autoMate: an MCP tool source that keeps notes, files and memory on your own disk
Like Manus, Computer Use Agent(CUA) and Omniparser, we are computer-using agents.AI-driven local automation assistant that uses natural language to make computers work by themselves
At a glance
- What is it?
- autoMate is a local FastAPI service that exposes notes, a file vault, reminders, memory and 40-plus tools to any MCP-capable chat client. The interesting part is the storage layer, not the agent loop.
- Who is it for?
- Adopt autoMate if you already run an MCP-capable client and want notes, files and memory in one SQLite database under ~/.automate/ that survives switching clients. Do not adopt it as a general computer-use agent for unattended work, and do not adopt it if you need cross-machine sync, since the README describes a single local store.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 21 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The gap autoMate targets: tool calls that do not outlive the chat
Chat clients can call tools. What they generally cannot do is hold state that a different client can read tomorrow. The README states the problem directly: most vendors can call tools, but none of them remember anything across vendors, store your files, or ping your phone. autoMate is positioned as the layer behind the chat, described as a warehouse you own, with the storage sitting in ~/.automate/ as SQLite plus Fernet encryption.
The intended user is someone who already has a preferred client (OpenClaw for IM, Claude Desktop for work, Cursor for code) and wants the same notes, files and facts available from whichever one is open. That is a narrower audience than the repository topics (agent, computeruse, rpa, manus) suggest. The topics advertise a computer-using agent. The README v4.6 describes a persistent personal data store with an agent loop attached. Those are different products, and the mismatch is the first thing to understand about the project.
The data model: notes, a content-addressed vault, and key-value memory
Four namespaces carry the personal data. notes.* holds markdown documents with tags, search and pinning. files.* is a content-addressed blob vault with deduplication and a configurable storage path, which the README says can point at an external SSD or NAS mount. memory.* is described as long-term key-value facts that any AI can read and write. reminders.* is backed by a scheduler thread that fires Web Push to the installed PWA.
retrieval is a single call. search.find is documented as Coze-style hybrid retrieval over SQLite FTS5 BM25, spanning notes and files together. FTS5 with BM25 is a lexical ranker, not a vector index. The README calls the retrieval hybrid but does not describe a second ranking stage, so treat search quality as keyword-dependent until you check the code. Content-addressed storage for files is the strongest design decision here: identical uploads collapse to one blob, and moving the vault to another disk is a path change rather than a migration.
Install paths and the first run
Four install routes are documented. pip install automate-hub for the Python package. A standalone binary for Windows, macOS and Linux from the releases page, for machines without Python. Docker with docker run -p 8765:8765 ghcr.io/yuruotong1/automate:latest, aimed at a headless box or NAS. And a browser extension in extension/ for driving your real Chrome.
After any of those, the command is automate, and the README notes that double-clicking on Windows or macOS does the same thing. A browser opens to http://127.0.0.1:8765 and a wizard walks through model selection, key entry, and optional client wiring.
The client side is where autoMate differs from most agent projects. Settings then Connect to AI clients offers a Copy install text button that produces a single markdown blob with the MCP URL and bearer token already filled in, plus per-client sections for OpenClaw, Claude Desktop, Cursor, Cline, generic MCP and non-MCP gateways. The README gives three ways to use that text: edit the client config by hand, paste it into another AI and ask it to configure itself, or for OpenClaw paste it under bundle-mcp. Once the client picks up the server it gets the individual tools plus a top-level automate tool that runs autoMate's own agent loop on demand. That last detail is the real architectural point: the same backend is reachable as a flat tool list or as a single delegated agent, and you choose per client.
Where the design gets thin: executors, IM channels, and the Pro tier
The local executors are the part to think hardest about. shell.*, script.* (Python, Bash, Node), desktop.* via pyautogui, browser.* via Playwright, and bx.* which drives your real browser through the Chrome extension. The README lists them without describing a confirmation step, a sandbox, or an allowlist. If a remote model can call shell.* and desktop.* through an MCP server on your machine, the trust boundary is the model provider, not autoMate. Nothing in the supplied material says otherwise, and that is a gap worth resolving before you point a hosted model at it.
IM support is deliberately not built in. The README says the project does not ship per-platform bots and instead expects you to run OpenClaw alongside, then drop an auto-generated bundle-mcp snippet into ~/.openclaw/openclaw.json5, run openclaw config set channels.wechat.* for apiKey, proxyUrl and webhookHost, and start the gateway with openclaw gateway start before scanning a QR code. The legacy adapters in automate/bots/ (telegram, wechat_oa, wecom) still exist as direct-webhook options and are labelled advanced and legacy in the UI. That is an honest arrangement, but it means the messaging story depends on a second project you also have to operate.
The tiering is also only half-documented. audio.transcribe is marked Pro in the README, and auth.py is described as an autoMate Cloud session with a Pro tier hook. Which of the other 40-plus tools sit behind that hook is not stated. Check before you plan around a tool.
autoMate against a plain MCP filesystem server
The obvious alternative is the reference MCP filesystem server, or a notes app with its own MCP endpoint. The difference is scope of state. A filesystem server exposes directories and nothing else: no notes with tags, no key-value memory, no reminder scheduler, no FTS5 index across notes and files, no single search.find call. If your need is read and write access to a folder, autoMate is more machinery than the job requires, and a lighter server will be easier to audit.
If your need is the opposite, a fact you want every client to know, a file you want indexed alongside your notes, and a push notification when something is due, then a filesystem server cannot express any of it. The trade is that autoMate asks you to accept an agent loop, a set of executors and a cloud session hook in exchange for that state layer. The comparison to make is not autoMate versus another computer-use agent. It is autoMate versus the combination of a filesystem MCP server, a notes app with its own API, and a separate scheduler, which is three integrations instead of one.
Maintenance, versioning and the MIT licence
Release cadence is visible and fast: v4.5.9, v4.5.10 and v4.6.0 all landed on 2026-04-28, and the last push to the default branch is 2026-08-25. Three patch-level releases in one day suggests either rapid iteration or release churn, and the material does not distinguish between them. The README itself anchors client-connection behaviour to v4.5.7, so the install text and the MCP wiring are recent additions rather than long-settled interfaces.
Upgrade cost concentrates in three places: the client config snippet, which you may need to regenerate after a version bump because the URL and token are embedded in it; the SQLite store under ~/.automate/, which is your data and should be backed up before any major upgrade; and the 25-provider LLM catalog plus 31 SaaS connectors, which are the parts most likely to need attention when an upstream API changes. Deduplication in the file vault means a backup is smaller than the sum of what you uploaded, but it also means a corrupt blob affects every note that references it.
The licence is MIT, which permits commercial use, modification and redistribution with the copyright notice and permission notice retained. That is the extent of what can be said here; whether the Pro tier, the autoMate Cloud session in auth.py, or the bundled SaaS connectors raise separate terms is not answerable from the README, and this is not legal advice.
Who should install it, and what to check first
Install it if you run an MCP-capable client already and you want one local place for notes, files, reminders and cross-session facts, with the option to hand the whole agent loop to a client as a single automate tool. The pip install automate-hub route plus the Copy install text button is the shortest path, and the configurable files.* storage path is what makes it workable on a NAS or an external SSD rather than only on a laptop.
Skip it if you want unattended computer use. The repository topics promise computeruse and rpa; the README describes executors with no documented confirmation step and no sandbox, which is not a foundation for letting a model click through your desktop unsupervised. Skip it as well if you need the same store on two machines, because the material describes a single local ~/.automate/ store with no sync mechanism.
Two things to verify before committing. First, whether the tools you depend on are Pro-gated, since only audio.transcribe is explicitly marked as such. Second, whether your client version handles MCP-over-HTTP with a bearer token, because that URL plus token is the entire integration surface. If both check out, the store is yours, it is SQLite, and switching chat clients later costs you one config paste.
Editorial conclusion
Adopt autoMate if you already run an MCP-capable client and want notes, files and memory in one SQLite database under ~/.automate/ that survives switching clients. Do not adopt it as a general computer-use agent for unattended work, and do not adopt it if you need cross-machine sync, since the README describes a single local store. Before installing, verify two things: that the tools you actually need are not gated behind the Pro tier (audio.transcribe is documented as Pro), and that your client version supports MCP-over-HTTP with a bearer token, because the whole integration is a URL plus a token.
Community notes