Glean MCP Server vs Coworker MCP (2026): Tools, Setup and Differences

Glean MCP is Glean's managed, remote Model Context Protocol server: it connects Claude, ChatGPT, Cursor, VS Code and other MCP clients to your company's Glean knowledge graph and to the tools your admin enables, with each person's own Glean permissions applied. It is built into Glean and on by default for eligible tenants, so for a company that already runs Glean it is the simplest way to bring Glean search, chat and agents into the AI tools people already use.

Glean and Coworker both put a precomputed company graph behind a hosted MCP server, and both publish benchmarks showing it beating per-app connectors. I read Glean's docs, blog and product pages on October 6, 2026, and each Glean fact links to its source.

What does "Glean MCP" refer to?

The phrase covers four related things:

How does the Glean MCP server work?

Glean's MCP security overview describes the server as a thin protocol adapter over Glean's existing APIs. Your MCP host connects over TLS to an endpoint shaped like https://{your-backend-domain}/mcp/{server}, with an OAuth or API token tied to you. Inside your company's Glean tenant, the server validates the token, runs the tool as you against Glean's Knowledge Graph, and streams back search results, snippets, answers or full documents.

Each customer gets a dedicated tenant in its own cloud project or VPC, and indexing, permissions and query processing stay inside it. The server never elevates access: if someone cannot open a document in Glean, no MCP host can reach it for them. Once results land in the host, the host's own data policies apply.

Glean's MCP Gateway announcement argues that "MCP is a protocol, not a context layer." Its server reads from a precomputed index and knowledge graph, so the model does not have to reconcile separate Jira, Slack and GitHub results at runtime. For breadth, Glean's connectors page lists 275+ native and MCP-based connectors.

What tools does the Glean MCP server expose?

Glean provides a default server with its managed toolset: search, read_document, code_search, employee_search, user_activity, memory, memory_schema and web_search. Glean serves each tool only where the deployment supports it; code_search, for example, needs a connected code repository (create and manage MCP servers).

Admins can add more built-in tools, including chat, gmail_search, outlook_search, meeting_lookup, data_analysis and knowledge_graph_query, plus toolsets for artifacts, image generation and on-demand skills. Glean agents can be exposed as tools too, except agents that use write tools or human-in-the-loop steps (agents as tools). The MCP Gateway adds connector, custom and third-party tools on top, more than 2,000 by Glean's count.

Admins can run several servers, each with its own URL and tools, such as an engineering server with Jira write tools. Glean advises against switching every tool on in one server, because long tool lists make the host model choose tools less accurately.

Which AI tools can connect to Glean MCP?

Glean lists 22 tested hosts and says any other MCP-compliant client can connect with the server URL (About Glean MCP server).

ChatGPT and Claude for Teams/Enterprise are admin-managed, so an admin adds the server URL in that product's admin console. Cursor, VS Code and Claude Code are set up by each user, and stdio-only hosts need a bridge such as mcp-remote. Glean also has plug-ins for Cursor, Claude Code and Codex.

Before a rollout, note two host limits from Glean's guides. In ChatGPT, only search and fetch equivalents are available, and users switch the connector on in each new conversation. Claude Desktop needs your firewall to allow Anthropic's connections to your Glean instance. Our ChatGPT MCP and Claude MCP guides cover the client side.

How do you set up Glean MCP?

For most companies the admin side is already done: Glean's setup page says the server and its OAuth authorization server are on by default for new customers and for those who had not configured MCP before. Admins manage servers under Admin console, Platform, Glean MCP servers, and can push host configuration to managed devices with MDM scripts.

For users, per Glean's user guide:

1. In Glean, open your profile, then Your settings, then the Third party apps and MCP tab.
2. On the Connect to your AI apps with Glean MCP card, keep the recommended server or pick another, and copy its URL.
3. Pick your host for tailored steps, or paste the URL into any MCP client.
4. Sign in with your company SSO when prompted.
5. Test with a request that names Glean, such as a search for a document you know exists.

In Claude Code that is one command, following the pattern in Glean's developer guide:

claude mcp add glean https://<your-backend-domain>/mcp/<server-name> --transport http --scope user

The backend domain usually looks like company-be.glean.com. If you cannot find the Configurator, ask an admin to check that MCP is enabled.

Is Glean MCP secure?

Glean's preferred sign-in is its own OAuth authorization server (OAuth 2.1 with PKCE) behind your SSO, such as Okta, Microsoft Entra ID or Google Workspace (About Glean MCP server). Tokens from your own identity provider also work, and user-scoped API tokens are a fallback for hosts without OAuth support. Access tokens last 7 days and refresh tokens 180 days by default, and admins can change both (MCP security).

Admins can restrict access to company-managed ChatGPT Enterprise or Claude Enterprise workspaces, instead of personal accounts, by turning off dynamic registration and issuing static OAuth clients. Glean Protect+ can screen tool calls for prompt injection and malicious code, and check that each call matches what the user asked for.

Glean treats local MCP servers as development-only. Its old local server, gleanwork/mcp-server, is archived on GitHub, and the README points new users to the managed server. For the wider risks, see our guide to MCP security.

Glean MCP vs Coworker MCP: how do they compare?

Both put a precomputed company graph behind the protocol.

Sources: Glean's MCP docs, connectors page and homepage for compliance; Coworker's MCP page, OM2 page and benchmarks.

*Statistically significant benchmarks comparing Coworker MCP vs. Claude Native Tooling: Claude with Coworker MCP and OM2 against Claude with its native connectors, 100 tasks across 7 categories, blind pairwise review. The 9x and 64% figures are the best case (retrieval-heavy categories, with Coworker Learning); across all 7 categories the same setup was 75.5% cheaper and 45.6% faster. Glean's MCP evaluation used about 175 queries in Claude Cowork on Claude Sonnet 4.6. Neither study tested the other vendor's server, and the methods differ, so the numbers are not directly comparable.

What Coworker MCP does differently

Glean builds its graph on its search index: its post on building a context graph describes inferring entities such as projects, customers and teams from indexed content, then adding activity signals like views and edits.

Coworker starts from facts. "You can think of OM2 as a context graph," Coworker's CEO wrote in the OM2 launch post. Documents, messages and tickets are broken into atomic facts linked to entities and relationships, then ranked by proximity, recency, entity centrality and relationship strength. The launch post says this supports enumeration, negation, counting, structural ranking and multi-hop traversal, with source permissions enforced on traversal and provenance back to each source. Coworker Learning, part of OM2, improves retrieval as a team uses it.

Glean also exposes knowledge_graph_query over MCP, so both can answer relationship questions. If most of your questions are about finding the right document, Glean's search-first design fits. If they sound like "which accounts never replied" or "how many tickets mention this bug", test both on counting and negation with your own data.

When is Glean MCP the better choice?

• You already run Glean. The server is part of your instance, on by default for eligible tenants and priced under Glean's usage-based FlexCredit terms, and your permissions already live in Glean.
• You need connector breadth: 275+ connectors against Coworker's 50+. If a system you depend on is only on Glean's list, that settles it.
• You need documented MCP governance at fleet scale: MDM rollout, per-team servers, per-user tool access and an MCP Insights dashboard.
• You want one governed endpoint for third-party MCP servers too, which the Gateway provides.

Glean also documents far more MCP admin controls than Coworker's MCP page covers, so ask about any control you need.

When does Coworker MCP make more sense?

• You do not run Glean and want company context inside Claude, ChatGPT, Gemini, Perplexity, Cursor, Windsurf or Claude Code.
• Your questions need structure: counts, lists, things that did not happen, or answers that chain across records.
• Token cost on retrieval-heavy work matters, which is where the benchmark above shows the largest gains.
• Your stack fits the 50+ connectors, such as Salesforce, HubSpot, Slack, Jira, Google Drive, Notion and Gong.

The trade-off is breadth, since Glean's connector list is much longer. For the full platform comparison, see Coworker vs Glean.

How do you set up Coworker MCP?

Your admin enables MCP access for your company. You then add the Coworker MCP server in your AI client and sign in with your existing Coworker login and SSO. Per the Coworker MCP page, setup typically takes under 30 minutes and queries respect your existing permission model.

If you are choosing what sits behind your team's AI tools, test both on the same questions from your own data. Coworker will run its benchmark on your company's data and hand you a Context Impact Report, or you can book a demo to see Coworker MCP in Claude or Cursor.

Frequently asked questions

Does Glean support the Model Context Protocol (MCP)?

Yes, in both directions. Glean's managed remote MCP server exposes search, chat, documents, people lookup and agents to clients such as Claude, ChatGPT and Cursor. Glean Assistant and Agents can also call external MCP servers, and Glean lists 178 supported remote servers.

Is Glean an MCP server?

No. Glean is an enterprise search and AI platform that indexes your company's apps into a knowledge graph. The Glean MCP server is one way to reach that index from other AI tools, alongside Glean's own apps and REST APIs.

What tools does the Glean MCP server expose?

The default server includes search, read_document, code_search, employee_search, user_activity, memory, memory_schema and web_search. Admins can add chat, mailbox search, meeting lookup, data analysis, knowledge graph queries, Glean agents, and read and write tools through the MCP Gateway.

How much does Glean MCP cost?

Glean's documentation says its MCP server is subject to usage-based pricing under its FlexCredit terms. Glean's pricing URL redirects to its homepage, which lists no prices. See our Glean pricing breakdown for details.

Is there a local or open-source Glean MCP server?

Not a supported one. Glean's original local MCP server repository on GitHub is archived and points new users to the managed remote server. Glean's security documentation treats local MCP servers as development-only.

What is the Glean MCP Gateway?

It adds tools from outside Glean to your Glean MCP servers: read and write tools from connectors, custom tools, and proxied third-party MCP servers, more than 2,000 governed tools by Glean's count. Tools inherit source permissions.

What is the difference between Glean MCP and Coworker MCP?

Both are hosted MCP servers backed by a precomputed company graph. Glean MCP reaches Glean's index and Enterprise Graph across 275+ connectors, with extensive admin controls. For large organizations, Coworker MCP reaches OM2, Coworker's organizational memory of atomic facts and relationships across 50+ connectors. Companies already on Glean should start with Glean MCP.

Related reading

• Coworker MCP
• Coworker vs Glean
• Glean pricing
• Glean integrations
• Glean Agent Builder
• Glean vs Coworker vs Guru
• What is MCP?
• MCP security
• What is an MCP gateway?
• Organizational memory (OM2)