What shipped, and what it proves.

Each dashboard action is a tool the agent can propose, calling the same code as the button, with the same checks and the same audit record, labelled as the agent acting for you. Nothing changes until you confirm it, and after each confirmed step it carries on to the next. Links with a token in them are shown to you and never to the model, and credentials are never passed through a conversation.

The agent can write to a teammate or to a customer's registered contact, and to nobody else: it composes the words, Meetext decides the address. Every email is a draft you read in full before it is sent. With replies set up, a customer's answer comes back into the conversation and the agent drafts a reply, looking only at that customer's information.

Ask for a recurring check, such as a morning summary of which customers need attention, and the agent runs it by itself and emails who you named. Each run acts with its creator's permissions as they are at that moment, and proposes any change rather than making it.

On the Team page, single permissions can be turned on or off for one teammate beyond their role, and an invitation can carry them. The API, the dashboard and the agent all read the same answer, so turning something off for a person turns it off for the agent working on their behalf.

Ask which customers need attention, what one of them costs this month, or what an incident means. It answers with tools over the same records the dashboard shows, and says where each fact came from. It acts with your role and never beyond it, and a change it proposes appears in the conversation with Confirm and Cancel: nothing runs until a person presses one. On its first live run of five eval tasks it chose the right tools every time, for under five cents in total.

On the customer's portal it sees only that customer's installations. It can read a pasted screenshot of an error, explain what is waiting for approval, and propose pausing an installation. It cannot approve or deny anything on their behalf.

A capability that deletes data now waits for a person in the customer's workspace every time, and on new installations a write waits too. The request appears in the Slack, Teams or Google Chat thread that asked, and on the customer's portal. The person who asked can confirm a write; a deletion needs somebody else. An approved call runs exactly once however many times the button is pressed, and the vendor can see every decision but cannot make one.

Email addresses, phone numbers, US social security numbers, payment cards that pass the Luhn check, IBANs that pass their checksum and IP addresses are masked in every tool result, whichever client made the call, under a policy the customer sets. Optional tokens let the model act on a masked value without ever seeing it: the value is put back only inside Meetext, just before the call runs, and the token expires after an hour. Google Cloud Sensitive Data Protection can be added for names and street addresses.

Each customer can have a monthly budget for the hosted model, checked before every model call rather than once per answer, with alerts at 80% and 100%. Either side can pause a deployment at once; only the side that paused it can resume it. Evaluation runs are now metered and capped like every other model call.

When an incident opens, the agent reads its recorded facts and writes a likely cause and a next step, shown as its unverified reading. It never changes whether a fix is attributed, and it never acts.

Each request carries a W3C traceparent through to the vendor's own server, and the capability record keeps the trace id, so a row in Meetext and a line in the vendor's logs can be joined without guessing at timestamps.

A question asked in a customer's chat tool is resolved to exactly one installation, sent to an endpoint the vendor runs, and the reply is relayed into the same thread. Streaming and single replies both work: a streamed answer is posted on the first chunk and the same message is edited as it grows, rather than posting fragments or waiting in silence. An answer that fails partway keeps what arrived and appends the reason, and the vendor's own error text goes to their event log rather than into their customer's channel.

Every question carries that customer's own MCP endpoint and token, so the vendor's assistant can act inside the workspace with the customer's credentials, bounded by the package frozen at their install and by what they are entitled to. Conversation history is neither sent nor stored: a stable thread reference goes instead, and the assistant keeps its own memory against it.

An assistant Meetext runs, given the same per-customer endpoint as its tools, so it answers from capabilities the vendor approved and the customer bought rather than from general knowledge. The name on the bot is still the vendor's.

A control on the product page sends a question through the same customer-scoped path and returns the reply in the dashboard, reporting latency, streaming behavior and which customer's capabilities it received. Teams can verify an assistant before inviting customers into the rollout.

Every page and every main component carries an icon that says, in plain words, what the thing is and what to do with it. Written for somebody who has not yet met the words entitlement, canary rotation or confounded incident, which is most people on their first day.

Salesforce, ServiceNow, Atlassian Cloud and HubSpot now follow the same deployment contract as Slack and Microsoft. The implementation accounts for the details that matter in production: Salesforce field visibility, per-instance ServiceNow registrations, Atlassian site access and renewed HubSpot consent after a scope change.

Every destination exposes installation, permissions, validation, credential lifecycle, versioning, failure handling and monitoring through the same product model. Provider-specific requirements stay visible without leaking internal release process into the customer experience.

OpenAPI reference resolution, repository authorization detection and discovery error messages were tightened so large specifications produce useful capabilities and failures explain what the developer can fix.

The privacy policy explains what is stored, how credentials are isolated, and what deliberately is not retained. The terms define account security, acceptable use, customer responsibilities, service availability and termination in plain language.

A document whose uniqueness mattered gets an id derived from the fields that made it unique, so a second write with the same identity updates one document instead of creating a duplicate.

Ids are time ordered and monotonic within a millisecond, so two facts recorded in the same instant read back in the order they happened rather than reshuffling between reads.

Sign in, create an organization, and reset a password in one page. Organization names are validated rather than silently slugified into a counter.

Observed by Meetext, derived by Meetext, asserted by the vendor, or observed by the customer. Every renderer preserves the distinction, so an inference can never be read as an observation.

A sign-off covers the exact posture the reviewer saw. When the posture moves, the review goes back into renewal instead of quietly covering whatever shipped next.

Share links are hashed, expiring, and revocable. Revoking one kills the page without touching the snapshot it pointed at.

Who owns the next move is a different question from what is blocking progress, and collapsing them made a waiting customer look like a broken deployment.

Time nobody can account for is reported as unattributed rather than distributed across the parties. A coverage figure says how much of the total is explained.

A pattern is only reported as verified once it has been checked at least five times and held at least eighty percent of them.

A client's own connector check verifies that it can reach the endpoint and list tools. That is the protocol layer and nothing more. It used to be reported as health.

An outage at Slack degrades the deployment and says so, rather than reporting the vendor's own product as broken.

A deployment is healthy when a real read capability ran against the vendor's system, not when a token exists.