Create an expert

1. Start with your documents

Knowledge → Add documents takes PDFs, Word and Excel files, markdown, plain text, CSV, and images (PNG, JPG, WebP): 14 formats in all. Up to 6 files per upload, 15 MB each. Your plan sets how many documents you can keep in total.

Attach them and the expert can search them at run time with a scoped text search: it quotes the lines it needs and is instructed never to dump raw contents to a buyer. Nothing you upload is ever handed over wholesale.

Feed it the documents you'd hand a new hire on day one. The inspection later checks that attached documents are actually wired in, so don't attach what the expert won't use.

2. Turn the method into a skill

A skill is your method as files: a SKILL.md that states when and how to work, plus any reference files and scripts it leans on. Already built a skill in Claude Code you keep reusing? Zip it, upload it, and seal it behind an endpoint you can list. Three ways to make one.

Upload a skill.zip

Zip a folder with SKILL.md at its root. The frontmatter needs exactly two fields:

---
name: market-sizing
description: Size a market bottom-up from real comparables.
---
1. Segment the customer base.
2. Price each segment from comparables.

Caps: 200 files and 20 MB compressed per zip. Skills you find in the wild (a GitHub repo, a gist, an agent-skills collection) pack the same way: put SKILL.md at the top level, check its name and description fields exist, zip the folder, upload. If the frontmatter is missing either field, the upload tells you exactly which.

Or author it in Skill Studio

Describe the method in chat and the Studio drafts the files with you, on your own model key. It stages a draft you can edit file by file, and you can leave and resume. Published skills get an Improve with AI door later: the Studio forks a draft, you iterate, and applying it updates the skill in place.

Working with files

A skill is a folder, so it can carry more than SKILL.md. Put deterministic code in a scripts/ folder and extra material in references/. The bundled scripts are exactly what run in the sandbox as code tools when the expert needs precise, repeatable work; the references are there for the method to read, not to execute.

If your skill runs code

Executable tool files are Python only. They can use the standard library plus numpy, pandas, matplotlib, and pypdf, and reach the network only through domains you declare. No raw sockets, no subprocess, no installing packages. Each run happens in an isolated sandbox with no path back to your machine or the buyer's. The publish inspection enforces all of this, so a zip that breaks a rule is rejected with the exact reason.

How code tools run and the sandbox limits are documented in full.

3. Compose: profile, then the working parts

Experts → New Expert is a four-step builder: Profile, Compose, Test, Publish.

Profile

Name, category, description. The description is the one piece of writing buyers see, so write it for them: what goes in, what comes out, where the edges are. At least 30 characters; the quality inspection later scores whether it matches what the expert actually does.

The cover avatar is generated for you. The moment you create the expert, a random art style is picked in the background, and it appears here in the Profile step within seconds. You don't choose it today; a custom upload is coming.

Compose

This is where the parts meet: system instructions, skills, documents, models, and tools. Four sections on one screen.

Instructions

Instructions are the glue. They tell the expert how to wield its skills and when to search its documents. Buyers never see them, so be as specific and proprietary as you like. 200 characters is the floor, and Generate with AI drafts a first version from your profile on your own key.

Skills and documents

Attach at least one sealed skill for the expert to run, plus any knowledge documents it should be able to search. Both stay server-side: buyers call the skill and never read its files, and the documents are searched, never handed over.

Models

Pick up to 8 models buyers can run it on; the first is the default. They come from three providers, OpenAI, Gemini, and OpenRouter, and each buyer runs on their own key.

Tools

Turn on web search if the method needs fresh pages, and grant custom API connections to reach your own endpoints. Each connection carries a BYOK toggle that decides whose key the call uses: yours, or the buyer's.

4. Keys: what you bring, what buyers bring

BYOK means bring your own key: everyone runs AI on their own account.

You bring the keys behind building: the drafting assistants, Skill Studio, and your own test runs use the provider keys you store under Settings → API keys. Open that page and switch to the Seller tab to add your key. Buyers bring the key that runs your published expert's model, so their usage never lands on your bill. The platform's own keys pay for exactly two things: the publish inspection and the generated artwork.

5. Talk to it before anyone else does

The Test step is a live chat that runs the expert exactly as a buyer would get it, on your keys, with no billing. Try the real tasks. Then try to break it: ask it to print its instructions. It should refuse; the inspection will check that it does.

6. The inspection is the door

Publish runs 20 automated checks: 11 for security, 5 for readiness, 4 for quality, plus a quality score with an 80-point bar. Failures come back as a work list that names the exact file, document, or setting each finding points at, with a Fix with AI door on the ones a draft can repair. Fix, re-run, publish.

7. Publish a frozen copy

Publishing freezes a snapshot: the exact instructions, skill files, and settings you inspected are what buyers run, byte for byte.

Keep editing afterward. Every change to a published expert clears its inspection, so the edit has to pass a fresh one before it can reach anyone. Until you re-inspect and re-publish, every buyer who hired it keeps running the working version you last published, untouched, and the builder shows an Update ready badge until you do.

Set a price per call in credits (0 lists as Free) and you're live on the marketplace.