Kriko answers one question about the product you are looking at: what is known to go wrong with this exact one. It runs on your computer — a desktop app for Windows, with an Apple Silicon macOS build made from source, and a browser extension — and keeps the source beside every finding. Installed catalog lookups work offline. Research uses the agent or model you choose, and a paid provider bills your own account.
Download the Windows installer · Browser extension · Website · Known problems by product (no app needed)
The real panel, from a live run: every finding names the page it came from. Watch the 28-second film · How to use it, step by step
0.3.14 adds two catalogs — gaming monitors (17 models, 50 sourced claims)
and condensing combi boilers (12 series, 38) — grades every source a claim
came from and records how well each catalog's claims are backed, and rebuilds
the agent plane: OpenCode and Hermes run as resident workers, a signed-out or
used-up account is reported once instead of paid for on every call, page reads
respect the budget you set and are reused within a job, and a turn that runs
out of time keeps the pages it read.
0.3.13 set the app in Inter with Chivo Mono titles, added an off-by-default
Share my product queries checkbox (only the product model you looked up and
whether the catalog had it), and checks each claim against its own source
before it enters a catalog. The chat turn orchestrator and the API app factory
were also split into named pieces.
0.3.12 made the desktop CLI answer in about a quarter of a second — --version
no longer loads the provider stack, and the read commands (prefs, costs,
sites list, drafts list, operations, verify --list) read the app's own
state directly, so they answer whether or not an engine is running — and adds
an Estimate screen for equipment purchases: lines from your installed
catalogs with quantity and unit price, known issues per line, a research
button, the cost of the next run, a depreciation and NPV forecast over 1–30
years, and Excel or CSV export. A chat turn whose agent session stalls no
longer spends its whole 90-second budget on the stalled session (#337). The
companion extension stays at 0.5.0 and mobile at 0.15.x, in their own
repositories. See the changelog for the release scope.
Compare in the desktop app: three listings, the same checks, the sources beside each. |
Known problems by product on kriko.cc: no app needed, every report links to its page. |
Ask about any product; the answer cites what it read. |
Fires, explosions and shocks, by product: the reports worth knowing before you plug it in. |
| Artifact | Use |
|---|---|
kriko-0.3.14-x86_64-setup.exe |
Per-user Windows installer; no administrator rights required. |
kriko-0.3.14-win64-portable.zip |
Extract the whole archive and open kriko.exe. |
- Open Kriko, choose an agent in More → Agents and press Check connection.
- Open Catalogs → Find release catalogs and install one, or import a
.kpackfile. Research also works with no catalog installed. - Add Kriko to Chrome, Edge or Brave, keep the desktop app running, and press the toolbar button or Alt + K on a product listing.
The installer is unsigned, so Windows may warn before it runs. Published desktop
artifacts are Windows x86_64. A native Apple Silicon build is made from source
by the desktop recipe — kriko-ui/package-macos.sh — and is not a release
download yet. Details: Windows installation,
macOS installation.
- Asks about a product, compares options and continues the same conversation.
- Shows sourced problems, exact variants, specifications and coverage gaps — beside the listing through the extension, or in the app.
- Watches a price, ranks a shortlist to one pick, and asks later whether a purchase held up.
- Keeps catalogs, conversations and research history on your computer.
Kriko Agent is experimental and runs on your own provider keys; shared access and the trial are in shared access. The app's screens and how to grow knowledge are in Usage; matching, the catalog lifecycle and layering are in How it works.
Kriko Agent is a fixed chain — route the question, read Kriko's own records, research the web when they do not answer, cite the pages, then act in the app — and it is compared against the previous code rather than described.
- Routing: the product-failure question reaches the installed records on 42 of 42 test messages, against 33 of 42 before. Ordinary chat and buying questions do not move.
- Claim extraction: 8 of 8 documents keep their expected claims, against 5 of 8 before. The old extractor stopped at the first 6,000 characters.
- One live turn: 11.45 s, eight pages read, a corrected premise (that engine has a belt, not a valve-train chain), the reader's preference saved in For you, and $0.0060 spent.
- The honest cost: the same tokens cost about 44% less on Mercury 3, but Mercury 3 can choose more searches and erase that advantage. The eager profile in the product-search harness did exactly that.
The full method, the small-sample caveats and the reproducible commands are in the benchmark note. This is a measured development stack, not a scored quality distribution.
Everything lives in %USERPROFILE%\.kriko on Windows (~/.kriko for a source
installation): knowledge.sqlite holds installed catalogs, app.sqlite holds
history, conversations and settings, and an update or uninstall keeps the folder.
Nothing leaves the machine except the question a research turn sends to the
provider you chose.
Source work needs Python 3.13+, Node and a Rust toolchain for the native shell.
tools/setup.sh
cd kriko-ui && npm ci && cd ..
tools/gate.shThe gate covers Python, the wheel, the extension and website tests, the desktop app's TypeScript and behavior, and the native checks; a missing native check is incomplete, not a passing release. See Architecture, the desktop build recipe, and report a problem.
MIT.
Our own inference service, a fine-tuned model, and statistical knowledge extraction from forums.


