View Categories

MCP Server

Table of Contents

The PCR MCP is a secure “translator” that sits between your AI assistant (such as Claude) and your PCRM database. It enables your AI to query your candidates, jobs, placements, and other PCR data directly. It also allows your AI to execute on tasks, bringing entirely custom automation, reporting, and processing capabilities to your recruiting workflow.

Connecting the MCP #

The PCR MCP is typically available with upper-tier PCRM contracts. Before connecting, verify that your PCR account has access.

The method of connection will vary from one AI agent to another, and across different versions or interfaces. Check your AI tool’s documentation for current steps.

Claude Desktop #

  1. Choose Settings
  2. In the settings panel, select Connectors 
  3. Use the ‘Add’ button and choose Add Custom connector
  4. Click Add
  5. Click Connect when prompted to make a connection with your database. The ‘Connect to PCR’ window will appear.
  6. Paste in your PCRM account login URL.
  7. Select the database and log in as normal.

ChatGPT Desktop #

  1. Choose Settings
  2. In the settings panel, select Plugins
  3. Click Browse Plugins
  4. Use the ‘+’ button at the top right to create a New Plugin
  5. Click Create
  6. Click Connect when prompted to make a connection with your database. The ‘Connect to PCR’ window will appear.
  7. Paste in your PCRM account login URL.
  8. Select the database and log in as normal.

If you want to connect the MCP to multiple databases, you will need to complete the above procedure for each. Give each a separate name AND replace the ‘www2’ in the URL with a unique value for that data source. For example, ‘https://abcrecruiting-sales.pcrecruiter.net/mcp’ and ‘https://abcrecruiting-healthcare.pcrecruiter.net/mcp’. These links point to the same server, but using a unique subdomain causes your AI assistant to view it as an independent source.

Prompting Tips #

Creating a successful prompt typically relies on:

  • Name the record type(s) (candidate / Company / job)
  • Describing the filter (who or what, specifically)
  • Say what you want back (which details, how sorted, what format)
  • Being explicit about whether you want the AI to change something or just look

Beyond that, it’s all normal conversation. If a request comes back not quite right, just refine it in your next message (e.g. “narrower than that,” “only the ones in Ohio,” “sort by salary instead”).

Ask for Specific Records #

The assistant is at its best when you give it something concrete to search for or filter on. The more precisely you describe who or what you want, the better the results. Think in terms of narrowing down:

  • Weak: “Show me some candidates.”
  • Better: “Find candidates in Ohio with ‘nurse’ in their title.”
  • Strong: “Find candidates whose résumé mentions RN or BSN, located within 30 miles of zip 44094, and show me their name, email, and mobile number, newest first.”

Precision when describing your intended record target is also important:

  • Weak: “Find clients with open jobs.”
  • Better: “Find people with Manager Status who are the contact for an open job.”
  • Strong: “Find people with Manager Status who are the contact for Available jobs and list their name and email address along with the title of their associated positions.”

A few kinds of filters that work especially well:

  • By detail: name, title, location, salary range, or when they were added (“candidates entered since January”).
  • By skills: you can search the full text of résumés (“mentions Python or AWS”) or the keywords or tags your team has applied. Searching résumé text is the most thorough.
  • Combining conditions: “Sales titles OR marketing titles, in Texas, added this year” is completely fine. Mix and match freely.

Ask For Specific Output #

By default you’ll get a sensible summary, but you can steer it. Tell the assistant which fields matter and how to order things:

  • “Give me first name, last name, and email, sorted by most recent activity, 25 at a time.”

This keeps results clean and readable instead of too little info or a wall of text containing every field.

You can also tell the assistant exactly what format you want the answers in. This can be as simple as “return the results as a CSV” or more complex like:

  • “Give me a list of candidates placed since November, sorted alphabetically and grouped under the job title they were placed for. Output this as a PDF using Calibri font and use my brand colors from this uploaded logo.”

Drill Down #

Once you’ve found someone or something, that result is held in the context of the conversation, so you can query further based on the result you’ve already got.

  • “Pull up everything on this candidate including work history, education, notes and a summary of activity this week.”
  • “What’s the full description and salary range for this job?”
  • “Show me the recent activity and notes on this client and cross-reference with my Gmail.”

Making Changes #

The MCP cannot Delete. This means you can’t remove records, take records off of lists, merge records, remove them from pipelines, etc. You can ask the MCP to add custom fields or keywords to records that make them easy to find for a user to delete later, but the MCP will never take action to remove any data on its own.

You can, however, ask the assistant to change things. Because these edits touch your live data, it’s still worth being clear and deliberate:

  • Adding: “Create a new candidate record under Talent Pool for Jane Doe with this email and phone number.”
  • Updating: “Update this candidate’s title to Senior Analyst and add their new mobile number.”
  • Moving: “Find all of the jobs that Lance Garrison is associated with and move him to the Withdrawn status in the Out of Process step on all of them.”
  • Cleaning: “In this uploaded CSV, make sure all phone numbers and email addresses are correctly formatted, then attempt to identify names on the list which are not already in my database.”
  • Adding context: “Add an Activity to this candidate using the current date/time indicating that Jake spoke with them this morning about their CV.”

A good rule of thumb: phrasing like “show me” or “find” is a safe read, while “create” and “update” changes real data. It’s smart to have the assistant confirm the details with you before it acts by saying “show me what you’re about to do first.”

Caveats & Tips #

  • We do not recommend using AI to make qualitative judgements, as this can result in not only potentially false outcomes, but can also raise legal and ethical concerns. Avoid asking the AI things like “which candidate is the best match for this job” and instead ask “find candidates whose resumes contain phrases that align well with this job description and summarize your reasoning for each selection.”
  • As with any AI-enabled process, the way your prompts are interpreted and executed will vary based on not only which AI tool you’re using, but which model you have selected within that tool. Experimenting with different models can help you identify methods that get you the best results.
  • Double-check the work in PCRM rather than assuming that your assistant has done exactly what you have asked. It may say with confidence that it has performed the task, but if it made any assumptions it may have performed a different task than intended.
  • Be mindful of memory use. The context window is an AI’s “working memory”. It dictates the maximum amount of information (input text, conversation history, instructions, and the generated output) a model can process at once. Once this limit is reached, older information is typically dropped to make room for new inputs. Passing a large amount of data, such as resumes or import files, will fill up the context window more quickly and lead to poorer results over the length of the conversation.
  • Whenever possible, ask your AI assistant to build a utility that completes the task at hand, rather than directly executing commands itself. By creating a utility, you will reduce token use and also have a reusable non-AI based method of repeating the same task.
  • No not pass secure data such as passwords and other critical information through your AI tool’s chat.
  • ZIP Radius Search is a feature of PCR’s internal search features and isn’t exposed via the API or MCP. The AI can approximate a radius search based on ZIP prefixes. For example, if you ask for records “within 20 miles of Cleveland” the AI may look for people with ZIP codes starting with 441 (the standard Cleveland-metro prefix, covering downtown plus inner-ring suburbs). These results will differ from the latitude/longitude-based radius search inside of PCR.

Troubleshooting #

  • If the assistant is trying to reuse an expired connection, you may need to revoke the access token inside of PCR, thereby forcing a new one.
    1. Click the ‘Settings’ gear (System menu).
    2. Search for ‘API’
    3. Select ‘API/OAuth & API Access’
    4. Locate the App Name and User combination that represents your connection.
    5. Choose ‘Revoke’ from the Action dropdown.
  • Some information about the connection may be store in local cookies or storage. If you are having a connection issue, you may be able to reset it by logging into your AI assistant’s website interface in an incognito browser window and then removing and recreating the connection.