Get grounded answers with Chat and MCP

Grounded answers use workspace knowledge as evidence instead of relying only on a model's general memory. This journey verifies the same knowledge through Usable Chat and an external MCP-capable client.

Choose a question and source

Start with a question that has a known answer in a workspace you can access. Identify one canonical fragment that should support the answer.

Good test question:

What is our documented process for reviewing a public fragment before publication?

Avoid testing with vague questions whose answer is spread across unrelated workspaces.

Get the first grounded Chat answer

1. Open Usable Chat.
2. Add the relevant workspace, fragment, or collection as context.
3. Ask the task-oriented question.
4. Request the sources used.
5. Open the cited fragments.
6. Compare the answer with the full source content.

Workspace, fragment, and collection pages can expose Ask Usable links that open Chat with context already selected.

A grounded answer should identify its evidence, remain within the selected workspace boundary, and distinguish uncertainty from fact.

Improve a weak answer

If the answer is generic or unsupported:

• narrow the workspace scope;
• attach the canonical fragment directly;
• use the exact product or process term;
• ask for source titles before conclusions;
• split a compound question;
• improve the source fragment's title, summary, headings, or content.

Do not treat a fluent response as verification.

Connect an MCP client

OAuth-capable remote MCP clients should use:

URL: https://usable.dev/api/mcp
Transport: Streamable HTTP
Authentication: OAuth 2.1 browser sign-in with PKCE

From the dashboard, choose Connect MCP, copy the remote URL, add it to your client, and complete browser authentication. Clients without OAuth support can use a dedicated, least-privilege personal access token where supported.

Verify MCP access

Ask the client to:

1. list accessible workspaces;
2. search the intended workspace;
3. retrieve the full source fragment;
4. answer using that source;
5. name or link the evidence.

The reliable pattern is:

search → inspect candidates → fetch full fragments → answer → cite

Search snippets and summaries are discovery aids, not complete evidence.

Compare Chat and MCP results

Both interfaces should reach the same canonical facts when they have the same workspace access and source material. Differences can come from selected context, token scope, model behavior, search query, or whether the client fetched the full fragment.

Check:

• workspace identity;
• source fragment and updated time;
• selected context;
• client permissions;
• whether the answer cites actual retrieved content;
• whether one interface is using a stale conversation.

Allow writes only when needed

Start MCP clients with read-only fragment permissions. Add create or update permission only for a defined workflow.

Before an AI writes:

1. search for an existing canonical fragment;
2. fetch it in full;
3. confirm the target workspace and fragment type;
4. draft the change;
5. review public/private safety;
6. apply the smallest update;
7. refetch and verify the result.

For public documentation, require release verification and human review.

Troubleshooting

The client does not authenticate

Confirm it supports remote MCP over Streamable HTTP and OAuth/PKCE. For PAT-based clients, confirm the token is active and configured as a bearer credential.

The workspace is missing

Check the signed-in account, invitation/subscription state, token workspace scope, and role.

Search returns no useful source

Broaden the query, remove strict filters, verify workspace scope, and use exact listing/filtering when you know metadata. Then fetch candidates in full.

A tool is visible but fails

Tool visibility does not guarantee permission for every workspace/action. Distinguish authentication failures from authorization and domain errors.

Chat and MCP disagree

Start a new conversation, pin the same source, fetch its current full content through MCP, and compare the evidence rather than the wording.

Related pages

• Using Usable Chat
• Context and grounded answers
• Connect an AI client to Usable
• Search and ground agent answers
• Create and maintain knowledge safely