AI connections · MCP
Connect your AI to your hiring workspace
Included on every plan, at no extra cost.
Point your own AI assistant at one of your workspaces and ask it about your jobs and applicants in plain language, or have it draft a job posting. You bring the client and the subscription; we run no model on your behalf. For now we set up the connection with you — contact us at [email protected].
Four things people ask it on day one
Your AI client talks to ATS Hiring over an ordinary encrypted connection, using a token you create in the app. Your assistant gets nine tools: seven that read your workspace, and two that create or edit a job draft. Everything it can do is bounded by your own role in that workspace — a connection never grants anything you could not already do in the app.
Ask about your pipeline
"How does the backend role stand?" comes back as counts by stage and by status for that one job, drawn from the same rows the pipeline board draws.
Toolget_job_pipeline_summarySearch applicants by stage and score
Structured filters, not free text: one job, a stage, a status, a grade band, a minimum score. "Everyone past screening scoring 80 or better" is one call.
Toolsearch_applicantsDraft a job posting
Nine fields — title, description, location, employment type, seniority, industry, skills, experience, education. It lands as a draft in the job wizard for a person to finish and publish.
Toolcreate_job_draftCatch up on the week
What moved across the whole workspace over a window you choose, up to ninety days, so a Monday morning starts with a summary instead of a scroll.
Toolsummarize_pipeline_activity
The nine tools
That is the whole list. Nothing is held back for a higher plan, and nothing is added by one.
Seven read tools
- get_workspace_summary
- list_jobs
- get_job
- get_job_pipeline_summary
- search_applicants
- get_applicant_summary
- summarize_pipeline_activity
Two job-draft tools
- create_job_draft
- update_job_draft
What the connection cannot do
- It cannot publish a job, unpublish one, or change its status. Publishing stays a human click in the app, and a prompt asking the assistant to publish has no tool to call.
- It cannot move an applicant between stages, change a status, or disqualify anyone.
- It cannot email anyone, and it cannot download or read a CV file.
- It never returns a candidate's email address or phone number — those columns are not even fetched, so no shaping mistake can leak them.
- It cannot see another workspace. A token names one workspace, and a question about anything outside it answers exactly as if that thing never existed.
What each plan gets
The connection itself is not a paid feature and never becomes one. Drafting is the only part a plan changes, and it changes by a daily allowance rather than by a locked door.
Reading your workspace
Every plan
Free, Trial, Pro, Business and Enterprise all carry it, at no extra cost. The seven read tools are not rationed by plan — the published ceilings below are abuse control, and they are the same for everyone.
Creating job drafts
Per workspace, per day
New drafts a day
- Free · Trial · Pro0
- Business25
- Enterprise100
A created draft is metered because an agent loop can write rows faster than a person can read them. The allowance resets every day.
Editing a draft
Uncapped
Only creating one is counted. A cap on edits would punish iteration, which is the whole point of drafting with an assistant, so updates cost nothing on any plan.
So a Free workspace can read everything through its assistant on day one and drafts nothing; a Business workspace adds twenty-five new drafts a day. Nothing on this page is sold to you as an unlock of the connection.
How to connect, in three steps
You create the token in the app; we set up the rest of the connection with you. Nothing changes for anyone else in your workspace.
- 01
Create a connection token
- Open your workspace and go to its AI connections page.
- Create a connection and give it a label naming the machine and the client — "work laptop — Claude Desktop". That label is the only thing that makes a revoke list useful six months later.
- Copy the token. It starts with atsh_mcp_ followed by a long string.
The token is shown once and never again
We store only a one-way hash, so there is no "show me the token" button and support cannot recover it for you. If you lose it, revoke that connection and create a new one — it takes ten seconds.
Three things worth knowing before you paste it anywhere
- One token is one workspace and one person. Two workspaces means two tokens and two connections. There is no company-wide key, deliberately: a shared key cannot be tied to a role, and it keeps working after someone leaves.
- It follows your role, live. If your role changes, the assistant's tools narrow or widen on the very next call. If you are removed from the workspace, the token stops working immediately.
- It expires, and rotation means creating a new one — an existing token is never extended. The connections page shows each token's expiry and when it was last used.
- You can hold up to 5 live connections per workspace and create up to 10 per hour. Revoke one you no longer recognise rather than collecting them.
- 02
Contact us
The program your AI client needs is not available to download yet, so we set up the connection with you — contact us at [email protected].
- 03
We connect it with you
We set up the connection with you, on the machine and AI client you use. You keep the token; we will never ask you to send it to us.
Never put the token on a command line
The token belongs in your AI client's own configuration file and nowhere else. Your shell history is not a secrets store: a token typed at a prompt is written to your history file, is visible to other accounts through a process listing while the command runs, and is the first thing pasted into a support ticket by accident.
ChatGPT
Not supported. ChatGPT's custom connectors accept remote HTTPS MCP endpoints only, not a local stdio bridge, and we do not run a hosted endpoint.
What bounds it
The parts a risk review asks about, before it asks. All three are properties of the code rather than promises about it.
The published ceilings
Code constants, so a change is a reviewed commit in our history rather than a silent dial. Two bound how fast a connection may work; two bound how many connections you may hold.
- 120 calls per minute per member.
- 300 calls per minute per workspace.
- 60 calls per minute per tool, per workspace — so one hot tool cannot starve the other eight.
- 10 new connections per hour per member, and at most 5 live connections per member per workspace.
The verbs that do not exist
The strongest thing we can say about what your assistant will not do is that it has nothing to do it with. There is no publish tool, no status tool, no stage tool, no email tool and no file tool in the list above — so a prompt asking for one of those, whether yours or one smuggled into a job description, finds nothing to call.
Your token, and the log
- The token is a workspace credential. Treat it like a password: it belongs in your client's own configuration and nowhere else. Never paste it into a chat, a ticket, or a shared document.
- The connection program never logs it, never echoes it in an error, and never puts it on a command line.
- We store only a hash. A leaked configuration file is a problem you fix by revoking one connection; your other connections and your browser session are untouched.
- Calls that reach a tool are recorded against your workspace's audit log. Each row carries the tool's name and what happened, and nothing else: a refusal records its refusal code and which ceiling it hit, a failure records that it failed. The arguments you sent are never recorded, and neither is a candidate's identity. Three cases write no row on purpose: a call that fails authentication, a call to a tool that does not exist, and repeat refusals once a rate ceiling has already been recorded for that window.
- Revoke anything you do not recognise, immediately, from the connections page. Revocation is instant and affects only that one connection.
Questions people ask before they set it up
Is it really free?
The connection is, on every plan including Free, at no extra cost — it is not an add-on, not a trial, and not something a salesperson unlocks. What you bring is your own AI client and your own AI subscription; we never run a model on your behalf, and every answer is assembled from what your workspace already stores. The one part a plan changes is how many new job drafts your assistant may create in a day.
Which AI clients work with it?
We set up the connection with you, for the AI client you use — contact us at [email protected]. ChatGPT is not supported: its custom connectors accept remote HTTPS MCP endpoints only, and we do not run a hosted endpoint.
Can it publish a job or reject a candidate?
No, and not because we ask it nicely. There is no tool in the list that publishes, unpublishes, changes a status, moves an applicant between stages, disqualifies anyone, or sends an email. Publishing stays a human click in the app. A prompt asking your assistant to publish — including one hidden in text somebody else wrote — has nothing to call.
What do you log when my assistant makes a call?
Calls that reach a tool are recorded against your workspace's audit log, and each row carries the tool's name and what happened: a refusal records its refusal code and which ceiling it hit, a failure records that it failed. The arguments you sent are never recorded, and neither is a candidate's identity. Three cases write no row on purpose — a call that fails authentication, a call to a tool that does not exist, and repeat refusals once a ceiling has already been recorded for that window.
Can it see candidate email addresses or phone numbers?
No. Contact details are never returned, and the reason is stronger than a filter on the way out: those columns are not fetched at all when applicants are read, so there is no shaping mistake that could leak them. It also cannot download or read a CV file, and it cannot see any workspace other than the one its token names.
It will not connect. What should I check?
Three things, in this order. If your client reports an authentication failure, the token is wrong, revoked or expired — you cannot recover a token, so revoke that connection and create a new one. If it connects but describes jobs you do not recognise, the token belongs to a different workspace; one token names exactly one workspace and one person. And if a call comes back refused rather than failed, read the refusal: it names either a role you do not have in that workspace, or a ceiling you have reached and can wait out.
Ready to connect?
Create your first connection from your workspace, or create a free workspace if you do not have one yet.