Skip to content
Analyse an idea

How it works

Seven agents. Every claim labelled.

IdeaGuard doesn't produce one long answer. It runs a small team of specialised agents as a dependency graph, checks their citations against what was actually retrieved, and shows you which parts of the report are evidence and which are reasoning.

The pipeline

  1. First

    01Idea Analyst

    Everything else depends on a precise restatement of the idea.

  2. Then

    02Research Agent

    Live web search. Sources are numbered and kept, even if a later step fails.

  3. In parallel

    03Competitor Agent

    04Feasibility Agent

    05Risk & Governance Agent

    Independent views of the same evidence, run concurrently.

  4. Then

    06Critic

    Needs at least two of the three above. Told exactly what's missing.

  5. Last

    07Product Strategist

    Needs at least three upstream inputs. A stance, an MVP and what to test first.

Reading a report

Evidence

Backed by a retrieved page. The citation links to it, and the citation was checked to exist.

Inference

Model reasoning without a source. Often useful, never presented as fact.

○ Unavailable

A step didn't complete. The section says so, and nothing is filled in for it.

Guardrails

  1. 01

    Evidence is checked in code, not trusted

    Agents may only cite sources that were actually retrieved. Every citation is checked against that list after generation; an unsupported claim is relabelled as inference, and a competitor that no source names is marked unverified. The model can't promote its own guesses to facts.

  2. 02

    Gaps stay gaps

    When a step fails, the report says so, and downstream agents are told what's missing instead of guessing it. Nothing fills the hole with plausible text. Retrying re-runs only that step and what depends on it.

  3. 03

    Structured, validated output

    Each agent answers in JSON matched against a schema. One repair pass sends back the exact validation problems; a truncated answer gets one retry with more room. After that it stops and shows why.

  4. 04

    Built for free-tier limits

    Calls share a tokens-per-minute budget per provider, learned from the provider's own rate-limit headers. Rate limits are waited out rather than failed, and a second provider can take over when the first is unavailable.

  5. 05

    Runs survive the tab

    On Vercel an analysis runs as a durable background workflow. Each agent is a persisted step, so closing the tab or losing the connection doesn't lose the run; reopening the project picks up where it was.

  6. 06

    Your ideas stay in your browser

    There are no accounts. Projects, versions and experiment notes live in local storage. A report leaves the browser only when you share it, and the shared copy omits experiment notes and other versions. You can withdraw a link at any time.

What it can't do

  • Market sizes, prices and growth figures are only as good as the pages search returns. Check the cited source before relying on a number.
  • Signals (demand, competition, effort…) are judgements, not measurements. They're shown as levels and never combined into a score.
  • Without a search key, research runs on model knowledge only, and the report labels every section accordingly.
  • The critique is adversarial by design. A hard stress test is a reason to run the experiments, not to drop the idea.
Analyse an idea