Your agent can edit files and run commands. Some work needs outside systems. Query an error in Sentry, pull a ticket from Jira, read a row from PostgreSQL, or check a status from GitHub. MCP servers in Tarsk are the bridge that connects your agent to those tools, databases, and data sources, and their tools become calls the agent can make during a conversation.
What problem this solves
Copying a stack trace into Sentry, pasting the response back, and repeating that for each issue wastes your time and the agent’s. MCP changes the pattern. You connect a server once, and the agent calls its tools directly, reads the results, and uses them to answer your request.
Tarsk supports two kinds of MCP servers. A local server runs as a process on your machine. A remote server runs at a URL that your agent reaches over HTTP. You can install a ready catalog entry or add your own server with a URL, headers, and optional OAuth.
Requirements and setup
Open Settings → MCP Servers in the project where you want the connection. The screen shows the servers configured for that project in an Installed MCP Servers section, with a marketplace of catalog entries above it.
You need one of these to connect:
- A catalog entry you can install with one click.
- A local server command, such as an npx package, for a stdio server.
- A remote server URL, plus credentials if the service requires them.
Each installed server exposes tools. The agent loads those tools when the tools are enabled for the project, and MCP tools load by default alongside the core coding tools. Server changes apply on a new thread, so start a fresh chat after you add or change a server to use its tools.
Install an MCP server
1. Browse or search the marketplace
In Settings → MCP Servers, the Available MCP Servers catalog lists entries such as Sentry, GitHub, Context7, Playwright, Supabase, PostgreSQL, and Notion. Use the search box to filter, then choose a server.
2. Install into the project
Click a catalog card to install that server into the current project. A remote server with a real URL installs directly. A few entries, such as Zapier, open the Add Server dialog prefilled so you can paste your personal URL or token.
3. Sign in when requested
A remote server that uses OAuth shows a Sign in button on its row. Select it, and Tarsk opens the service’s sign-in in your browser. Approve, and the flow completes on a local callback page. For a server such as Sentry that supports dynamic client registration, you sign in with your account and reuse the session across future runs. Some services need a client you register separately; you enter its Client ID, secret, and callback port in the server editor first.
Signed-in credentials are encrypted and stored locally, scoped to the project. Signing out revokes them.
Add a custom MCP server
When the catalog does not have what you need, select Add Server and choose a tab.
Local (stdio)
Enter a Command (for example npx), its Args (such as -y @modelcontextprotocol/server-filesystem and an allowed path), and any Environment variables. Tarsk launches the server as a local process. Local servers commonly take authentication through environment variables, including a personal access token where a service asks for one.
Remote (HTTP)
Enter the server URL, pick a Transport (sse or streamableHttp), and add HTTP headers or environment variables when the service needs them. For an OAuth-protected server, fill the OAuth client section with the Client ID, secret, and callback port from the service’s developer settings if it does not support dynamic registration.
A server must have a URL or a command, never both, and never neither. A local server requires an args array.
A practical example: connect Sentry
You want the agent to inspect an error before you fix it. In Settings → MCP Servers, find Sentry in the marketplace and install it. The added server uses the Sentry MCP endpoint and streamable HTTP transport.
On the server row, select Sign in. Tarsk opens Sentry’s OAuth screen in your browser. Complete the sign-in, and the server reads Authorized.
Start a new chat so the agent loads the Sentry tools. Ask it to look up the latest issue in a project. The agent calls the Sentry tool, reads the stack trace and event context, and explains the failure and a likely fix.
Keep the scope tight when you set up Sentry. A server URL bound to one organization or one project lets the agent reach the data you care about, and avoids granting a wider footprint than needed.
Advanced tips
Sign in once, reuse across sessions
For a server that supports dynamic client registration, sign in once and the credential renews across sessions. No static token to paste and no shared credential to rotate by hand.
Keep test and production separate
Connection choices are per project. Use a personal token in a local project and a scoped token or the correct OAuth account in the project where production work runs.
Test the connection on a fresh thread
Pipe commands that inspect or modify data run against the connected server. Verify the server authenticates and exposes the tools you expect, then continue in a fresh thread so the tool set matches the server configuration you saved.
Common pitfalls
Sign in is missing or fails
A remote server that does not support dynamic client registration needs its client registered in the service’s developer settings first. Add the Client ID, client secret, and matching callback port in the OAuth client section. A client secret without a client ID is rejected, as is a callback port already in use.
The agent does not see the new tools
MCP tools connect at conversation start. Start a new chat after installing, changing, or signing in to a server so the agent loads its tools. Check that MCP tools are not disabled for the project.
Tool calls authenticate with the wrong account
Remote credentials are stored per project. Confirm you signed in with the account you want for that project, and sign out from a server row to revoke a credential you no longer want.
Summary
MCP servers connect your agent to external tools, databases, and data sources. Install a server from the marketplace or add your own local or remote server, sign in once when the service uses OAuth, and start a new thread so the agent can call its tools.
Begin with one catalog entry for a service you use daily. Install it, sign in, and ask the agent to query that system on your next task.