// the spec layer between intent and code
SparkVibeAI is a calibrated spec authorship platform. You describe what you want built — in the browser or straight from your coding agent over MCP; a house committee writes the build spec; a chained review pipeline — independent cross-model reviewers, deterministic structural checks, and a reviewer-derived score — pressure-tests it at write time so you know what you're getting before you build it.
Most "AI writes my spec" tools share a single problem: nothing tells you whether the spec is any good. The model wrote 2,000 confident words. Maybe they describe your idea. Maybe they're plausible-sounding boilerplate the model would have generated for any similar prompt. You find out which by trying to build it.
SparkVibeAI exists because that bet is too expensive. The fix isn't a better writer — every model gets better every quarter. The fix is independent scoring at write time, calibrated so the scores actually track what matters. That's the entire thesis.
Every accepted spec passes through a chained review pipeline: two independent reviewers audit it, a deterministic validator checks its structure, and a reviewer-derived score reads out of those findings. The signals run on different mechanisms, so where they converge you have real signal.
These signals run on different mechanisms — two cross-model reviewers, deterministic structural code, and a score derived from their findings — so when they converge, you have real signal. When they diverge, the divergence IS the signal: an unresolved blocking finding points at exactly where the spec is overstating itself. The score is an explainable derivation of the reviewer and structural results today; calibration against real build outcomes is future work, not a number we hide a claim behind.
The strongest version of independent review is reviewers from different model families — different training corpora, different RLHF philosophies. SparkVibeAI is built on a vendor-agnostic LLM client layer so the committee can route the two reviewers to different providers — today one Anthropic, one OpenAI. When two models trained on different corpora confidently agree, the agreement is much stronger than two calls to the same model.
This isn't theoretical. It's api/committee-runs/llm-clients/ — three vendor clients today, one router, designed from day one to add more without restructuring.
You don't have to leave your editor. SparkVibeAI ships an MCP server — add it to Claude Code, Cursor, Codex, or Grok and your agent commissions a spec, polls it to a reviewer-derived score, pulls the body, and can push the selected spec into a private GitHub repo, all inside the agent loop. It's the same calibrated pipeline as the web app, and the score travels with the spec.
The core loop is four tools — commission_spec → get_commission → get_spec → push_to_github — plus read tools for the spec body, subscription usage, and the pre-spec interview. Agentic MCP access is included on Pro and up; generate a key from your account and paste the config. See the setup →
Folsom Holdings, LLC is the legal entity behind the Platform — the operator, the data controller, the recipient of subscription fees, and the party responsible to users under the Terms of Service and Privacy Policy.
We're deliberately small. The product is built by a focused team with strong opinions about what good infrastructure looks like — versioned schemas, drift markers across duplicated code, audit responses that land within hours of a finding, deterministic baselines that don't depend on any model staying online. The cross-model review architecture is what we'd want a calibrated AI service to look like before we ever paid for one.
Your ideas are yours. We don't take, build, sell, or train models on what you're actually making — the substance of your spec is never ours. We may analyze specs structurally — completeness, how they score, where the committee converged or struggled — in aggregate, to sharpen the platform's scoring and the committee itself. That's the shape and quality of specs, never the idea inside them. We retain rights only in the platform: the scoring infrastructure, the rubric, and the Aletheia overlay. The full commitment is in Section 5 of the Terms.
The full public API contract lives in docs/public-api-contract.md in the repository. Privacy details are in the Privacy Policy.
Questions, feedback, press, partnership, or anything else: support@sparkvibeai.com or the contact form.
For legal questions specifically: see the Terms of Service or contact us at the same address.
SparkVibeAI is a service mark and product of Folsom Holdings, LLC. All rights reserved.