowns for rate limits, a first-token timeout watchdog, a degradation chain (primary model fails → backup takes over → explicit failure message), and no crashes
• Traceability + data sovereignty → every answer carries source citations; all data lives on his server, exportable and deletable, not used to train models; external requests go through zero-data-retention endpoints

Operating shape: Next.js canvas frontend + Starlette backend + ~24 MCP tool servers behind a gateway (BM25 search finds the right tool instead of stuffing all schemas into context); answers render as movable cards, each labeled with model, cost, and latency.

What old behavior it replaces

For the author personally, three kinds of old actions:

Explaining yourself over and over to different AI programs. Switching models or tools meant re-communicating context; a unified system that self-updates, remembers preferences, and keeps data private ends that.

Long chat threads. One task, one conversation, scrolling forever. Now it is movable cards on a canvas, each labeled with which model did it, what it cost, and how long it took.

Dependence on a single model. Switching vendors used to be a migration; now it is a one-line config change.

For the general reader, it replaces "subscribing to a pile of AI services" with an alternative: rebuild a private version for $50–80/month with all data in your own hands.

Business model

None. Not a product — a personal system plus an open-source blueprint (the site says "almost all of it is open source"). The author's reconstruction estimate: ~$20/month coding assistant, ~$20/month small server, $10–40/month metered model calls, about $50–80/month total.

Read: this should not be evaluated as a business model. The real question is whether the architectural blueprint can be reused by others — if it can, it becomes a valuable reference for the self-hosting AI community.

Hard numbers

• Running totals: 13,053 runs, 26,515 model calls, 9 models used
• Performance: median model-call latency 3.1s, p90 latency 14s
• Knowledge base: ~14,000 documents / 196,000 text passages
• Users: 1 (the author); explicitly not aimed at enterprise or multi-tenant
• HN 5 points / 0 comments (2026-08-11) — essentially undiscussed
• Team: 1 person, with AI coding agents as development partners

Four-way read

The call

This is a "single-user architecture reference," not a product, and the data says Unproven. HN 5 points, zero comments, no users, no commercial intent — but it should not be graded as a product. It should be graded as a blueprint.

Three designs worth stealing: First, deterministic cost routing — a rule (complexity score 1–5 × budget usage), not a model, decides which model tier to use, avoiding the self-contradiction of "let the LLM decide how much money to spend." Second, model-agnostic as a first principle — swapping models touches neither memory, data, tools, nor routing; "no vendor lock-in" is architecture fact, not a slogan. Third, tools behind a gateway with BM25 lookup — 24 MCP servers never enter context; retrieval finds the right tool on demand. Same idea as mcptoon, different implementation.

The biggest question mark: the reliability numbers (3.1s median latency, 13,053 runs) and the architecture diagram are self-reported by a non-programmer writing with AI agents; no third-party verification and no direct repo link visible on the site. So it is a "reference design worth reading," but every number must be discounted.

What to watch next

① Whether the code actually goes open source (downloadable, reproducible) — a blueprint's value depends on being reusable ② Whether monthly cost stays around $50–80 — the cost claim is its most falsifiable assertion ③ Whether anyone builds a team version or an open-source project from it — reuse is the evidence that this architecture actually works

What you can take from it

Product logic: for deciding "which model to use," do not let the LLM decide — route with a deterministic scorer (complexity score × budget level): controllable, explainable, auditable. Any multi-model product team can copy this routing logic directly.

Pricing structure: no business model, but the broken-down cost items (~$20 coding assistant + ~$20 server + $10–40 model calls) are themselves a pricing reference — the selling point of self-hosted options is "total cost below the sum of subscriptions."

Verdict

Unproven, but worth filing. Not a product — a complete design document for "one person plus AI agents builds a private AI system," with three genuinely transferable designs: cost routing, model-agnosticism, and tool retrieval. All data is self-reported and no direct code link is visible, so the grade is Unproven by the data — but this blueprint deserves a read by anyone planning a self-hosted setup. In three months, check whether it is actually open-sourced and reused.