YC-backed Almanac launches an always-on AI agent with its own computer and a self-updating company wiki

An agent you text like a colleague

Almanac, a startup from Y Combinator's S26 batch, has launched an AI assistant built to be delegated to over ordinary messaging. According to the company's launch post on Hacker News on August 31, 2026, the agent runs continuously on a machine of its own, stays signed into the tools you connect, and messages you over text or Slack when a job is finished.

The launch page leans on workplace vignettes to illustrate the idea. In one, an employee asks the agent to sweep a customer support channel on Slack and open GitHub issues for newly reported bugs; it replies with progress updates and eventually files two issues with screenshots and reproduction steps, linking them back to the original conversations. In another, someone asks what the team had promised a customer on pricing, and the agent answers with the commitment, the date it was confirmed by email, and a warning about a related, still-open bug — then drafts a briefing and sends it to a colleague.

The wiki is the memory

The heart of the product is a wiki that maintains itself. Almanac connects to services such as Gmail, Slack, and meeting-notes apps, and compiles what it observes into pages describing a company's people, customers, and projects. Before carrying out a task, it reads that wiki, which the company presents as the key difference from a chatbot plugged into integrations: an integration fetches data only when asked, while Almanac keeps a durable record and starts every task already caught up.

The wiki is also meant for humans. Every line links back to the source it came from, so any claim can be checked, and an edit or correction on a page becomes part of what the agent knows from then on.

A computer, not just API calls

The second design choice is infrastructure. Almanac runs on its own hosted computer with a browser, files, and logins of its own. For tools that offer no integration, the agent signs in and operates the software much as a person would. Because the machine is always on, work continues after you close your laptop — the company contrasts this with running a local model or coding agent yourself, which stops the moment the lid shuts.

The agent is also meant to act unprompted: using its connections and the wiki, it notices work worth doing, does it, and tells you. One example on the site has it spotting unreceipted charges in a billing tool and attaching the missing receipts itself.

Privacy and guardrails

On privacy, the company says an account you connect stays usable only by you; what flows into the shared wiki is conclusions and decisions rather than raw inboxes, and shared accounts must be added explicitly by the organization. Every connection is visible and can be revoked.

On autonomy, Almanac works on what you hand it and on tasks it identifies on its own, but it pauses at logins, payments, and decisions it should not make alone — either asking first or handing you its live browser. Each run can be watched step by step. The company positions the product as sitting alongside Notion, Linear, and Slack rather than replacing them, acting across the tools a team already uses.

Why it matters

Every capability described here comes from a single launch announcement, so these are the company's claims rather than independently tested results. Even so, the design shows where the agent category is heading. Persistent, auditable memory in the form of a source-linked wiki attacks the most common complaint about assistants: they forget everything between conversations. Handing the agent an actual computer, rather than a fixed set of API hooks, extends its reach to the long tail of software that will never ship an integration. And the explicit pause points before logins, payments, and judgment calls suggest vendors have learned that autonomy inside a company only sells when it comes with visible brakes. Whether colleagues accept an agent reading their Slack and mail, and whether the wiki stays accurate as a company grows, will decide whether this model holds up.