Sentl.ai Go to the app →

Connect Sentl to your AI assistant

How to connect Sentl MCP to an AI client like Claude or ChatGPT, with a step-by-step guide. This page is public: you can share it or paste it into your AI assistant.

Essential connection details

Server name
Sentl MCP
Connection type
HTTP / Remote MCP / Streamable HTTP
Authentication
OAuth login or Bearer token
MCP endpoint (same for both methods)
https://app.sentl.ai/mcp
Authentication header (token method only)
Authorization: Bearer <TOKEN_SENTL>
With the login (OAuth) method no header is needed: usually you just paste the URL and the client takes care of access.

Are you reading this page as an AI assistant? To connect, add a remote HTTP MCP server at https://app.sentl.ai/mcp. If you support OAuth no token is needed: the user will complete login and consent in the browser. Otherwise ask the user for a Sentl token and pass it in the Authorization: Bearer header. Either way the connection sees all of the user's organizations: use list-organizations to find out which ones and with which role.

A compact Markdown version of this guide is available at https://app.sentl.ai/mcp/setup.md.

What MCP is and what it's for

In short: a secure bridge between your AI assistant and your Sentl data.

MCP (Model Context Protocol) is the standard that lets an AI assistant (for example Claude or ChatGPT) use Sentl: read your projects and reputation analyses and, if your role allows it, start runs or update projects.

Instead of opening every screen by hand, you ask the assistant in natural language («summarize the latest analysis for project X») and it contacts Sentl on your behalf, through tools with well-defined permissions. The assistant never gets your password: it only gets limited access, tied to your role in each of your organizations, which you can revoke at any time.

Sentl exposes a remote MCP server via URL: it works with AI clients that support remote HTTP MCP servers, with nothing to install on your computer.

The assistant can browse projects, analyses and results across all the organizations you belong to and, where your role allows it, start runs, update projects or create new ones. If you work in several organizations you don't need to tell it which one you're in: when you mention a project or a run it figures it out on its own.

New projects are created paused: saving does not automatically start a run. To begin, ask for a manual start or activate the project for scheduled runs.

Before you start

Check these three things before configuring the AI client.

  1. Sign in to Sentl with your account.
  2. Check your role: where you are Owner or Admin, the assistant will also be able to write.
  3. Open the AI app and look for a section called MCP, Connectors, Tools, Integrations or Custom MCP server.
Sentl exposes a remote MCP server via URL. If the AI client only asks for a local stdio command, that is not the right setup for this connection.

Two ways to connect

You can connect Sentl in two equally supported ways. Pick the one your AI client supports.

Login (OAuth)
No token to paste. Add the server URL, you get redirected to Sentl and sign in. Ideal for clients that support OAuth (e.g. Claude, ChatGPT). Revoke it from the Connected apps page.
Token (Bearer)
You create a token in API Tokens and hand it to the client securely. Suited to clients that don't support OAuth. Revoke it from the API Tokens page.
Either way the connection applies to all your organizations, and in each one it can only do what your role there allows. The only difference is how the client proves your identity.

Connecting via login (OAuth)

The simplest way if your client supports it: no token to copy.

  1. In the AI client, add a new MCP server / connector.
  2. Enter Sentl MCP as the name.
  3. Enter https://app.sentl.ai/mcp as the URL.
  4. Save or enable the server: the client redirects you to Sentl (a sign-in page opens in the browser).
  5. If you don't already have an active session, sign in to Sentl.
  6. On the consent screen check the organizations listed and confirm the authorization.
  7. You're taken back to the client: the connection is active, without pasting any token.
Sentl consent screen: the access request title, the requested mcp:use permission, the list of organizations the connection will have access to, and the Authorize and Deny buttons.
The consent screen from step 6. Below the requested permissions you'll find the list of organizations the connection will have access to, with your role in each: that's what the AI assistant will be able to see.
A single connection covers all your organizations: the consent screen lists them before you authorize. If you later join a new organization, the connection sees it without you having to authorize again.

From the terminal, with one command

If your client has a command line, connecting takes a single command. Copy the one for your client and paste it into the terminal. There's no command that works for all of them: each client has its own syntax.

Claude Code
claude mcp add --transport http sentl-mcp https://app.sentl.ai/mcp
Codex CLI
codex mcp add sentl-mcp --url https://app.sentl.ai/mcp
Gemini CLI
gemini mcp add --transport http sentl-mcp https://app.sentl.ai/mcp
VS Code
code --add-mcp '{"name":"sentl-mcp","type":"http","url":"https://app.sentl.ai/mcp"}'

The command registers the server, not the connection: authorization happens on first use. In Claude Code type /mcp and choose sentl-mcp; in other clients the first request to the assistant is enough. That's where the login and consent described above start.

If the name is already in use, first remove the previous connection — for example claude mcp remove sentl-mcp — and try again.

For Cursor, Windsurf, Claude Desktop and other clients without a command line, use the JSON configuration further down or the add-server dialog described in Configure the AI client.
OAuth login is enabled for the official clients (e.g. Claude, ChatGPT). A client pointing to an unexpected domain can't complete self-registration: in that case use the token method below.

Connecting via token (Bearer)

The token is the personal key that lets your AI assistant access Sentl, for clients without OAuth.

The token is personal: it works across all your organizations, with your role in each one.

  1. Sign in to Sentl.
  2. Open Integrations → API Tokens from the user menu (top right).
  3. Click Create token.
  4. Give it a recognizable name, for example Claude Desktop, ChatGPT or Cursor.
  5. Set an expiry date if you want to limit how long the token lasts.
  6. Click Create and copy the token shown on screen right away.
The token is shown only once. If you close the window without copying it, revoke it and create a new one.

Configure the AI client

Screens vary from client to client, but the steps are almost always similar.

  1. Open the AI client's settings.
  2. Look for MCP, Connectors, Tools or Integrations.
  3. Choose Add server, Add connector or Add custom MCP server.
  4. Enter Sentl MCP as the name.
  5. Enter https://app.sentl.ai/mcp as the URL.
  6. If asked, choose HTTP, Remote MCP or Streamable HTTP.
  7. If the client supports OAuth: save and let it redirect you to Sentl for login and consent (see Connecting via login).
  8. If you use the token: choose Bearer token and pass the Sentl token; if the client asks for headers manually, add Authorization: Bearer <TOKEN_SENTL>.
  9. Save the configuration and enable the MCP server.
Some clients ask for confirmation before using an MCP tool. Only confirm when you want the assistant to read or use your Sentl data.
WARNING: NEVER paste the token into the AI client's chat if it asks for it; always ask for a secure way to pass the token.

JSON configuration, if the client requires it

Some clients don't have a guided screen and require a configuration in JSON format.

With OAuth login the URL is enough: no header. On first use the client opens the browser for login and consent.

OAuth login configuration
{
    "mcpServers": {
        "sentl-mcp": {
            "type": "http",
            "url": "https://app.sentl.ai/mcp"
        }
    }
}
The configurations below are for the token method: they pass the Sentl token in the Authorization header.
Basic configuration
{
    "mcpServers": {
        "sentl-mcp": {
        "url": "https://app.sentl.ai/mcp",
        "headers": {
                "Authorization": "Bearer <TOKEN_SENTL>"
            }
        }
    }
}

If the client explicitly asks for the server type, use an HTTP value. The exact key may vary, but it's usually something like type or transport.

Configuration with HTTP type
{
    "mcpServers": {
        "sentl-mcp": {
        "type": "http",
        "url": "https://app.sentl.ai/mcp",
        "headers": {
                "Authorization": "Bearer <TOKEN_SENTL>"
            }
        }
    }
}
Configuration with an environment variable
{
    "mcpServers": {
        "sentl-mcp": {
        "url": "https://app.sentl.ai/mcp",
        "headers": {
                "Authorization": "Bearer ${SENTL_MCP_TOKEN}"
            }
        }
    }
}

Check that it works

After saving the configuration, open a new chat in the AI client and type:

Test prompt
Use Sentl MCP and show me the available projects.

If the connection works, the assistant should contact Sentl and show you the projects in your organizations. If you don't see any projects, check that you're a member of an organization that has some.

What the AI assistant can do

With Sentl MCP the AI assistant can help you work on reputation data without opening every Sentl screen by hand.

  • Read example

    List my Sentl projects and tell me which ones have the most recent analyses.

  • Analysis example

    Open the project with the lowest reputation and summarize the main issues raised in the latest analyses.

  • Comparison example

    Compare the AI models' responses in the latest run and tell me where the brand is described worst.

  • Write example

    Start a new analysis on project 123.

What it can read (all roles, including User)

  • List the available AI research models.
  • List the organizations you have access to, with your role in each.
  • Show an organization's credits: available, balance and reserved by runs in progress.
  • List the projects across all your organizations with their latest analysis.
  • Show a project's details: dimensions (archived ones included), questions and grading rubrics.
  • Get the link to edit a project in Sentl.
  • Estimate what an analysis on a project would cost in credits, and whether it can start now.
  • List a project's analyses (runs) with score, dates and the reason when one went wrong.
  • Read an analysis's results: model responses and judgments/sentiment.

What it can change (Owner and Admin only)

  • Start a new manual analysis on a project.
  • Create or update a project: name, frequency, rounds, models, activation (new projects start paused).
  • Update a project's questions and evaluation dimensions, and archive or restore a dimension.
Rounds repeat every question within the same run, to show where the models' answers change. The run's cost is multiplied by the number of rounds.
If you're a User in all your organizations, the write tools aren't even available: the assistant only sees read operations. If you're Owner or Admin in at least one, the assistant sees them, but where you're a User it gets a refusal.

Permissions and roles

The connection inherits your role's permissions, organization by organization.

User

Can read projects, analyses, responses, questions and grading rubrics.

Can't start new analyses or edit projects.

Owner and Admin

Can read, start new analyses, create or update projects and edit questions.

Permissions follow your current role, checked on every request: if your role changes (for example from Admin to User), the connection's behavior changes immediately, with no need to reconnect.
The connection reaches all the organizations you belong to, with the role you have in each: it can therefore start runs where you're an Admin and stay read-only where you're a regular user. The assistant doesn't need to ask you for the organization when it talks about a project or a run: it derives it from them. You only need to tell it when it creates a new project.

Connected apps (OAuth connections)

This is where you see and revoke AI apps connected via OAuth login.

Open Integrations → Connected apps from the user menu (top right): you'll find the list of apps (e.g. Claude, ChatGPT) you've authorized via OAuth login. The list belongs to you, not to the organization: you see the same apps from whichever organization you're in.

Click Revoke to cut off an app's access: all its tokens are invalidated and the connection is removed. The client will need to sign in again to reconnect.

Connected apps manages OAuth connections. Connections created with a token are revoked from API Tokens instead.

Security

How to protect your connections.

If you use OAuth login

  • You don't need to copy or store any token: the client takes care of it.
  • Only authorize official, trusted clients.
  • Check Connected apps regularly and revoke anything you no longer use.

If you use a token

Treat the token like a password.

  • Don't paste it into an AI chat.
  • Don't share it via email or insecure messages.
  • Don't publish it in screenshots, documents or repositories.
  • Set an expiry date when possible.
  • Revoke tokens you no longer use.
  • Create separate tokens for different clients.

To revoke a token, open Integrations → API Tokens from the user menu, find the token and click Revoke.

Common issues

The AI client can't find Sentl

Check that the MCP server is enabled, that the URL ends with /mcp and that the client supports remote MCP over HTTP.

OAuth login doesn't start or the client doesn't register

OAuth login is only enabled for the official clients' domains (e.g. Claude, ChatGPT). If your client points to a different domain, it can't self-register: use the token method.

The assistant is working on the wrong organization

No need to reconnect: just tell it. If two organizations have projects with similar names, ask it to list the organizations and point it to the right one by name.

Error 401 Unauthorized

Credentials weren't sent or aren't valid. With a token, check that the header is Authorization: Bearer <TOKEN_SENTL>. With OAuth, sign in again from the client: the session may have expired, and a login unused for 30 days always has to be repeated.

The assistant says it can't do something

The connection works, but Sentl refused that specific action and the assistant reports the reason. The most common ones:

  • Insufficient permissions: in that project's organization your role is User. Ask an Owner or Admin there to change your role or to take the action.
  • Insufficient credits: the run costs more credits than the organization has left (the message shows both numbers). Top up that organization's credits, or lower the rounds or the models.
  • Subscription not active: the organization's subscription must be reactivated before starting new runs.
  • Run already in progress: the project already has an open run, and Sentl doesn't start a second one. The message names the open run: ask the assistant to follow that one and start a new run once it has finished. A run that stops making progress is closed automatically after a while, with a warning that explains what happened.
  • Project setup incomplete: the project can't run yet, and the message says what's missing: at least one researcher model, or at least one active question. Add it, then start the run again.
  • Project or run not found: it doesn't exist, or it belongs to an organization you're not a member of.

I don't see the write tools

That's expected if you're a User in all your organizations. Write tools appear as soon as you're Owner or Admin in at least one.

Error 405 Method Not Allowed

The client is probably using the endpoint as a generic URL. Configure it as an HTTP MCP server.

The assistant doesn't see the data I expect

The connection sees all the organizations you belong to: if something is missing, you're almost always not a member of the organization that contains it, or the project is archived. Ask the assistant to list your organizations to see what it can actually reach.

Frequently asked questions

OAuth login or token: which is better?

They're equally supported: choose based on your client. If it supports OAuth, login is more convenient (no token to manage); otherwise use a token.

With OAuth login, do I need to paste a token?

No. Just enter the server URL and complete access with login and consent in the browser.

Can I connect more than one AI assistant?

Yes. Authorize each client via OAuth or create a separate token for each one.

Can I use the same connection for several organizations?

Yes. A connection (OAuth or token) works across all the organizations you belong to, with the permissions of your role in each. If you join a new organization, the connection sees it right away.

Can the AI assistant change my data without permission?

Changes are only allowed with write permissions, usually for Owner and Admin. Many AI clients also ask for confirmation before running tools that change data.

Should I create a token with an expiry date?

It's not required, but it's recommended. An expiry date reduces the risk if the token gets copied to the wrong place.