Frao Advisor: Expert Counsel, on Demand

Frao Advisor is the MCP server that gives every Claude Code session a standing review board: seven expert personas, four tools, one model, wired directly into the development workflow. Over eleven days it has produced 459 pieces of advice for $2.06 total. This post covers what it is, how it runs, and what the first 459 records say about where review effort actually goes.

The review gap

Code that compiles is not code that is right. Between "it builds" and "it is correct" sits the entire value of review — and review is the first thing a small team drops when time is short. The reviewer is busy, the question feels awkward, the fix is "probably fine."

A second opinion on every change has always been the obvious fix and always been too expensive. A senior reviewer costs more than the change, reviews slowly, and sees only a fraction of the diff. The gate gets skipped not because nobody cares but because the economics never worked.

Frao Advisor inverts that. When a reasoning model answers a focused review question for a fraction of a cent, the cost of the gate stops mattering. The constraint becomes process, not price.

What it is

Frao Advisor is a Go binary that speaks the Model Context Protocol. Claude Code launches it as a subprocess; it answers JSON-RPC requests over stdin and stdout. It calls DeepSeek’s native API directly — no router, no proxy, one provider, one model (deepseek-v4-pro).

The design deliberately leaves things out. No multi-model consensus: the same model wearing different hats is good enough. No conversation memory: each request is self-contained, and what is worth keeping — the advice record itself — is persisted for the dashboard. The result is a zero-dependency static binary that runs anywhere and costs nothing to operate.

Seven personas

Each persona is a system prompt that pins a role and a bar. The same model becomes a security analyst, a debugging specialist, or a technical lead depending on the lens you hand it.

These are roles, not models. A code review from the code-reviewer and one from the security analyst read differently because the questions differ — one hunts for correctness, the other for exposure.

Four tools

Three of the four are thin wrappers over the personas. The fourth is where the value compounds: frao-multi-perspective fans a question out across several experts, then a synthesis step reconciles their findings into a single recommendation. For deployment strategy, major refactors, and security audits, this is the default tool.

The protocol

An advisor that sits unused is furniture, so the protocol is opt-out, not opt-in. Every session runs with standing rules: after writing a file past a few lines, a code-reviewer pass; before declaring a task complete, an independent consult; for architectural decisions, the architect; for anything touching auth, network, or configuration, the security analyst. High-stakes calls get the full deliberation.

The effect is procedural rather than aspirational. Review is not something a developer remembers to schedule — it is part of the definition of done. Advisory feedback has to be addressed before a task can close, and critical findings block progress until they are resolved.

The numbers

Since 2026-07-26 the advisor has recorded 459 pieces of advice across twelve tracked sessions.

The per-persona split says where the tool is actually pointed.

Code review dominates — 189 of the 286 reviews. The advisor is primarily a gate that runs after the author believes the work is done, and it exists to be disagreed with. Architecture is the second busiest lane at 69, then security at 22.

The daily activity chart shows the rhythm of a tool that stopped being rationed.

The dashboard

Every advice record is published to a standalone dashboard service, so a single view shows what every session has asked and what every answer cost. It runs in Docker, aggregates from its own SQLite store, and renders six live views plus a per-record drill-down. The screenshots below were taken live on 2026-08-05.

Overview: the four KPIs — 459 pieces of advice, $2.06 total, 2,478,720 tokens — the 30-day cost trend, and the consult / review / deliberation split.

Timeline: every record, filterable by type and persona, with cost, tokens, and latency per call.

Experts: per-persona volume and cost charts.

Deliberations: each multi-perspective run with its synthesis, expandable to the individual expert analyses.

Costs: spend by model and by record type. The entire history fits comfortably in the column.

Metrics: a per-day heatmap of activity across personas, plus latency by day.

Detail: the drill-down for a single record — full question, full response, tokens, and latency.

The cost of a second opinion

2,478,720 tokens of expert reasoning, $2.06 total. The average call runs about 5,400 tokens and costs about half a cent.

At $0.0045 per call the "should I bother?" question disappears. The advisor runs on every change, including the ones you were confident about — and the bugs it catches are rarely the ones you were worried about.

What it changed

The measurable output is volume: 459 reviews in eleven days, 36 of the 286 expert reviews on a single long-running project’s phase work. The qualitative effect is harder to table but harder to miss — the review gate is now a constant, and constants change behavior.

Work that used to end at "it builds" now ends at "the reviewer agrees." Multi-phase work on the robo-chef simulation passed every phase through a review, and the plan was reviewed before it was even presented to a human operator. Architecture decisions got a deliberation before they got a commit.

$2.06 bought a lot of opinions. That is the entire point.

Want to Install It?

This post covered what Frao Advisor is. For the how-to — building and registering the server on Linux, macOS, and Windows — see Frao Advisor: The Installation Guide.