Attensira changelog

The agent connects to a repository through a GitHub App rather than a personal token, and opens its changes as pull requests you review and merge. This is the path for sites and docs whose content lives in a repo.

Workspace API keys can be created and revoked from the account area, and authenticate requests to the public REST API. Keys are scoped to the workspace that issued them.

The agent's proposed work now lands in a queue instead of executing silently. Each item can be approved or declined, and the API exposes the same list, approve and decline operations the screen uses.

Credit balances can be topped up by a custom amount, and a rule can refill the balance automatically when it falls below a threshold you set. Refunds are recorded as a first-class ledger operation rather than an adjustment.

The metric was renamed everywhere — UI, API and docs — because "share of voice" implies a share of a fixed total, and this number is not one. It is the proportion of successful runs in a window that named your brand, computed against the same denominator as every competitor's rate. The rates are independent and do not sum to 100%.

Crawler-fetch data is now one of the tools the agent can call, so it can tell an access problem from a content problem: a page no AI crawler has fetched cannot be cited, and that is a different fix from rewriting it.

Scheduled readings are covered by the plan's prompt slots and response budget. Credits are spent on agent work, not on the measurement that runs on a schedule.

The tracking vocabulary was renamed across the product to match what the objects actually are: you track prompts, group them into topics, and read the result per platform. Old URLs and API fields were migrated in the same change.

Server-side collection records which AI crawlers fetched which pages. AI crawlers do not execute JavaScript, so a client-side beacon never saw them; the product generates the install for the server-side path.

Facts about your business can be stored as knowledge and composed into the context the agent works from, so the same correction does not have to be repeated in every session.

The same capabilities as the MCP server, over plain HTTP, authenticated with a workspace API key.

Plans are metered on how many prompts you track and how many model responses are collected against them, with sampling depth as the difference between tiers, rather than on a single opaque allowance.