Skip to content

For agents

Ask with your own agent

This site runs a small MCP server. It’s read-only, uses no language model, and answers from the same sourced facts as the pages, so your agent gets numbers with their sources and an honest status, not a sales pitch.

Connect

In Claude Code:

claude mcp add --transport http solai https://solaimuthu.vercel.app/mcp

Any client that speaks Streamable HTTP can use https://solaimuthu.vercel.app/mcp. Limit: 60 requests a minute. No writes, no accounts, nothing stored.

Tools

get_profile
Who Solaimuthu A (Solai) is: role, experience, education, achievements, the stack he has shipped with, and how to reach him.
list_builds
Every build in his ledger, newest first, with dates, team, his part, one verified number and an honest result.
get_case_study
A case study: problem, his part, team, stack, verified metrics with sources, result, links and the full write-up.
get_fact
One number from the facts ledger with its source and status (verified, self-reported, pending, student estimate, derived).
list_facts
List ledger facts (id, label, display, status), optionally by id prefix.
search_sketchbook
BM25 search over the site's own content. Returns passage ids to cite with check_claim.
check_claim
Run Redline's verifier (code, not a model) on a sentence about Solai: are the cites real, are its numbers and names in the cited passages, and does it credit him with a teammate's work?
run_demo
Run one of the site's demos as a deterministic simulation of recorded numbers: why-graph-guard, rag-budget, projectx-funnel or mission-energy.

A real call

Computed when this page was built, by the same function the Why-Graph demo on Sheet 05 uses.

→ run_demo {"demo":"why-graph-guard","timeoutMs":300}

← {
  "note": "Deterministic simulation of recorded numbers; see the source.",
  "source": "https://github.com/redwing-381/we-make-devs-hack (staged incident on the astro-shop demo app)",
  "change": "shipping → quote timeout",
  "risk": "high",
  "apiRisk": "high",
  "safeMinMs": 453,
  "lesson": "L-1",
  "cites": [
    "L-1",
    "INC-1031",
    "D-1"
  ]
}

Without MCP

How the sketchbook agent scores

Measured on 34 cases: answerable questions, refusals, prompt injections and attribution traps. Published only after a real run.

ModeDateAnswerableRefusals + injectionsHonestyGrounded claimsp95$/run
evidence-only (no model)2026-10-19100%100%100%100%2 ms$0.0000

The live-model row appears once the live eval has been run against the model.