Drushya MCP

• Meeting dates: started_from includes its timestamp; started_before excludes it. Both accept RFC 3339 timestamps with explicit offsets, for example 2026-10-24T00:00:00+05:30. Dates filter meeting start times.
• Search: use query with AND between words, uppercase OR, minus exclusions, or quoted phrases. Search is based on text matches, not semantic similarity.
• Pagination: pass a returned nextCursor as the next call’s cursor, preserving the filters. Stop when it is null. Transcript calls also accept limit and max_characters.
• Task creation: supply meeting_id and title. Optional fields are assignee_name, due_date (YYYY-MM-DD), and priority (low, medium, high). New tasks start incomplete.
• Task updates: also supply task_id and at least one change. Use completed to complete or reopen a task. Null clears assignee, deadline, or priority. Both task writes return meetingId and items containing all current tasks in that meeting, including unchanged ones. Each write targets one task.
• Templates: creation uses name and field_ids. Updates use template_id with a name and/or field change. Listing returns items plus supportedFields; writes return template. Existing summaries are not regenerated.
• Template fields: executiveSummary, fullSummary, detailedContent, attendees, keyDiscussionPoints, decisions, and nextSteps. Custom section names and freeform template prompts are not supported.
• Tag visibility: pass tag_id and visibility set to private or shared. Adding/removing meeting tags uses tag_id and meeting_ids (1–100 IDs). Visibility cannot target individual recipients.
• Retries: task and template creation are not idempotent. If a response is interrupted, check the existing records before repeating the request.

Use list_meetings or search_meetings with include: ["summary"]. Ordinary lists remain lightweight. Both tools accept include_tag_ids and exclude_tag_ids; filtering happens before pagination. Tag filters require tags:read.

This list_meetings example covers October 27 through November 3, 2026, in India time. Call list_tags first and replace the placeholder IDs. Omit include_tag_ids to review all eligible meetings except Tests.

{
    "started_from": "2026-10-27T00:00:00+05:30",
    "started_before": "2026-11-04T00:00:00+05:30",
    "include_tag_ids": [
        "<Release-5.4-tag-id>"
    ],
    "exclude_tag_ids": [
        "<Tests-tag-id>"
    ],
    "include": [
        "summary"
    ],
    "summary_language": "en",
    "summary_format": "markdown",
    "limit": 25,
    "max_response_bytes": 64000
}

• Saved versions: summary_language accepts en, hi, or kn. Optional template_id selects a saved template. These options also work on get_meeting and get_meeting_summary.
• Formats: summary_format defaults to markdown. Request document for the rich-text document, or legacy for the former full summary object. Clients that relied on summaryDoc or fullSummary should request the corresponding format explicitly.
• Byte budget: max_response_bytes defaults to 64000 and accepts 4096–256000. It bounds the serialized MCP result, including its text and structured copies, but excludes HTTP and JSON-RPC framing. Bytes are not a model-independent token count.
• Meeting pages: limit is a maximum; the byte budget can return fewer records. Pass nextCursor as cursor with the same filters and summary selection until it is null. Ordering is newest meeting first, with ID as the tie-breaker.
• Oversized summaries: contentStatus: "deferred" includes a retrieval reference to get_meeting_summary. Follow its cursors and concatenate each page’s content. If contentEncoding is json, parse only after assembling every page.
• Changed summaries: continuation fails explicitly if the saved content or selected version changes. Discard the partial result and restart the summary request without a cursor.
• Coverage: check generationStatus, contentStatus, and availableVersions. Retained content can accompany a failed or unfinished regeneration. Do not describe it as a newly completed summary.
• Compatibility: existing tag_id remains supported and is merged with included tags using “any” matching. Unknown IDs in the new arrays produce an error. Use either structuredContent or its JSON text copy once, rather than sending both to the model.

Check that your AI client supports remote MCP and that its plan and workspace policy allow custom servers. A team administrator may need to enable or add the connection. Use the provider documentation linked above for its current interface.

The /mcp URL is a protocol endpoint, not a website. Opening it directly can show an authentication error or Method not allowed. Paste it into your MCP client and complete OAuth sign-in instead.

Use the exact HTTPS server URL and automatic OAuth client registration (DCR). Sign in with the Drushya account that owns or can access the intended workspace. If a login attempt expired, start authentication again from your AI client.

Check that Drushya is enabled in the client and in the conversation. Check the client’s MCP status for errors, then reconnect if authorization has expired. In Claude Code or Codex, /mcp shows connection status.

The operation needs write permission as well as access to the underlying record. Ask your client to request the relevant scopes and complete consent again. Task writes need both tasks:read and tasks:write; built-in templates cannot be edited.

Refresh the client’s Drushya tool discovery or reconnect the MCP connection, then start a new conversation if needed. The current tool list includes get_meeting_summary and the meeting-list options include_tag_ids, exclude_tag_ids, and include.

MCP reads saved versions only. Check availableVersions, choose an existing version, or generate the required version in Drushya before requesting it again. A read does not generate or translate a summary.

A response can return fewer meetings because of its byte budget or because fewer meetings match your filters. Follow nextCursor for more meetings and use the summary’s retrieval reference for get_meeting_summary. Continue until every cursor is null; deferred content has not been silently cut off.

The content or selected version may have changed. Discard the partial summary and restart without a cursor. For meeting pages, reuse the same filters, language, template, and format.

Check the Drushya account and workspace selected during authorization, then check date and tag filters. Follow pagination for older records. Deleted meetings and records outside your authorized access are unavailable.

Drushya searches keywords and phrases. Try a distinctive term from the discussion, retrieve the matching notes or transcripts, and ask your AI client to summarize that evidence. Searching by meaning alone is not supported.

A cloud-hosted AI client may reach Drushya from the provider’s network rather than your computer. Ask your administrator to check the workspace’s allowed networks and the provider’s connection requirements.

No. The Drushya account you choose during browser sign-in determines which workspace data the connection can access.

Not with a client that supports automatic registration. Drushya supports OAuth discovery and dynamic client registration. Choose automatic registration / DCR rather than supplying credentials manually.

Not to read meeting content already saved on the server. MCP retrieves your saved workspace data; it does not start a recording or stream live audio.

The documented write tools manage tasks, personal summary templates, tags, and tag assignments. They do not edit transcripts or summaries, delete meetings, or control recordings. Each action is limited by the permissions you approve and your access to the record.

No. Saved language and template reads leave the meeting’s title and selected version unchanged. Versions in any of the 46 supported languages can be requested when they have already been saved.

No. Tag visibility controls whether the label is private or workspace-visible. It does not grant access to meeting content or share a label with selected email addresses.

Templates use Drushya’s predefined summary fields. Ask the client to list summary templates and supported fields first. Arbitrary section names such as wins and blockers are not accepted.

It returns the information stored in Drushya’s People records, such as name and email. It does not enrich contacts with external role or company information.

Yes. Implement Streamable HTTP MCP with OAuth 2.0, PKCE, and dynamic client registration. Use the same server URL, authorize a Drushya workspace, and discover the tools through MCP.

Disable or remove Drushya in your AI client’s connector or MCP settings. This removes it from that client’s available tools; it does not delete your Drushya meetings.