LLM and MCP integration: Coodesh inside your AI stack

MCP (Model Context Protocol) is an open standard that lets an AI assistant, such as Claude or ChatGPT, securely access tools and data from other systems, with the user's explicit authorization. Connecting Coodesh via MCP means your assistant can see your workspace: jobs, applications, assessments and results. Instead of copying names and scores into a chat, you ask the assistant directly "who are the top five in this week's backend assessment?" or "invite the candidates who passed to the AI interview", and it looks things up or takes action in Coodesh on your behalf.

The assistant has access to read and write actions, listed separately on the authorization screen. For reading, it looks up jobs, applications, assessments, results and reports. For writing, it does what you would already do manually, such as creating assessment invitations and triggering AI interviews. The principle is simple: the assistant never does anything you couldn't do with your own login. It speeds up the recruiter's work; it doesn't create a user with superpowers.

Any client that implements the MCP standard with OAuth authentication. Today that includes Anthropic's Claude Desktop and Claude Code, whose flows are documented step by step in the Help Center, as well as in-house agents built with frameworks that support MCP. Because the protocol is open, the same server works with other assistants as they adopt the standard, without Coodesh having to build a separate integration for each vendor. If your company uses its own agent, the connection is the same: server address, browser authorization, done.

In five steps, with nothing to install on the Coodesh side. In your assistant, add an MCP server with the address provided in the Help Center. The assistant opens the browser on the Coodesh authorization screen. You log in, choose which workspaces the assistant can access (it can be more than one) and click Authorize. The connection is active right away. To reconnect later, the screen opens with the workspaces already selected and you adjust whatever you want.

Authorization follows OAuth and inherits your permissions, never expanding them. If you can't see talent data in a workspace, neither can the assistant. Access applies only to the workspaces you selected. In My account, under the Security tab, the Connected assistants card lists each authorized app, its workspaces, permissions and recent activity. Disconnecting takes one click, and access is cut off immediately. For teams, we recommend that each person connect their own assistant with their own login instead of sharing a connection, so that the action history is attributed to the right person.

Yes, what the assistant reads from Coodesh passes through the model you chose, and that's a decision your company makes as data controller. Coodesh doesn't send data to any model on its own via MCP; it responds to the calls your assistant makes, within the permissions you authorized. What happens to the data at the LLM provider depends on your company's contract with them: enterprise plans from Anthropic and OpenAI usually offer zero retention and don't use content for training, but that needs to be in your contract, not ours. Before connecting, it's worth aligning with your DPO on which assistant is approved, which account should be used and which workspaces can be exposed.

It means Coodesh's AI features, such as automated criteria-based review and the conversational interview, are built on a routing layer that doesn't depend on a single model vendor. For enterprise accounts with data residency or approved-vendor requirements, Coodesh can run these features with the model and infrastructure defined by the company, keeping candidate data within the AI policy it already has. This avoids lock-in and resolves the most common information security objection: "our data can't go through a model we haven't approved". Setup is done with the Coodesh technical team, case by case.

No. Candidate data, answers, transcripts and results in your workspace are used to operate the contracted service, not to train language models. The platform's AI features work through instructions and criteria defined at the time of use, and your company's content doesn't go into the training data of any model, whether Coodesh's or its vendors'. If your contract or security questionnaire requires this statement in writing, the Coodesh team will provide it.

The REST API is for system integrations: your ATS, data lake or internal portal calls Coodesh programmatically with an API key and receives webhooks when something happens. MCP is for people working with an assistant: the recruiter chats, and the assistant decides which calls to make and executes them with that recruiter's permissions. In practice, many companies use both: the API to sync candidates with the ATS and MCP for the recruiting team to handle day-to-day work in natural language. Both share the same permission model and the same activity log.

The ones that save the most time are repetitive tasks and those that require cross-referencing information. Common examples: asking the assistant for a summary of an assessment's results with each candidate's highlights; inviting everyone who passed a technical test to the AI interview in one go; comparing two candidates based on their transcripts; generating a feedback message from the report of someone who didn't advance; and answering manager questions like "how many data candidates do we have in screening right now?" without opening the platform. The gain isn't replacing the recruiter; it's taking lookup and copy-paste tasks off their plate.