MCP

The Kolena MCP server connects AI assistants such as Claude and ChatGPT to your Organization through the Model Context Protocol (MCP). Once connected, you can ask the assistant in plain language to find Agents, read results, edit Prompts, and process documents, and it does so on your behalf.

This feature is in beta and is enabled per Organization. Contact Kolena if you’d like it turned on for your Organization.

Enabling MCP Access

MCP access is off by default. Organization admins can turn it on from the MCP Server tab on your Organization page. The same tab shows the server URL and the setup steps for each supported assistant.

Installing the Kolena Plugin

The Kolena plugin is the recommended way to connect. Along with the connection, it gives the assistant guidance for processing documents and designing Prompts. The steps below are a quick reference. Each assistant’s own documentation, linked in each tab, has the current and complete instructions.

ChatGPT & Codex
Claude

Search for Kolena in the plugin directory, or open the Kolena plugin.
Install it, and sign in to Kolena when asked.

If your workspace restricts plugins, ask a workspace admin to allow Kolena under Workspace settings › Plugins.See Plugins in ChatGPT and Codex.

If you connected Kolena manually before, remove that connection after installing the plugin so the assistant doesn’t have two Kolena connections.

Manual Setup

Other MCP clients, and assistants where plugins aren’t available, can connect to the server directly:

https://agents-api.kolena.com/mcp

The steps below are a quick reference. Each assistant’s own documentation, linked in each tab, has the current and complete setup instructions.

Claude Code
Claude Desktop & claude.ai

claude mcp add --transport http kolena https://agents-api.kolena.com/mcp

See Connect Claude Code to tools via MCP.

On first connection, the assistant opens a browser window to sign in to Kolena. No client ID or secret is needed. Assistants and scripts that send their own credentials can use an API Key as a bearer token instead of signing in. The assistant then acts as the Key’s user or service user.

Network Access

Uploading and downloading documents is done by the assistant itself, as ordinary HTTPS requests to agents-api.kolena.com, the same host as the server URL above. This is a separate permission from the MCP connection. Assistants that run code in a sandbox apply a network policy to it that usually does not cover the MCP connection, so the connector can be working — listing Agents, reading Runs — while uploads and downloads fail. If your assistant restricts outbound network access, allow agents-api.kolena.com. Where that setting lives, and whether you can change it, depends on the assistant and your plan with it. Check the assistant’s own network or sandbox settings. If everything works except uploading and downloading documents, check network policy configuration first.

What You Can Do

Each response links to the Agent, Prompt, or Run it refers to in Kolena.

Entity	What You Can Do
Workspaces	See which Workspaces you have access to, as a starting point for finding Agents.
Folders	See how Agents are grouped, set up new Folders, rename or reorganize existing ones, and move Agents between them.
Agents	Find an Agent and see exactly how it is set up: its Agent Guidelines and every Prompt’s instruction and settings. Create a new Agent, or change an existing one’s name, objective, or Folder.
Prompts	Add Prompts of any output type or control step to an Agent, reword an instruction, change a Prompt’s settings or when it runs, or remove a Prompt you no longer need.
Runs	Send new documents to an Agent, follow a Run until it finishes, and read its results. Find past Runs by document name, who submitted them, date, or by what the Agent found, such as “Risk Level is High” or “Total between 10,000 and 50,000”. Re-run some or all Prompts on existing Runs, or get a link to download a Run’s documents.
Usage	See how many credits were used each month, split by Workspace and by quota, prepaid, or overage, and check prepaid balances. Drill into the individual charges behind any figure, by Agent, Run, or user.

Examples

Task	Example Request	What Happens
Find exceptions	”In the Invoice Review Agent, show Runs from last week where Approved is False, with vendor and total.”	Finds matching Runs and lists their results, with links
Summarize results	”Summarize this month’s lease abstracts: count by property type and the five largest annual rents.”	Reads results across Runs and summarizes them in the chat
Process documents	”Run these 12 PDFs through the Claims Intake Agent.”	Uploads the files as one Run, waits for processing, then shows results
Review Prompts	”Review the Prompts in the Contract Review Agent and flag unclear instructions.”	Reads the Agent’s setup and comments on each Prompt; changes nothing
Fix and re-run	”Make the Effective Date Prompt return dates as YYYY-MM-DD, then re-run it on the 20 most recent Runs.”	Edits that one Prompt and re-runs only it on those Runs; uses credits
Set up an Agent	”Create a Purchase Order Agent in the Finance Workspace with Prompts for PO number, vendor, line items as a table, and total.”	Creates the Agent and its Prompts
Organize	”Group the Agents in the Operations Workspace into Folders by department.”	Creates Folders and moves Agents into them
Check spend	”How many credits did each Workspace use in September, and which Agent used the most?”	Reports monthly usage by Workspace and the charges behind the top Agent

For changes or re-runs across many Agents or Runs, ask the assistant to list what it will do and confirm before it starts.

Permissions and Credits

Same access rules as the platform. The assistant runs with your permissions, including permissions set by fine-grained access control. It cannot see or change anything you can’t.
Credits work as in the web app. Creating a Run and re-running Prompts consume credits; reading Agents, Runs, results, and usage does not. Usage is attributed to you and appears in credit usage with the Agent and Run it belongs to.
Actions are labeled. Kolena tells the assistant which actions only read, which change data, and which cannot be undone. Many assistants ask you to approve changes before making them; this depends on the assistant and your settings.
One sign-in per person. Each person should connect with their own Kolena login so that permissions and credit usage are attributed correctly.
Tool calls are logged. On the Enterprise plan, every tool call an assistant makes is recorded in MCP access logs, which admins can review.
Your assistant’s data policies apply. Anything the assistant reads from Kolena, such as documents and results, becomes part of your conversation with it and is handled under your assistant provider’s terms. If your organization has data privacy or retention requirements, check that the assistant account you use meets them.

What the MCP Server Can’t Do

The following are available in the web app only:

Ground Truth and Validations
Templated Outputs
Adding or changing files attached to a Prompt
Editing Agent Guidelines (they can be read)
Deleting Agents or Runs

Uploading and downloading documents requires an assistant that can send and fetch files itself, such as Claude Code or Codex. Assistants that cannot will link you to the Agent or Run in Kolena instead.

Get Started

Managing Your Organization

FAQ

How-To Guides

Enabling MCP Access

Installing the Kolena Plugin

Manual Setup

Network Access

What You Can Do

Examples

Permissions and Credits

What the MCP Server Can’t Do

​Enabling MCP Access

​Installing the Kolena Plugin

​Manual Setup

​Network Access

​What You Can Do

​Examples

​Permissions and Credits

​What the MCP Server Can’t Do

Enabling MCP Access

Installing the Kolena Plugin

Manual Setup

Network Access

What You Can Do

Examples

Permissions and Credits

What the MCP Server Can’t Do