Six weeks after OpenClaw crossed 300,000 GitHub stars, Google walked onto the I/O stage and announced Gemini Spark. The two events are not actually connected in any technical sense, but they have been connected in my head all week, because I have been running OpenClaw on my own machine since the 280K mark and I have some thoughts about what the Spark announcement means for people like me who already have a local agent stack that they trust.
Spark is, on paper, a 24/7 personal agent that runs on Gemini 3 (Google’s latest large language model, the one that finally started beating benchmarks that used to be reserved for human experts) and lives in the Google app on your phone. It is supposed to be able to take actions across your apps, schedule meetings, summarize your inbox, and remember things about you for weeks. The demo was the usual I/O demo. The latency was good, the voice was natural, and every developer in the room pulled out a phone and tried it on the way home.
I have been on OpenClaw since late March. It is a 14,000-line Python codebase with a small Go runtime, and it is the first open-source agent (a long-running AI process that takes actions on your behalf, not just answers questions) that I have not eventually abandoned. I picked it up because I was tired of the “we have a privacy-first agent” SaaS pitches where privacy-first meant “your data lives in our SOC 2 datacenter in Virginia.” OpenClaw runs on a Raspberry Pi 5 I had in a drawer, it talks to my local Ollama instance (a tool that runs large language models on your own hardware), and it does about 60% of what Spark promises to do. The other 40% is the part I want to talk about.
What OpenClaw actually does well
The thing OpenClaw gets right is the part nobody else even tries: it is honest about what it does not know. Every action it takes comes back with a confidence score, the source it pulled the answer from, and a one-sentence plain-English explanation of why it took that action. When I asked it last week to schedule a meeting with my accountant, it paused, told me it could not find a calendar entry that matched what I described, and asked me to confirm the name. Spark, in the demo, just scheduled the meeting. I am not sure which behavior I would prefer at 9 PM on a Friday, but I am sure I want the option.
Three other things OpenClaw does well that the announcement made me appreciate more than I had a month ago:
- It runs offline. I have a cabin in the woods with no cell signal. OpenClaw still works there. Spark does not.
- It is auditable. Every action is a log line in a SQLite file I can read with
sqlite3 ~/.openclaw/actions.db. I have caught two bugs in my own configuration this way, both involving a tool that was supposed to delete a draft email and instead deleted a real one. - It is small enough to read. The whole codebase is about 14,000 lines of Python and a small Go runtime. I have read most of it. The first time I read a line of code from an agent that was about to take a real action on my behalf was a small revelation.
- It is a community project, not a product. There is a Discord with about 8,000 people in it, a maintainer who answers questions at 2 AM his time, and no roadmap. The roadmap is whatever 30 people decide to build next month. This sounds chaotic. It is, in fact, chaotic. It is also the only reason the project has not been bought by a large company and turned into something with a free tier that gets worse every year.
What OpenClaw does not do well
This is the part where I have to be honest. OpenClaw is not Spark. It is not even close.
Gemini Spark has voice in and voice out. OpenClaw has neither. The voice part is a small thing in theory, but in practice it is the difference between “I asked my agent to do a thing” and “I had a five-minute conversation with my agent about a thing.” For the kind of scheduling triage Spark showed in the demo, voice is the whole product.
Memory is the other gap. OpenClaw remembers things for a session, and it remembers things you explicitly tell it to remember, and that is it. Spark, by Google’s account, remembers things about you across sessions for weeks, and it does the “this is what I learned about you from last week” summary that OpenClaw has been promising for a year and has not shipped. Memory is the kind of feature that sounds boring in a demo and turns out to be the entire reason people stay with a product.
Tool coverage is the third gap. OpenClaw has about 40 tools, mostly the ones the core team uses: calendar, email, file system, shell, web search, GitHub, Notion. Spark has whatever Google’s first-party integrations team has built, which is a lot more, and it has the agent-to-agent protocol (a standard that lets agents from different companies talk to each other) that was announced at I/O. OpenClaw does not have that.
What Spark is, and what it is not
Spark is a personal agent for people who live in the Google app. It is good. I tried it for a weekend on a fresh phone with a fresh Google account, and the email summarization alone is worth the price of admission, which is currently zero, because the free tier is the whole product. The voice is the best I have used. The memory is actually good, not the “we store your data and call it memory” thing that other vendors do.
People who do not want to live in the Google app are going to have a different experience. If your email is in Fastmail, your calendar is in iCloud, your files are in a Synology NAS (a network-attached storage box that sits in your house and runs Linux), and you are not going to switch any of those things, Spark is going to be a worse fit for you than it is for the I/O demo audience. The Google app is the substrate. You can build outside the substrate, but the substrate is doing a lot of the work.
The data question is a separate one, and it is the reason OpenClaw exists. Every action Spark takes is a Google API call. Every memory it has is stored on Google’s servers. This is fine for most people, and I am not going to tell you it is not. It is also why 300,000 people starred an open-source alternative.
Trade-offs
OpenClaw is not free in time. Setting it up the first time took me a Saturday and most of a Sunday. I have spent about 4 hours a month on maintenance since then, plus another full day in April when I upgraded Ollama and broke the embedding model (the AI component that turns text into numbers the system can search by) that OpenClaw depends on. I am not a systems person by training, so the maintenance is real to me; an experienced Linux user would do it faster.
Money is the second cost. OpenClaw itself is free. The Raspberry Pi was 80 dollars. The SSD I put in it was 60. The UPS (uninterruptible power supply, a battery that keeps the box running when the power flickers) was 40. Power is about 4 dollars a month. None of this is Spark-scale money, but it is not zero, and the math stops working if you want to run a 70-billion-parameter model (a large AI model that needs a real GPU, not a Pi) on it.
Features are the third cost, and the biggest one. I gave up voice in and voice out. I gave up cross-session memory. I gave up most of the agent-to-agent stuff. I gave up first-party integrations with anything I do not already use. For the things I do use, the integrations are good. For the things I do not, they are not there.
Upside is real too. OpenClaw runs on hardware I own, in a place I choose, with a code I can read. I can take the SD card out of the Pi and put it in a drawer and my agent stops existing, on its own, with no account to delete, no email to send, no data export to wait 30 days for. That is the trade I am making, and the trade I will keep making for as long as the project stays alive.
Bottom line
People who already use the Google app for email, calendar, files, and the rest of it should install Spark. There is no open-source project that will give you a better experience in that environment this year, and the price is right. Set the privacy settings to whatever you are comfortable with, and use it.
People who do not live in the Google app, or do not want to, should look at OpenClaw. It is a real option in 2026, not a curiosity. It will not match Spark on voice, memory, or first-party integrations. It will match it on auditability, portability, and the small daily things that an agent can do on a Pi that costs less than a year of Netflix. Install it on a weekend you are willing to lose, expect to read a few hundred lines of Python, and keep your expectations grounded.
Anyone who is somewhere in the middle should wait six months. The open-source agent space is moving fast, and the gap is going to narrow. Spark is not the last word from Google, and OpenClaw is not the last word from the open-source side. Pick a side when you have to, not before.