MCP Library

What this page is for

MCP — the Model Context Protocol — is an open standard for connecting AI agents to external tool providers. The MCP Library page is Exolvra’s built-in catalogue of community MCP servers, pre-vetted for quality and kept current. Use it to find, install, and manage MCP servers that extend what your agents can do.

See Tools, skills & MCP for the conceptual difference between built-in tools, skills, and MCP servers.

Layout

The page has two views, toggled from the header:

1. Library — the catalogue. Browse every available MCP server by category, search by name, filter by runtime requirements and transport type.
2. Installed — the MCP servers you’ve already installed. Manage credentials, verify connectivity, disable, or uninstall.

Runtime status

At the top of the Library view, a small badge shows whether Node.js and Python are detected on the host. Green = installed; grey = missing. Most stdio MCP servers need one or both runtimes. HTTP MCP servers don’t need any.

If runtimes are missing, the install button on stdio servers is disabled with a tooltip explaining what’s needed. See MCP setup for install instructions.

Catalog cards

Each MCP server in the catalogue shows:

• Name and description
• Category — database, SaaS, search, cloud, dev tools, productivity, etc.
• Transport — stdio, HTTP, or SSE
• Runtime — Node, Python, pipx, or none
• Verified by — Exolvra team, community, or official vendor
• Install button

Click any card for the full detail page with the server’s capability list, environment variable requirements, and installation instructions.

Common tasks

Browse and install a server

1. Open the Library view
2. Filter or search for what you want (for example, “github” to find the GitHub MCP)
3. Click Install on the card
4. A dialog prompts for any required credentials or configuration (API keys, endpoints, tokens)
5. Fill in the values — they’re encrypted at rest before saving
6. Click Install
7. Exolvra spawns the server process (for stdio) or verifies connectivity (for HTTP) and adds it to the Installed list

Verify an installed server

From the Installed view, click any server and hit Verify. Exolvra tests the connection: for stdio servers it spawns the subprocess and checks that it responds to the MCP handshake; for HTTP servers it makes a request to the endpoint. A green check means ready; a red error means something’s wrong and the error message will tell you what.

Common verification failures:

• Credentials missing or wrong
• Runtime not installed (stdio)
• Network blocked (HTTP)
• Server package not yet downloaded (first run of npx -y packages can take 30+ seconds — retry verification)

Update credentials

Open the server detail page, click Edit config, update the fields, save. The new values are encrypted and applied on the next tool call.

Control per-agent access

Every installed MCP server has a per-agent access control list. From the Installed view, open the server detail page and click Access. You’ll see a list of every agent in your instance with a checkbox for each — check to grant, uncheck to revoke. By default, new MCP servers are not granted to any agent. You have to explicitly enable them.

Uninstall a server

From the server detail page, click Uninstall. Confirm. Exolvra stops the subprocess (if stdio), removes the configuration, and wipes the encrypted credentials. Any agents that were granted access lose it on their next turn.

Transport choices

MCP servers come in three flavours:

• Stdio — Exolvra spawns a subprocess that reads and writes on stdin/stdout. Needs a local runtime (Node, Python). Fast, low latency, runs on your host.
• HTTP — Exolvra makes requests to a remote HTTP endpoint. No local runtime. Runs on someone else’s server.
• SSE — Server-Sent Events over HTTP. Same as HTTP but supports streaming. No local runtime.

Stdio is common for open-source community MCPs that wrap local things (databases, git, file system). HTTP is common for SaaS MCPs (Zapier, Pipedream, Cloudflare, Notion).

If you’re running Exolvra in an air-gapped or tightly-sandboxed environment, HTTP MCPs are the only option — they don’t need a local runtime.

Credentials are encrypted

Every environment variable or credential you provide at install time is encrypted at rest using the platform’s data-protection system. The plaintext is never written to the database. On the way out to an MCP server, the values are decrypted in memory, injected into the server’s process environment, and never logged.

That said — anyone with admin access to the Exolvra host can read the decryption keys. The encryption protects against at-rest snooping (database dumps, backups); it doesn’t replace host-level security.

Common pitfalls

Installing a stdio MCP without its runtime. The install button is disabled on stdio servers when the required runtime is missing. If you install one manually via the API and the runtime is absent, the server will refuse to start on every call. Install the runtime first — see MCP setup.

Not granting per-agent access. A newly-installed MCP is not automatically granted to any agent. Install is step one; granting is step two. If you install and then wonder why no agent is using the server, check the access list.

Passing sensitive credentials through an untrusted HTTP MCP. HTTP MCP servers run on someone else’s infrastructure. If you send customer data through one, that data leaves your perimeter. Only use HTTP MCPs from vendors you trust with the data you’re sending.

Where to go next

• MCP setup — install Node / Python so stdio MCPs work
• Tools, skills & MCP — the conceptual model
• Skills — the in-process alternative to MCP
• Agents — per-agent access is set in the tool grants