Skip to content
decosa
decosa
Get an API keyTry a tool

Industries

LegalBriefs, depositions, discovery, records, client calls.HealthcareThe visit, then notes, audits and denials.Finance & insuranceAlerts, SARs, disputes, claims, recorded calls.HRInvestigations, leave and I-9 audits.Compliance & trustSecurity questionnaires, control evidence, AI audits.OperationsForms, shop pages, imports, listings: any industry.Life sciencesTrial numbers, safety cases, labels, papers.

More industries

Public sector 2 toolsSales and marketing 6 toolsScience and research 6 toolsMusic 5 toolsFilm, TV and games 6 toolsAll industries

Create and tools

CreateTools

Developers

Decosa APIQuickstartAPI referenceAPI pricingBuilding blocksDecosa agentAssembly promptsFor AI agentsSelf-hosting docsVerify a runAccuracyStatus

Trust

Where your data goesHosted APISelf-hostOn a MacConfidential optionDecosa for ChromeAccuracyVerify a runStatusSecurityRegulation watchCompare ways to run it

Company

PricingAboutApps by DecosaContact

For AI agents

Using Decosa from an agent

Decosa has 86 tools on open-weight models (73 live, 13 in preview). Every hosted model call gets a signed receipt, and every tool can be self-hosted (on request while it's in early access). This page is for coding agents, browser harnesses and scripts that read or drive the site. The same facts are in plain text at /llms.txt and /llms-full.txt.

Machine-readable endpoints

Prefer these to scraping. They are built from the same data as the pages.

/api/catalog.json
Every tool, industry, product and building block, with status, verification and links.
/use-cases/<id>.json
One tool: stack, tiers, verification, buyer facts, where its data goes (data_handling, explained at /data), console samples and prompt URLs. Example: /use-cases/privilege-log.json
/api/metrics.json
Per tool: eval_summary (headline metrics with split, n, dataset, held out or not, caveats, date), QA-sweep verification, cost per run, rehearsal bundle and models. The page is /metrics; live nightly results come from GET /verify/status on the API.
/api/hardware.json
The hardware catalog (memory, bandwidth, FP8/FP4, price, datasheet links) and, per tool, the memory each tier needs (component by component, with its basis: measured, stack or estimate) and the verdict on common boxes. The interactive version is /self-host/hardware.
/prompts/<id>-selfhost.md
The self-host prompt, raw markdown. Also -hosted.md and -assemble.md (the Stack tab's assembly prompt).
/api/contract.json
The decosa-api routes, indexed from the contract. The contract itself: /api/contract.md, rendered at /docs/api.
curl -s https://decosa.ai/api/catalog.json \
  | jq '.use_cases[] | select(.industries | index("legal")) | {id, name, status}'

Driving the pages

  • Landmarks: one header, nav labelled “Main”, main and footer. One h1 per page, headings in order. Every button, link and field has an accessible name, so role-and-name locators work.
  • Links are real hrefs. Search results, catalog cards and menus all navigate by URL.
  • Tool pages have tabs addressable by hash: /apps/<id>#stack, #live, #watch, #build, #self-host. Arrow keys move between tabs.
  • Catalog filters live in the query string: /tools?industry=legal. Params: industry, job, input, deploy, status, verify, collection, q, sort, view.
  • Console deep links: /console/<id>?sample=<n>&autorun=0 preselects sample n (1-based, or a sample id). autorun=1 also starts it: a live run, or on slow tools the recorded real run first; add &mode=live or &mode=replay to choose. Example: /console/privilege-log?sample=1&autorun=0.
  • Add ?reduce-motion=1 to any URL to turn animations off for the rest of the tab session (?reduce-motion=0 turns them back on). prefers-reduced-motion is respected too.
  • Every copy button has its text visible next to it, so you never need clipboard access. The same text is at the /prompts URLs.
  • Nothing is shown only on hover.

Stable test hooks

These data-testid values are kept stable across redesigns. Tab panels that are not shown stay in the DOM (hidden), so scope locators to the visible panel, e.g. #panel-build.

data-testidWhat it is
use-case-card, use-case-rowA card (or list row) in the /tools catalog. Carries data-use-case-id and data-status.
catalog-search, catalog-filterThe catalog's search field and filter chips (data-facet, data-value, aria-pressed).
search-openThe header Search button (also ⌘K or Ctrl+K anywhere).
search-inputThe search combobox. Results are links with data-testid search-result and data-href.
tab-stack, tab-live, tab-watch, tab-build, tab-self-hostThe tabs on /apps/<id>. The URL hash selects them too.
copy-prompt-selfhost, copy-prompt-hosted, copy-prompt-assembleCopy buttons for the prompts. The text is in the <pre> beside each (data-testid copy-text), and at /prompts/<id>-<kind>.md.
rehearsal, rehearsal-download, rehearsal-expected, rehearsal-commandThe "rehearse on mock data first" section on /self-host and each Self-host tab (data-use-case-id): the bundle zip, its expected.json and the rehearsal command. Also in /use-cases/<id>.json under rehearsal.
copy-snippet, copy-textCopy buttons for code snippets and keys, and the visible text beside any copy button.
get-api-keyOpens the key form. Then key-name, key-create, and key-value for the new key (shown once).
console, console-picker, console-picker-optionThe console root (data-use-case-id) and the use-case switcher; options are links.
console-sampleThe sample picker in a console.
console-run, console-stop, console-recordStart and stop a console run; console-record starts the microphone on live audio tools.
console-outputThe result area on consoles that stream a result. aria-busy is true while a run streams.
watch-sample, watch-play, watch-seek, watch-speed, watch-outputThe Watch replay: pick a recorded run, play, seek, speed, and the replayed result.
data-badge, data-badge-hostedWhere the data goes, on use-case headers and catalog cards (data-self-host: nothing, identifiers or content; data-hosted: operator, operator-plus-third-party or demo-only). Both link to /data#<id>.
metrics-table, metrics-row, metrics-search, metrics-filter-industry, metrics-filter-status, metrics-filter-nightly, metrics-filter-cost, metrics-filter-heldout, metrics-sort, metrics-countOn /metrics: the table, one row per tool (data-use-case-id, data-status, data-nightly: pass, fail, partial or none, data-nightly-source: nightly or qa, data-held-out, data-cost) and its controls. Filters and sort are in the query string (industry, status, nightly, cost, held_out=1, sort, q).
metrics-detail, metrics-eval, metrics-eval-metric, metrics-caveats, metrics-nightly, metrics-history, metrics-selfhost, metrics-bundle, metrics-models, use-case-metrics-linkOn /metrics/<id>: the eval (data-held-out; each metric has data-split), the nightly check (data-source, data-result) with its run history, self-host verification and rehearsal bundle, models. use-case-metrics-link is the link from /apps/<id>.
data-level, data-flow-rowOn /data: the three levels (data-level: self-host, hosted, sealed) and one row per tool (data-use-case-id).
hardware-check, hardware-check-row, hardware-summary, hardware-summary-rowThe hardware check on each Self-host tab (data-use-case-id; each row has data-device, data-verdict: runs, smaller, no or unknown, and data-tier) and the per-box summary on /self-host.
hardware-fit, hardware-use-case, hardware-mode-catalog, hardware-mode-custom, hardware-device, hardware-count, hardware-custom-arch, hardware-custom-vram, hardware-custom-count, hardware-custom-ram"Help me customise for my hardware" on each Self-host tab and on /self-host/hardware (root has data-use-case-id and data-verdict). The standalone page keeps its choice in the query string: ?use=<id>&hw=<device>&n=<count>, or ?use=<id>&vram=<GB per GPU>&gpus=<n>&arch=<blackwell|hopper|ada|ampere|cdna3|rdna3|apple|cpu>&ram=<GB>.
hardware-recommendation, hardware-tier-row, hardware-tier-pick, hardware-substitutions, hardware-substitution, hardware-speed, copy-prompt-hardwareIts output: the recommended tier (data-tier, data-verdict), one row per tier (data-tier, data-fit: runs, changes, no or unknown) with a radio to pick it, the model swaps, measured speed (data-measured: yes or no) and the setup prompt with the hardware plan.
hardware-catalog, hardware-device-rowThe device table on /self-host/hardware (data-device-id), with memory, bandwidth, FP8/FP4, price and source links.
guide-open, guide-open-menu, search-ask-guide, guide-nudgeOpen Ask Decosa, the site guide: the bottom-right button, the item in the mobile menu, the last option in the ⌘K palette (it asks what you typed), and a one-time hint on the home page. window event "decosa:open-guide" with {q} or {profession} opens it too.
guide-panel, guide-opener, guide-profession, guide-input, guide-send, guide-stop, guide-close, guide-starterThe guide panel (not modal: it stays open while pages change; html[data-guide=open] while open; Escape closes), the "What do you do?" opener and its profession chips (data-id), the message box, send and stop, close, and suggested questions.
guide-job, guide-nav, guide-undo, guide-strip, guide-expand, guide-collapseJob chips after a profession (data-id = tool id), the "Taking you to …" line when the guide opens a page, Undo, and on phones the one-line strip the panel folds to after a navigation, with Chat to reopen.
guide-request, guide-request-send, guide-request-done, guide-request-status, guide-feasibilityThe request card the guide drafts when no tool fits (the visitor reads it and presses Send), its preliminary feasibility read, the sent confirmation and the link to /requests/<token>.
guide-turn, guide-answer, guide-source, guide-link, guide-noticeA message and answer (guide-answer has aria-busy while it streams), the source chips under it, links inside it, and the notice when identifiers were removed. The same answers come from POST /api/guide (NDJSON, with nav and request events); an agent can read the pages it cites directly instead.

Common tasks

  1. Find a tool: read /api/catalog.json, or open /tools and pick a use-case-card, or press ⌘K and type.
  2. Route a creative request (music, music video, animatic, audio drama, characters, UGC ad, dubbing): start at Decosa Studio (one app with packs: Story, Music, Film, Voice, Characters & Worlds, Brand & Ads, plus Studio checks). In /api/catalog.json, the studio product lists its packs as branches and their tools, and each tool carries part_of (e.g. studio-music). Every tool runs through the consent ledger and adds C2PA credentials; the shared project (cast, world, look) is live for Story at /studio/projects.
  3. Get the self-host prompt: fetch /prompts/<id>-selfhost.md, or open /apps/<id>#self-host.
  4. Get a hosted API key: /account/keys, then get-api-key (if shown), key-name, key-create, and read key-value. Keys are shown once; three a day per browser.
  5. Run a sample: open /console/<id>?sample=1&autorun=1 and wait for console-output to lose aria-busy.
  6. Call the API: Authorization: Bearer dk_… against the base URL in /api/contract.json. Every model call returns a receipt id you can check at /receipts/<id>.

Please use synthetic or public data in the hosted demo. For real client or patient data, self-host.