Give Your Trae AI Agents a Custom Brain with CustomGPT.ai’s hosted MCP (Model Context Protocol) server

No. In the standard Trae setup, you do not run your own MCP server. Your team runs a lightweight local bridge with your project ID and MCP token, while the MCP endpoint is hosted by CustomGPT.ai.

The practical split is simple: the provider hosts the endpoint; you manage only the local bridge process and your credentials unless you choose a separate customer-hosted deployment. Choose the default hosted setup if you want the fastest connection and do not want to manage backend infrastructure. Choose a customer-hosted MCP endpoint only if your security, networking, or compliance rules require your team to host and control that endpoint, such as inside your own VPC.

That distinction exists because MCP supports both local and remote server transports in Anthropic’s spec. The same hosted-versus-owned decision comes up in Claude Desktop and Cursor. For scale, Lehigh University uses AI search across 400M+ words of newspaper archives without needing customers to stand up their own MCP infrastructure.

Trae gives generic answers after MCP connection when it falls back to its base model because the MCP tool was not selected, authenticated, or refreshed. A green connection only shows the server is reachable, not that retrieval is active.

Check the active agent first, then the project ID, token, and tool refresh. A 401 in the MCP log usually means the token is invalid or expired. A 403 usually means the token lacks access to that project. A project-not-found error usually means the project ID is wrong. If no tool call appears at all, Trae is likely still using a non-MCP agent or a cached tool list. In the MCP spec, clients discover tools through tools/list, so stale manifests can persist until reconnect or restart, including in Cursor and Claude Desktop. Fully quit and relaunch Trae. Then ask something only your CustomGPT.ai knowledge base can answer and confirm the MCP log shows a tool call.

Use Trae’s hosted MCP server by default. Choose the hosted option unless your security review requires internal audit logs, customer-managed credentials, private network routing, or regional data handling controls; in those cases, ask Trae support to confirm whether a customer-hosted MCP connection is supported for your workspace.

As of Trae Docs, “MCP Setup,” available in March 2026, only a hosted endpoint is described, not a customer-hosted connection flow. Anthropic’s Model Context Protocol documentation allows MCP servers over local stdio and HTTP transports, but protocol support does not mean Trae currently exposes customer-hosted MCP connections in the product. Hosted is therefore the lower-risk path for faster rollout, simpler support, and fewer auth, uptime, patching, and secret-rotation tasks. By comparison, Claude Desktop and Cursor publicly document broader MCP connection patterns. This recommendation is specific to Trae and should not be assumed for CustomGPT.ai or other platforms.

Yes. Trae can answer from your company documents when it is connected to an approved knowledge base. CustomGPT.ai can provide that knowledge base, and if your setup also uses MCP, Trae retrieves from that approved source at answer time; the hosted versus customer-managed MCP pattern depends on your deployment.

It can cite the exact passage, respect document permissions, and refuse to answer when evidence is weak. A common policy is to require one passage above a confidence threshold such as 0.8, or two matching passages, otherwise Trae returns “I don’t know” instead of guessing. For best results, upload text-based files with clear headings, run OCR on scanned PDFs, and split large manuals into topic-focused documents. Teams evaluating Glean or Microsoft Copilot usually care less about generic AI fluency than whether every answer is auditable and traced to cited source text. At MIT, a source-grounded deployment across 90+ languages reported zero hallucinations because responses were tied to cited source text.

No. MCP does not give Trae live browsing access to your intranet or private wiki. Trae can only answer from content that has first been added and indexed in the connected knowledge project; MCP lets it query that indexed content, not your network directly.

Usable inputs typically include PDF, DOCX, HTML, plain text, individual URLs, and website crawls, including authenticated site ingestion when a connector supports login. If your wiki can be exported as HTML or documents, or exposed as crawlable pages, it can work. If not, it will not be usable through this setup. For example, a team can export Confluence or SharePoint pages as HTML or DOCX, add them to CustomGPT.ai, and then let Trae answer over that indexed copy. Lehigh University queries 400M+ indexed words, which shows this pattern can scale to large knowledge collections. Like Glean or Elastic, this is indexed enterprise search, not a live intranet session.

Based on the reviewed setup materials, data access appears limited to the connected knowledge project and protected in transit over HTTPS, but those materials do not establish a full privacy policy. The configuration shown includes a project-specific HTTPS endpoint and bearer-token authentication, which indicates scoped access to that project.

For CustomGPT.ai, that supports project-scoped retrieval, not blanket privacy assurance. The setup screen alone does not confirm retention period, whether prompts or retrieved files are used for model training, who can access logs, SOC 2 scope, GDPR terms, or data-storage region. One technical point from RFC 6750: a bearer token works for whoever possesses it, so secure storage, rotation, and revocation matter as much as HTTPS. Before using Trae and MCP for sensitive data, request written confirmation on retention, training exclusion, log access controls, subprocessors, and regional storage. OpenAI Enterprise and Microsoft Copilot Studio publish these details more explicitly.