Read AI
by Read AI
Read Read AI meeting transcripts, summaries and action items from an AI agent. Anthropic lists 2 read tools; Read AI's own documentation describes 4, and the 2 omitted both write — one sends a recording bot into a live meeting, the other emails a meeting report to a third party. OAuth 2.1.
Verified connector
Listed by Anthropic as a partner connector in its Connectors Directory.
Connection checked by Agentman on .
Anthropic states this reflects the level of review a connector received, not a security audit.
Connect Read AI via MCP
https://api.read.ai/mcpWorks in any MCP-compatible client. In Agentman Studio it is one click — no config file to edit.
Read AI Tools & Capabilities (4)
Tool names from Anthropic's directory listing. This server requires sign-in, so we could not read tool descriptions or parameter schemas.
Limits
- Anthropic's listing omits both write tools. The directory carries two names and marks the connector
"permissions": "Read". Read AI documents four tools, and the two missing ones dispatch a recording bot and email a meeting report to a named person. Do not size this connector's risk from the directory entry. - We could not verify which list the live server serves. The endpoint returned HTTP 401 to an anonymous request, so no
tools/listresponse, no parameter schemas and no safety annotations. Whether the deployed server exposes two tools or four is unresolved from outside. - We do not publish machine names for the two write tools. Read AI's article gives them as prose headings only. Guessing at snake_case identifiers would fabricate a fact, so they appear here in Read AI's own wording.
mcp:executeis an undifferentiated grant. One of seven advertised scopes names a resource, and it is the read one. Nothing at the OAuth grant separates reading a transcript from sending a bot into a live call or emailing a report externally.- Both outbound actions are irreversible on the first call. A dispatched bot has been seen by the other participants; a sent invite has been delivered. Read AI's documented idempotency protects against duplicate retries, not against the first call.
- Consent in desktop and mobile mode is entirely yours to obtain. No bot appears, no participant list entry, and the "Read Stop" and "Opt Out" chat commands do not work. Read AI says so itself and provides recommended consent language, but cannot enforce it.
- Global Report Access widens the read surface to the whole workspace, and Read AI names the MCP server explicitly as a path that inherits it. We found no way to scope the connector more narrowly than the account it authenticates as.
- There is no delete, edit or correction tool. Nothing in either list removes a report, redacts a passage or fixes a misattributed speaker. Deletion runs through the Read AI app, an in-meeting "Opt Out", or account deletion.
- The model-training exception is opt-out, not opt-in. Read AI states meeting content is not used for training without explicit opt-in, except for a Customer Experience Improvement Program that is opt-out by default and which it says 10 to 15 percent of users are in.
- Token handling is awkward by Read AI's own account. It documents 10-minute access tokens and single-use rotating refresh tokens, warns this "may require manual intervention if token chains break", and reports some MCP clients failing against its authentication.
- Rate limit is 100 requests per minute per user, returning HTTP 429.
list_meetingscaps at 10 meetings per page. - The whole thing is an open beta. Read AI labels both the API and the MCP server as open beta, lists feature gaps including Ask Read and coaching, and says additional tools are planned.
- The vendor's own docs are unreachable to a normal HTTP client. Read AI's help centre sits behind a Cloudflare challenge that returns HTTP 403 to a plain request. Everything cited here came through the
help_centerJSON API, which Read AI'sllms.txtnominates for exactly this purpose.
Frequently asked questions
Read AI documents four; Anthropic's directory lists two. The directory names get_meeting_by_id and list_meetings, both reads. Read AI's own MCP Server article adds Create meeting agent and Share meeting report, numbered 3 and 4 on the same page. The two the directory omits are the only two that write.
Yes, per Read AI's documentation, though Anthropic's listing does not show that tool. Read AI documents a Create meeting agent tool that dispatches a Read AI bot to a Zoom, Google Meet or Microsoft Teams meeting from a join link, optionally scheduled for a future time. Anthropic's directory lists only the two read tools.
In web mode yes, in desktop mode no. Read AI states that in web mode the bot appears in the participant list, may announce itself in chat, and any attendee can type Opt Out to end recording and delete that session's data. In desktop and mobile mode audio is captured locally, no bot appears, and those commands do not work.
Read AI documents a Share meeting report tool that grants a named email address viewer or editor access and sends an invite. Read AI states that if the recipient has no account an email-based invite is created for them, and that a notify flag defaults to yes. Anthropic's directory does not list this tool.
Seven, of which two are functional. Read AI's RFC 9728 descriptor advertised openid, profile, email, offline_access, offline, mcp:execute and meeting:read on 2026-08-23. Only meeting:read names a resource. The write tools Read AI documents can only fall under mcp:execute, which names no object and separates nothing.
Full transcripts with speaker attribution, plus participant names and email addresses. Read AI documents the meeting object as carrying participants with name, email, invited and attended flags, an owner record, engagement metrics, and a URL to download the recording as an MP4. Those participants include people outside your company.
Yes, if Global Report Access is enabled. Read AI states that workspace owners and admins granted this setting can retrieve data for all meetings in their workspace through the REST API and MCP server, and warns to use it with caution as it could expose sensitive information. Without it, a user sees only their own reports.
Read AI documents 100 requests per minute per user, returning HTTP 429 when exceeded. It also documents access tokens expiring after 10 minutes and single-use refresh tokens that rotate on each refresh, and warns this may require manual intervention if token chains break. Both API and MCP server are in open beta.
Sources
- Read AI MCP Server support article (retrieved 2026-08-23 via the Zendesk
help_centerJSON API after the HTML returned HTTP 403 to our client; article last updated 2026-08-19). Source of the four-tool enumeration. · retrieved 2026-08-23 - Read AI API and MCP Overview (retrieved 2026-08-23). Source of open-beta status, prerequisites, rate limits and token behaviour. · retrieved 2026-08-23
- Read AI API Reference (retrieved 2026-08-23). Source of the meeting object schema. · retrieved 2026-08-23
- Read AI API Keys & Authentication (retrieved 2026-08-23). Source of the OAuth 2.1 registration flow and the requested scope string. · retrieved 2026-08-23
- How Consent Works on Web Mode vs. Desktop Mode (retrieved 2026-08-23) · retrieved 2026-08-23
- How to Stop or Opt Out of a Read AI Recording (retrieved 2026-08-23) · retrieved 2026-08-23
- What Does Read AI Do With My Data (retrieved 2026-08-23). Source of retention controls and the training opt-out. · retrieved 2026-08-23
- Global Report Access (retrieved 2026-08-23) · retrieved 2026-08-23
- Read AI Security & Privacy Overview (retrieved 2026-08-23) · retrieved 2026-08-23
- Read AI
llms.txt(retrieved 2026-08-23). Nominates thehelp_centerJSON API as the canonical machine-readable source and permits contextual explanation of Read AI's products while prohibiting model training;support.read.ai/robots.txtpermits/hc/en-us/articles/. · retrieved 2026-08-23 - Live OAuth posture check: anonymous initialize returned HTTP 401 with a
www-authenticateBearer challenge naminghttps://api.read.ai/.well-known/oauth-protected-resource/mcp; that descriptor and the authorization server metadata athttps://authn.read.ai/.well-known/oauth-authorization-serverwere both read (2026-08-23). Three control paths, including a nonsense resource path, all returned matching HTTP 404 JSON, so the descriptor is genuine rather than a catch-all. · retrieved 2026-08-23 - Anthropic Connectors Directory entry — , read from our directory snapshot dated 2026-08-16 · retrieved 2026-08-16
- Read AI support — · Privacy
Use in Agentman
Connect once and your agents call these tools on their own — on a schedule, in a workflow, with nobody at the keyboard.
Open in Agentman StudioServer Info
- Category
- Productivity
- Developer
- Read AI
- Tools
- 4
- Domain
- api.read.ai
Using Claude Desktop or another MCP client? Setup docs — the connection URL above works anywhere.