Decosa: AI tools that do one job and show their work.

What Decosa is

Decosa makes AI tools that each do one job a professional already does by hand: digest a deposition, check a brief before filing, appeal a denial, fill a form from your own papers. They are grouped by the work they serve, from legal, healthcare, finance and HR to compliance, operations, life sciences and creative work. Every one opens on a sample, so you can see a real run before you sign up.

Under the tools are open-weight models, named with their versions and licences on each tool’s page, so anyone can run the same models themselves. The tools are also an API in the OpenAI chat format: any OpenAI SDK or coding agent works once its base URL points at Decosa. Usage is paid in credits of one US dollar each, and a new key gets a small free daily budget to try it.

There are 86 tools, 73 of them live and 13 in preview. A separate Labs page lists 22 more working tools that have not cleared the bar yet, labelled that way.

A signed receipt for every model call

Every hosted model call from a Decosa tool comes with a receipt: which model served the request, hashes of exactly what went in and came out, and the token counts, signed with an Ed25519 key that Decosa publishes. Receipts hold hashes and signatures, never the text, and anyone can check one on the site or through the API.

The docs are plain about what that proves. A receipt shows that Decosa signed the record, that nothing in it changed afterwards, and what Decosa sent and received. It does not prove which weights an outside provider ran, or that the provider kept no copy. A self-hosted box signs its own receipts with a key it makes on first start, which is an attestation by that box, not a proof that the model ran.

Measured in public

Every tool shows its time per task, cost per task and accuracy on a test set, with the dataset, what was held out, the caveats and the date. A nightly check re-runs the hosted tools on their samples. On October 20, 85 of the 86 tools had an eval of their own, 69 were scored on cases held out from tuning, and 73 had passed a fresh self-host setup run end to end.

Two examples, with their limits. Check a brief before filing caught 17 of 19 planted fake citations (89%) in held-out excerpts of real Solicitor General briefs, and called 19 of 511 items in unaltered real briefs problems when they were not (3.7%). The set is small, so the intervals are wide. Fill a form from your papers wrote 1 wrong answer in 491 (0.20%) on 50 held-out synthetic PDF forms, and its eval says real flat forms will be harder.

Most tool evals use synthetic or public test sets. The accuracy page says so directly: a synthetic test set is not real-world performance, so run your own sample first.

Open models, including our own

Hosted calls go encrypted to Decosa’s API server. The open models are run by NEAR AI through OpenRouter, with Reka AI as the only fallback, under Decosa’s account and with zero data retention requested, and nothing is used for training. Each receipt names who served the call.

Where no open model was good enough under a licence that allows commercial use, my lab trained its own. Eleven small verifiers and readers are on Hugging Face under Apache-2.0, and two of them are already building blocks in the API: citation support and note-detail checking. Their held-out results and limits are on the models page.

Self-hosted, sealed or confidential

For data that cannot go to an ordinary hosted service there are three routes, at different stages. Self-hosting runs a tool and its model on your own NVIDIA GPU or Mac, with nothing sent to Decosa. It is in early access and on request, and its code is licensed AGPL-3.0-or-later, though the source is not public yet.

A sealed tier encrypted each prompt on the user’s machine with HPKE (RFC 9180) and decrypted it only on a model server Decosa operated, so Decosa’s API forwarded only ciphertext. It is paused at launch: hosted calls now run on outside providers, so there is no Decosa server to decrypt on, and the sealed endpoints are off rather than falling back to plain text.

It returns as a confidential tier, with the model and the decryption inside a hardware-attested enclave. In a one-day pilot on 27–28 September 2026, on a rented NVIDIA H200 in an Intel TDX confidential VM, 3,329 sealed requests ran with no errors, 33 of 33 tools ran end to end through the enclave on their samples, and 249 of 249 sealed receipts checked out offline. There is no permanent confidential node yet. It is a tested pilot, offered by arrangement, and not yet for patient data.

Apps by Decosa

Decosa, Inc. makes Everdust, the physics and music app. Five more apps come from the same team and keep their own products and accounts: Cue Mixtapes, Snug, AutoTalent, ElmoSEO and Dr. Grey. The hard part under each, its engine, is moving into Decosa one at a time, so every app and every Decosa tool can use it. Next are the mix cue engine from Cue, the room engine from Snug, the signed decision record from AutoTalent, the site-check engine from ElmoSEO and the evidence engine from Dr. Grey.

Everdust already uses the same kind of stack. Its scene Creator, in beta on TestFlight, builds scenes with open models that Decosa can serve, DeepSeek V4 Flash for its Quick mode, called through OpenRouter for now rather than Decosa’s own hosting.

What is not there yet

Several hosted services are switched off for now, among them live transcription, the document reader, translation and Studio renders. Tool pages that need one show a recorded run and say so. The sealed tier is paused, the confidential tier is a pilot, self-hosting is on request until the code and container images are published, and the site’s chat guide is off.

What exists is the public site, the API, the published evals with their caveats, and the open models. This page makes no claims about who uses them.