The MCP server
Still Needs a Human answers as a remote MCP server: you ask an AI assistant to check a file or a text, read the report, make a corrected copy, mark findings or draft a profile from a client's folder. The assistant acts as the person who connected it, with fewer rights: signing a report off and changing a profile stay with a person in the app.
Read the transcript
- Ask your assistant to run a check. Once an admin connects it under Settings.
- 1. Ask it to check a file against a profile. Attach the file to the message.
- It answers with the result and the report link. The report is the same one as in the app.
- 2. Ask for the confirmed fixes. You get a corrected copy. Your original is never changed.
- The run is in Reports like any other.
- Next: Read the report.
Server address: https://app.stillneedsahuman.com/mcp
Set it up in Claude
- In Claude open Settings, Connectors and choose Add custom connector. Name it What's Wrong and paste the server address. Leave the client fields empty: Claude registers itself.
- Press Connect. A Still Needs a Human page opens: sign in, pick the workspace if you belong to several, keep the permissions you want and press Allow.
- Turn the connector on from the tools menu of a chat and ask for a check. On a Team or Enterprise plan an owner adds the connector once; each person then connects and signs in as themselves.
- The same address works as a remote server in Claude Desktop and Claude Code. In a terminal the assistant can send a file to the upload link with curl.
Set it up in ChatGPT
- In ChatGPT open Settings, Apps and Connectors, Advanced settings and turn on Developer mode. On a business plan a workspace admin may have to allow it first.
- Choose Create. Name it What's Wrong, paste the server address and pick OAuth. ChatGPT registers itself, so no client id is needed.
- Sign in on the Still Needs a Human page and press Allow. In a chat pick the connector from the tools menu and ask for a check.
Sign-in and permissions
You sign in to Still Needs a Human itself, never to the assistant: OAuth 2.1 with a consent page that shows one box per permission, or a personal token from Settings, Connections, Assistants for your own agent. Access lasts an hour at a time and renews on its own for 30 days; you can disconnect an assistant in Settings at any time.
| Permission | What it allows | On the consent page |
|---|---|---|
| Read | Read your reports and profiles | ticked |
| Run | Run checks, which count against the workspace's AI allowance | ticked |
| Review | Mark findings Ignored or Fixed on your behalf, never sign a report off | not ticked |
| Build | Propose a profile from your files; a person approves it before anything changes | not ticked |
Which plans include it
The connector is on every plan, Free and the trials included; what it may do grows with the plan.
| Plan | Over the connector |
|---|---|
| Free | Text checks and every reading tool |
| Solo | Plus the file tools: upload links, file checks, corrected copies and rechecks |
| Team | Plus Dropbox files, marking findings and profile drafts |
| Business, Enterprise | Everything |
A trial has the tools of its plan. Below the plan a tool is absent from the list and a call answers with the plan's sentence. The profile builder over the connector is switched on per workspace, on request.
The tools
| Tool | What it does |
|---|---|
check_text | Checks a source text and its translation (up to 3,000 words together) with the workspace's profile; answers the result, the score, the counts, up to 10 findings and the links. |
list_profiles | The profiles this workspace can run, with languages, account and what is checked, in words. |
describe_profile | What one profile checks in plain words. A loose name works. |
list_profile_terms | One profile's glossary, brands and terms to keep and Fine as is, paged. A PM or above. |
list_reports | Your reports, newest first, with result, score and link; filters by profile, state, result, date and text. |
get_report | One report: result, score, counts, open counts and the download links, with up to 50 findings. |
list_findings | Every finding of a report, paged, with filters by severity, status, origin, file and segment range. |
get_segment | One segment by its id or number: source, target, lock, match and every finding in it. |
get_run_status | A run's state and progress; with a wait it answers as soon as the run is done. |
get_usage | The month's usage: AI words used and left, the reset date. |
request_upload | A one-time page to drop a file on, with the link, how long it works and a curl line. No cost until the check. |
get_upload | The state of an upload link: pending, received, used or expired. |
check_file | Checks one file from an upload link, a Dropbox path or inline content; answers the report. |
check_files | Up to five files in one call, each its own run or all as one batch. |
list_dropbox | The folders and files inside the Dropbox folders this workspace watches. |
apply_fixes | Writes the report's confirmed fixes into a copy of the file, never the original, with the change log and a one-hour download link; a dry run shows the log only. It writes the XLIFF family, Excel, CSV and TSV, software string files (a gettext .po, or a zip with the source and the translated string file) and packages of them. |
recheck | Runs a file you fixed yourself with the report's profile and answers what was resolved, what remains and what is new. |
set_finding_status | Marks findings Ignored, Fixed or Open again with your reason. Never signs a report off. (Review permission) |
scan_folder | Reads the zip of a client's folder from an upload link: each file's kind and languages, with the questions to ask. Writes nothing. (Build) |
propose_profile | Makes a profile draft from that folder for an account. A new proposal replaces your own open draft for the same account and says so. (Build) |
get_profile_draft | The draft in full: the files, per profile the exact lines it would write, the answers so far. (Build) |
answer_profile_draft | Your answers to the draft's questions; the draft is proposed again. (Build) |
apply_profile_draft | Builds the draft's profiles once you have said yes to the summary. (Build) |
discard_profile_draft | Discards an open draft once you have said yes. Nothing in a profile changes. (Build) |
The tool list an assistant sees holds only the tools your permissions, your workspace's settings and your plan allow. Every answer carries the link to the report.
The server also offers three prompt templates: check_and_fix (check the attached file, wait for the AI review, list the confirmed findings, ask before applying, then apply with a recheck), lqa_summary (five lines at most on one report) and batch_status (today's reports grouped running, failed, open criticals and clean).
Six things to ask
Type one line. The assistant picks the tools; you read the answer and open the report link.
- A file in the chat. "Check this file with the Northwind support profile." The assistant sends the file it has in hand to your workspace and runs the same check as the app, with your workspace's profile. Back come the result, the counts and the link to the report.
- A batch and the bill. "Check these five files from the Contoso drop with one report each and tell me what the month has cost so far." Up to five files go in one request, each as its own run or all as one batch; the assistant lists each report with its result and link. It then answers with the AI words used and left this month and the day they reset.
- The findings and the AI's verdicts. "Which findings in the Acme report did the AI agree with and which did it dispute?" The assistant reads the report's findings with the AI's verdict on each one. You get them by segment, with the source and the target beside each finding.
- Accept, fix a copy, recheck. "Mark the two glossary findings the reviewer accepted as Fixed, reason: corrected in Trados, then apply the confirmed fixes to a copy and recheck it." With the Review permission ticked, the assistant marks those findings with your reason; the app shows each mark as made through the assistant, under your name. It then writes the confirmed fixes into a copy of the file, never the original, hands you a one-hour download link with the change log and says what the recheck resolved, what remains and what is new.
- A folder on Dropbox. "What is new in the Contoso folder on Dropbox? Check the German file." The assistant lists the folders and files inside the Dropbox folders your workspace watches and runs the file you name from there. Nothing passes through the chat; the report link comes back as always.
- A profile from a client's folder. "Build a profile for Northwind from this zip of their folder." The assistant reads the zip, says what each file is (a style guide, a glossary, a checklist, bilingual files) and proposes a profile draft with the questions it still has. Nothing is written until you approve: you answer, read the summary and say yes; only then are the profiles built.
A file too big for the chat goes through a one-time upload link: say "I have a 2 MB package.", drop the file on the link the assistant gives you, say done and ask for the check.
Limits
| Text | 3,000 words | Source and target together, for a check of pasted text. |
| Pasted file | About 150 KB | A small text XLIFF can go in the chat. Anything bigger goes through an upload link. |
| Upload link | 30 minutes, once | One file of up to 50 MB per link, checked within two hours of the upload. |
| Several files | Five a call | A zip or a package counts as one file. An i18n .json, .resx, .strings, .stringsdict or .properties file goes in a zip with its source file; a .po file goes on its own. |
| Links | One hour | Report downloads and corrected copies. Ask again for fresh links. |
| Pace | 60 calls a minute | Per person. Checks, uploads and corrected copies have lower limits of their own. |
| AI review | Waits, then says so | When the AI review takes longer, the assistant says the result is not final yet and asks again. |
| Sign-off | In the app only | No assistant can sign a report off. |
Privacy
A file sent from an assistant is handled as any other upload and removed within 24 hours. A workspace can agree a longer keep for its uploaded files with us; we set it per workspace. The assistant receives what the tools answer: results, findings with the segment text they are about and links. The text you check goes through the same rules and AI review as in the app, on the AI providers your workspace allows. Reports stay in your workspace until an admin deletes them. The connector never signs a report off, never changes your original file and never changes a profile without a draft a person approved. Security and privacy.
Disconnect
In the app, Settings lists every connected assistant with who connected it and when it was last used. Disconnect ends it: its next request is refused and the reports it created stay.
Support
Something does not work or you want it on for your workspace? Tell us the time, the assistant and what you asked. Never send a token or a link from Settings.