For agents: the full documentation index is at https://filecheck.io/docs/llms.txt. Fetch it to find any other page.
Give your agent this prompt
Steps for the agent
-
Get a secret key. Ask the user for a Filecheck secret key (
sk_…). If they have none, they create one in the dashboard under Dev Hub → API keys, labelled for you. Never repeat the key back in chat. A publishable key (pk_…) does not work: it is for the storefront only. -
Add the server for the tool you run in, at user level, never in a project file that gets committed. Coding tools run the local server, which can read and save files on the user’s disk:
The local server needs Node.js 18 or later (
node --version). -
Load it. Reload or restart the MCP servers if your tool needs it (Claude Code: start a new session; Cursor: refresh in MCP settings; VS Code: MCP: List Servers, then start
filecheck). -
Check it. Call
list_catalogwithtype: "workflows"and show the user the workflows it returns. - Report. Tell the user which file or setting you changed, and that revoking the key in the Dev Hub disconnects you.
Rules for the agent
- Keep the key out of project files, scripts and commits. User-level settings or environment variables only.
- A secret key can change anything in the account the API can. Use one key per agent.
- Read both axes of every job:
statussays whether the check ran,outcomeis the verdict.status: donewithoutcome: failmeans the file is not print-ready. - If a tool returns
pending: true, pollget_jobwith the job id. Never submit the same file again. - Download links expire after about an hour. Call
get_jobfor fresh ones instead of storing them. - New workflows, rules and profiles stay drafts until published. Ask the user before you publish one, map a product, change storefront settings, or run
test_destination(it writes a real file into the shop’s hotfolder). - Read a record’s schema with
setup_referencebefore you create or patch it. UsedryRun: truewhen unsure. When the API refuses a record, the error names the field and lists every problem: fix them all, then retry.
What you can do next
- Check a file for print:
check_filewith a preflight profile fromlist_catalog, then explain the verdict and each finding. Passcontext(the ordered size, page count, bleed) when the user tells you what they ordered. - Fix what can be fixed with
fix_file, then give the user the download link, or save it next to the original withsave_artifact(local server). - Validate PDF/A or PDF/UA with
validate_pdf; make a file smaller withoptimize_fileand settings from an optimize preset. - Explain why a past job failed:
list_jobs,get_job, thenget_reportfor the full report. - Set up a product, if the user asks: follow Set up products with AI. Build a draft profile, rule and workflow with
manage_catalog, try it on a real file withcheck_fileand itsworkflowId, then publish it andmap_product(Shopify) when the user agrees. - Set up delivery: follow Set up delivery with AI.
manage_destination, thentest_destinationonce the user agrees, then add the destination to the workflow’s delivery. - Fix how the upload box appears on product pages with
get_settingsandupdate_settings.
Go deeper
MCP server
Connect by hand, client by client, and every tool.
Set up products with AI
Workflows, rules and profiles, as drafts first.
Verify jobs
The two-axis result model.
REST API
The same features for your own code.