Unabyss
Self-updating context layer that every AI tool you use can read
About Unabyss
Unabyss is a context layer that sits between the tools where your work actually happens and the AI tools you ask about it. It connects to Slack, Gmail, Notion, GitHub, Linear, Obsidian, Google Drive, Google Calendar, OneNote, X and LinkedIn, plus meeting recorders including Fathom, Fireflies, tl;dv and Granola, pulls what is in them into one structured store that keeps itself current, and serves that over the Model Context Protocol. Claude, Cursor, ChatGPT, Codex, Perplexity, OpenCode, VS Code and any other MCP client then read the same context.
The problem it aims at is the one where your coding agent knows your repository but not the Slack thread where the decision was made, your chat assistant knows neither, and every new tool starts from nothing while you re-explain the same background for the fourth time. The part of the design that deserves attention is the access control: permissions are set per item and per topic rather than per integration, so you can expose a work persona to a coding agent while keeping personal threads out of it entirely. It is closed source and hosted, with SOC 2 Type II listed on their security page as in progress rather than complete, so the trade is real: one vendor holds a structured index of your working life, and in exchange you stop repeating yourself.
You sign up and connect accounts through OAuth, one integration at a time. Unabyss extracts and structures what it finds rather than storing raw dumps, tags it, and keeps it in sync as those sources change. You then connect an agent. There is an official Claude connector that finishes signup and onboarding inside Claude itself, and any other MCP-compatible client is a single configuration line. Once connected, the agent queries your context the way it would query any other MCP tool, and you decide at the item and topic level what each connected agent is allowed to see. The plan tiers are drawn around how much of this you do rather than around features: Pro caps you at three connected agents and twenty connected accounts, while Max removes both caps and adds multi-account support for Google apps.
- •More Than Twenty Native Integrations - Slack, Gmail, Notion, GitHub, Linear, Obsidian, Google Drive, Calendar, OneNote, X, LinkedIn and four meeting recorders
- •MCP Native - Any MCP client connects with one configuration line, and Claude has an official connector that handles onboarding in-app
- •Item and Topic Level Permissions - Decide what each agent can see instead of granting or withholding a whole integration
- •Self Updating - The context tracks its sources continuously rather than being a snapshot you remember to refresh
- •Structured, Not Raw - Material is extracted and tagged into one live context instead of dumped as unindexed text
- •Team Workspaces - Shared workspace, invitation by email, and central seat and invoice management on the Team plan
People whose work is genuinely spread across many tools and who run more than one AI assistant against it: a coding agent, a chat assistant, and something for research, none of which share what they know. It is built for individuals first, with a Team plan for a shared workspace at a three seat minimum. Two things to weigh honestly. There is no free tier, only a seven day trial, so the cheapest way to evaluate it is time-boxed rather than open-ended. And it overlaps Composio, which solves an adjacent problem from the other direction: if what you need is an agent that can act in external systems rather than read context out of them, Composio is the closer fit and it has a hard-capped free tier. Anyone whose work already lives in one place, or who runs a single agent against a single repository, gets little from a layer whose entire point is spanning many sources.
