Decosa API
Every tool on this site and the open models under it, by API in the OpenAI format, with a signed receipt for each model call. Or run any of it on your own hardware.
Decosa API
A key, a first call, a receipt.
The hosted API speaks the OpenAI format, so the SDK or coding agent you already use works. It serves every tool on this site and the open models under them, and each model call comes back with a signed receipt.
Checking the API’s live status…
export DECOSA_API_KEY=dk_... # from /account/keys
curl https://api.decosa.ai/v1/chat/completions -i \
-H "Authorization: Bearer $DECOSA_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "qwen3.8-27b", "messages": [{"role": "user", "content": "Hello"}]}'
# The answer, and a signed receipt id in the x-decosa-receipt header.Who serves it
Hosted calls are processed by Decosa's API server, with the open models run by NEAR AI through OpenRouter, with Reka AI as the only fallback under Decosa's account. Each receipt names who served the call; the companies are listed below.
Where your data goesCredits
One credit is $1 of usage at list price: hosted API calls, the coding agent, Studio and tool runs all draw on the same balance.
$1 a credit, in US dollars. A new key has a free budget of $0.02 of usage a day; a key with credits has no daily cap. At zero, calls stop with HTTP 402 and a link to buy more.
API pricingReceipts
Each hosted model call is signed with Decosa's Ed25519 key: the model, and hashes of what went in and came out. A receipt shows the record wasn't changed after signing; it holds hashes, not your text.
Verify a run| Company | What it does for Decosa | Policy |
|---|---|---|
| OpenRouter | Routes Decosa's hosted language-model calls to NEAR AI (Reka AI as the only fallback), requiring zero data retention and no data collection on every request. | Privacy |
| NEAR AI | Runs Qwen3.8-27B in FP8 inside a TEE (a hardware enclave) for the hosted API. | Privacy |
| Reka AI | The only fallback, when NEAR AI is unavailable: Qwen3.8-27B in FP8, keeping no prompts. | Privacy |
| fal | Renders images and video for UGC ads (the render prompt and inputs); keeps request data 30 days by default, and outputs sit on its CDN for at least 7 days. | Privacy |
The services that need no GPU run on Decosa's own server: the tools' checkers and lookups, music similarity and beat alignment, Kokoro and Chatterbox speech, and the browser that runs site checks.
Coding agents: start at For AI agents (llms.txt, catalog.json, the MCP server).
The Decosa agent
A private coding agent, and Decosa in the agent you already use
decosa agent is a coding agent on the open Qwen3.8-27B model. Run it on your own GPU and nothing leaves your machine, or use it on Decosa's hosted service with a signed receipt for every call. The same install adds the Decosa tools to Claude Code, Cursor or any MCP client, so they can run a Decosa tool on your input, verify a receipt offline, plan a self-host install or check a GPU box.
1. Install (Python 3.10+)
# The Decosa agent is in early access: ask for it at https://decosa.ai/contact?topic=self-host2. Run the coding agent (Node 22.19+)
decosa setup && decosa agentOr add the Decosa tools to Claude Code (Cursor, Codex: see setup)
claude mcp add decosa -e DECOSA_API_KEY=dk_... -- decosa-mcpPrivate beta. The install line works for beta accounts with repository access; the public package is waiting on release approval.
What you can ask it
- Find the right tool for a job and show its working example call
- Run a tool on your input and save the full output and its receipts
- Verify a receipt offline against the pinned signing key
- Plan a self-host install for your GPU: pinned images, models and a smoke test
- 93%
- of 46 Decosa tasks done right by Qwen3.8 with these tools (100% with thinking on)
- 46%
- the same model as a generic agent with our docs and a raw HTTP tool
- $0.007
- per task at list price, a quarter of the model calls a generic agent makes
Our own eval, 46 tasks written for it, 28 Sep 2026: not a public benchmark. In Claude Code with the MCP server, 4 of 5 hand-run tasks were right the first time. The fifth (can an RTX 4090 run the coding model?) was wrong, and the tool now checks the GPU architecture.
Why developers use Decosa
- Open models, with receiptsEvery hosted call runs on pinned open weights and comes back with a signed receipt: which model, which weights, a hash of what went in and came out. You can check it yourself, offline.
- Prove what an endpoint servesPaying for an open model by the token? Check that the endpoint really serves the model and precision you're paying for, with a signed report you can hand to procurement.
- Run it yourselfEvery tool also ships as a self-host kit: pinned containers, a copy-paste prompt for your coding agent and a smoke test. Your code and data never leave your machine.
Jobs for developers
Make and do
- Run an open model behind the OpenAI APIAlready on the OpenAI SDK? Change the base URL.Qwen3.8-27B behind a standard OpenAI-style /v1 endpoint: try it hosted, with a signed receipt per call, or run it on your own GPU so nothing leaves the box. Hosted replies are capped at 2,048 tokens and don't take tool calls yet.478 / 500Coding benchmark, 5 core tasks, self-hosted stack (out of 500)Try it
- Turn a prompt and 50 examples into a small model you ownPlannedYour examples train a small open model, tested on held-out cases. The weights are yours, under Apache 2.0.
Check and prove
- Check if an open model can take over your promptMoving off GPT to cut cost?Replay prompts you've already logged on an open model. In minutes you see whether it gives the same answers, which examples differ and why, and what 1,000 requests would cost: go, no-go or not enough data yet.14 of 15 pairingsGo/no-go verdicts that match human experts' verdicts (held-out test)Try it
- Check your provider serves the model you pay forPaying a provider for “Qwen 27B at BF16”?Find out whether an OpenAI-style endpoint really serves the model it claims, or a smaller or more compressed one. You get a signed pass, drift or fail report you can put in the vendor file, and anyone can re-check it.~$0.18 per 100 auditsmeasured, at list priceTry it
- Catch AI answers your sources don't backYour bot answers from your docs. Does it make things up?Every sentence of an AI-drafted answer checked against the source passage it relies on, with the quote, and a pass, flag or block you can put in front of Send.0.966Unsupported sentences caught, strict gate (it flags about 30% of sentences for a look; held-out)Try it
- Prove what your agent didCompliance asks what your agent did, and why.A signed, step-by-step record of what your agent saw, decided and did, which anyone can re-check. Change one step later and verification names it.339/339Edited copies of a record caught, signing key pinned (synthetic set)Try it
Start to finish
Move a prompt off GPT
Check an open model can take over the prompt, call it through the OpenAI API, and check the provider you'll pay serves the model it claims.
- Step 1: Check if an open model can take over your prompt
- Step 2: Run an open model behind the OpenAI API
- Step 3: Check your provider serves the model you pay for
Ship an agent compliance will sign off
Block answers the client's documents don't back, and keep a signed record of every step the agent took.
See the workflow
My runs
No runs yet. Open a tool above and run a sample; it will show up here. Runs are saved on this device only.
Building blocks22
Every Decosa tool is made from these. Use them in your own product through the API or the Decosa agent.
Check what a model wrote
- GroundingChecks each sentence against its sources: supported, partial, unsupported or contradicted, with the span it rests on.API
POST /grounding/checkAsk the agent “Check this answer against these sources with Decosa.”Try it, with the request shape →Inside 38 Decosa tools - Numeric groundingEvery amount, date, count, span and percentage in a sentence recomputed in code from the table rows it cites; a mismatch comes back with the figure the rows do support.API
POST /sar/numbersAsk the agent “Check every number in this draft against the case data.”Try it, with the request shape →Inside 8 Decosa tools - Typed judgmentYes/no, one-of-N, score and label questions answered as typed values with a calibrated probability and a receipt.API
POST /judgment/judgeAsk the agent “Label these tickets with Decosa's typed judgment and give me the probabilities.”Try it, with the request shape →Inside 28 Decosa tools - Small checker modelsTwo small models of our own, Apache-2.0 on Hugging Face, that run on a CPU: citation support (does the cited passage back the sentence?) and note-detail checking (did a drug, dose or date change between transcript and note?).API Hugging Face: decosaaiTry it, with the request shape →
Prove what ran
- Signed recordA hash chain over every entry, signed with Ed25519, that anyone can re-verify in the browser.API
POST /record/verifyAsk the agent “Verify this Decosa record.”Try it, with the request shape →Inside 80 Decosa tools - Record AI decisions about peopleA signed, hash-chained record of every automated step your app takes about a person (score, screen, rank, filter, a message sent without approval) and the human decisions after it, holding hashes and your own ids, never personal data. Explain one person's steps in plain words and export a retention log. Open-source TypeScript; nothing is sent to Decosa.API
POST /record/verifyAsk the agent “Verify this decision-record export and tell me which entry changed.”Try it, with the request shape → - Agent flight recorderA signed, step-by-step record of what a browser or computer-use agent saw, decided and did.API
POST /flight/runsAsk the agent “Record this agent run with the Decosa flight recorder.”Try it, with the request shape →Inside 6 Decosa tools - Endpoint auditProbes an OpenAI-compatible endpoint for model identity, quality and speed, and signs the verdict.API
POST /audit/runsAsk the agent “Audit this endpoint: does it serve the model it claims?”Try it, with the request shape →Inside 2 Decosa tools - Content credentialsC2PA credentials and watermarks on rendered files, and a checker for any file.API
POST /provenance/checkAsk the agent “Check this image's content credentials.”Try it, with the request shape →Inside 17 Decosa tools - Consent gateA signed consent ledger for voices, faces and movement: every render is checked against a specific, revocable consent first, and the decision is signed either way.API
POST /consent/checkTry it, with the request shape →Inside 6 Decosa tools
Read, hear and act
- Document readerPDFs (born-digital or scanned), images, forms and handwriting read into elements with page and box: text in reading order, tables as cells with spans, form fields and checkboxes, and a signed receipt over the page pixels.API
POST /docreader/readTry it, with the request shape →Inside 7 Decosa tools - Evidence retrievalFinds the passages that answer a question in your documents: an embedder, BM25 and a reranker, each hit a chunk with byte offsets and a Merkle proof into a signed index snapshot.API
POST /retrieval/searchTry it, with the request shape →Inside 6 Decosa tools - Citation indexLooks up case citations in a local index of public court opinions first (a remote lookup only on a miss), so invented citations are caught without sending the brief out.API
Inside POST /preflight/checkAsk the agent “Check the citations in this brief with Decosa.”Try it, with the request shape →Inside 2 Decosa tools - Live speech to textStreaming transcription from a microphone or a sample script, sentence by sentence.API
POST /lang/transcribeTry it, with the request shape →Inside 11 Decosa tools - Speaker diarizationWho said what: transcripts with speakers from recorded audio.Try it, with the request shape →Inside 14 Decosa tools
- Language packTranslation into all 24 EU official languages with a number lock, the model chosen per language from measured scores: every number, unit, date, code, negation and required term checked in each language's own format, drift flagged with its source span. Speech in consented house voices, speech recognition, and a signed receipt for every model call.API
POST /lang/translateTry it, with the request shape →Inside 3 Decosa tools - Video understandingOne recording per request to Qwen3.8-27B, sampled at 1 frame a second (up to 240 frames, 32,768 video tokens), answers that cite time ranges, and a receipt whose request hash covers the video's sha256 and the sampling settings.Try it, with the request shape →Inside 2 Decosa tools
- Form fillingFills a PDF or web form from your own documents, shows the source line for every value, never types passwords or card numbers, and stops before Submit.API
POST /fill/pdf/runAsk the agent “Fill this PDF form from these documents with Decosa, and stop before submitting.”Try it, with the request shape →Inside 4 Decosa tools
Make
- Studio renderMusic, image and video generation on pinned checkpoints, queued, with a seeded render receipt.API
POST /studio/jobsTry it, with the request shape →Inside 12 Decosa tools
More
- Evidence tables from PubMedWhat the trials, meta-analyses and reviews in PubMed found for a supplement and an outcome: one row per study with its PMID, design, n, direction, finding and a quote checked word for word, each finding judged against its own abstract, graded by a published rubric in code, with a signed record.Eval and API →Inside 1 Decosa tools
- Song starts in long audioWhere each song starts in a mix, a radio show or any long recording, from a tracklist (with or without times), your own track files and the audio itself, on CPU with no model call: a chaptered .m4a, a timestamped tracklist and a signed cue sheet that names the audio's SHA-256.Eval and API →Inside 1 Decosa tools
- Site checksDeterministic checks for AI answers: robots.txt for AI crawlers by RFC 9309, llms.txt, JSON-LD property by property, and a review-reply guard that never confirms a patient. Open source, in TypeScript and Python.Eval and API →Inside 2 Decosa tools
Self-host prompts92
Each tool has a copy-paste prompt that tells a coding agent how to set it up on your own machine, with a smoke test.
Every self-host prompt
- Visit copilot
- Sales & meeting copilot
- Private code assistant
- Decosa Studio
- Field reports
- Live translation
- Tamper-evident record
- Provenance and consent
- Endpoint auditor
- Disclosed UGC ads
- Animated characters
- Grounding check
- Deposition and hearing digest
- Receipted hiring screen
- Filing pre-flight
- Promotional-claims pre-check
- Open-model migration check
- Typed-judgment API
- Security questionnaire answerer
- Agent flight recorder
- Verified end-to-end test runs
- Model-risk evidence pack
- Clinical AI assurance monitor
- Privilege review and privilege log
- Editorial-control ledger
- Report integrity
- Privileged drafting editor
- Public-records desk
- Structured oral assessment
- Patent claim-support checker
- Rights-cleared music generation
- Sample and lyric clearance pre-check
- Split-sheet and metadata checker
- Citation and claim checker for papers
- Signed lab notebook
- Green-claims substantiation check
- Auto F&I disclosure record
- Consent ledger and replica gate
- Consented creator dubbing
- Synthetic-performer disclosure and S&P pre-flight
- Script to animatic
- Audio drama and narrated story studio
- Music video from your track
- SAR narrative desk
- Insurance claims-file conduct pack
- Collections and servicing call QA
- Incident notification pack
- Filing tie-out and MD&A grounding
- Sanctions alert disposition record
- Medicare sales-call record
- Claim denial appeal packet
- Made-for-kids content pre-flight
- HCC evidence file and RADV defence
- Signed VEX triage
- Device complaint MDR triage
- CMMC / NIST 800-171 evidence map
- EU trial lay summary with number grounding
- Reg E dispute investigation file
- Disclosed virtual staging
- M&A due-diligence red flags
- Tariff classification memo
- CSR number-to-table verifier
- GPSR listing pack
- Medical chronology with page cites
- No Surprises Act IDR packet and eligibility screen
- Walkthrough-to-quote
- Expert-to-SOP
- Honest product imagery
- Pharmacovigilance intake
- Label consistency across PI, SmPC, CCDS and carton
- No-training receipts
- Storefront accessibility pass
- Fill a form from your papers
- Privileged call notes
- Check their brief
- Check their discovery responses
- Payer audit response
- Prior-auth pre-check and packet
- Notes from your own jottings
- Capture audit evidence from your admin screens
- Injury demand reader
- Certificate request check
- Vendor bank-change check
- Family interview film
- Music video starring you
- Our story film
- Settlement video from the case file
- Interview themes
- Mix cue sheet
- Site readiness for AI answers
- Review reply with patient privacy
- What studies found
Docs
- Getting started →Base URL, a demo session and your first receipted call.
- API reference →The full contract: every route, request and response.
- Receipts and verification →What a receipt contains and how to check it yourself.
- Self-hosting →Run the same stack on your GPU with Docker Compose.
Status: a scheduled check runs the hosted tools end to end and reports pass or fail, with the date; the feed is GET /verify/status. Test results.
More tools
- Run an end-to-end browser testA browser run of your plain-language test, judged only by your assertions in code, ending in a signed certificate that fails the build on a failure.Try it
- Draft VEX for scanner findingsA draft VEX statement for every scanner finding, with the evidence from the image itself, ready for a security engineer to sign off.Try it
Looking for something else? Every tool, as an index.