Building CoWork OS: free open source agent harness for any LLM (API or sub). I ship tools, test new models, and build AI apps.
Lisbon
Joined February 2026
- Tweets1.8K
- Following451
- Followers311
- Likes964
Pinned Tweet
CoWork OS 0.5.61 in 95 seconds, recorded in the real app with a local model (qwen3.6:35b on Ollama).
Install in one command, bring your Hermes agent, approve every action, give bots their own computer, and see only what needs you.
👏👏👏
Jensen closed his GTC Berlin 2026 keynote with this Hermes Agent demo as the finale.
He introduced it by saying the way to build an agent is to ask an agent to build it for you, so he installed and set up a customized Hermes in one prompt.
Then he used Hermes, running on @nvidia RTX Spark, to create a simulated training environment for the @huggingface Microduck. Hermes planned an obstacle course with Nemotron and had GPT-6 Astra build it using FreeCAD, Omniverse libraries, Blender, Newton, and Isaac Lab. Hermes trained thousands of simulated ducks on that course until the validated policy was ready to run on the real robot in the physical world.
At Nous, we believe open source is a practical tool as much as a philosophical ideal. It gives users the ability to build powerful, complex workflows chaining numerous local+cloud models, devices, and tools themselves instead of waiting for a company to enable them.
This demo runs on qwen3.6:35b through Ollama.
What would you want to see it attempt next? Give me a real task, including the part where your current agent gets stuck.
CoWork OS 0.5.61 is out. The idea from day one: an AI desktop that never acts without showing you exactly what it will do first.
New in this one: install with one command and run your first task free, no API key.
Try it and tell me what breaks.
Pick one painful workflow and finish it end to end. Put it in someone's hands. Their first failure will teach you more about agents than another weekend spent tuning the setup.
Rooting for the OpenCode team. Being open with users when things go wrong matters. Wishing you a steady recovery.
We had a security breach last night related to a large scale distillation attack that impacted 1.4% of OpenCode Console users
We have notified them all this morning via email. If you did not get an email you were not impacted, it was a subset of users who had external providers linked
The incident is ongoing so we cannot share more details at this point but we are working with Anthropic, OpenAI, and Google to remediate
The next software demo should start with the app closed. Give an agent a task, interrupt it halfway, then inspect what actually changed. That would tell me more about readiness for agents than another chat box beside the dashboard.
==[AI Executive Update]==
I agree headless software is the new game
This 1 minute-ish Instagram reel makes the point well:
instagram.com/reel/DeWzqDXJ3…
My $0.02, based on a lot of hands-on time with each market leading AI agents, is that this is exactly where the market is heading.
The important question for software companies is no longer just:
"Is our user experience good?"
It is increasingly:
"Can AI agents easily interact with our system?"
That means clean APIs, strong permissions, clear data models, reliable automation, and rich MCP interoperability.
The exciting part is that this creates a BIG reset.
Legacy incumbents may have years of UI advantage, but if their systems are difficult for agents to interact with, that advantage can disappear quickly.
This is a real opportunity for smaller, faster teams to leapfrog competitors.
My takeaway:
Design for humans, but architect for agents.
That is becoming a very important distinction.
--.
==[Helpful Hint]==
Cool prompt to copy/paste:
I did this movie recommendation prompt last night and it worked out well :-). Give it a try to uncover a few gems to add to your “watch list”
--.
Create a Movie Marathon Around My Current Mood
Act as an expert film curator and design a personalized movie marathon that starts from how I feel now and gradually moves me toward the mood I want by the end.
Context:
- Current mood: [tired / nostalgic / anxious / romantic / restless / lonely / excited / etc.]
- Desired mood: [comforted / energized / inspired / relaxed / hopeful / emotionally satisfied / etc.]
- Genres I enjoy: [genres]
- Genres I dislike: [genres]
- Favorite films: [examples]
- Films I do not want repeated: [examples]
- Era preference: [classic / 80s / 90s / 2000s / recent / any]
- Language preference: [English / international / any]
- Time available: [X hours]
- Number of films: [2 / 3 / 4]
- Watching with: [solo / partner / friends / family]
- Streaming services: [platforms]
- Intensity tolerance: [light / moderate / heavy]
- Hard limits: [no horror / no sad endings / no explicit content / etc.]
Build the marathon so the sequence has an emotional arc, rather than just listing movies that fit the same mood. Start with something that meets me where I am, then gradually shift tone, energy, or emotional intensity toward the desired end state.
For each film, provide:
1. Why it fits
2. Why it belongs in that position
3. Runtime
4. Emotional tone
5. One thing to pay attention to
6. Optional snack/drink pairing
7. A spoiler-free one-line pitch
Avoid generic or obvious recommendations unless they are genuinely the best fit.
Finish with:
- One wild-card pick
- One shorter backup marathon
- One comfort-only version
- One more adventurous version
- A recommended start time so the marathon finishes by [target time]
===.
Switching agents shouldn't mean setting up every app again. But a shared login isn't a delegation system. Software needs to know which agent is acting, on whose behalf, and what this particular task allows it to do.
As a software provider (Endgame.io) I love MCP!
But the real reason I don't like MCP in my personal workflows is, I use way too many agents.
I work daily with Claude, Codex, Goated, ChatGPT, Muse, Shelley, Pi. Honest! I really do this.
and if I auth a CLI they can _all_ use it without me painstakingly setting each one up and wasting time. and then re-logging sometimes at random for a random agent.
beyond that, my _code_ can also use it just by shelling out to bash. so I can write scripts and crons that check things via a CLI and wake an agent in headless mode IFF the state requires it.
MCP simply can't do that and never will (unless it's CLI-wrapped, like the endgame CLI does dynamically for our MCP server)
My agent should be able to book dinner with yours without either of us joining another platform. It should share availability and dietary needs, not our entire histories. Open protocols need selective disclosure built in.
I read every word of @andrewchen's article and here’s what I think.
If we enter a future with no network effects from using the same agents, then no matter which agent I use, it can talk to Andrew’s agent (built by a different company) to schedule a meeting for us, decide if @a16z should invest in @waku_agent, or book dinner with other founders, as agents smoothly share all context on preferences and schedules. Horizontal agents would have no moat and would not dominate.
Businesses may soon realize they should demand their own personal or business agents, hosted or built in-house, assembling components from various providers. As long as agents can communicate, research, and execute via shared open protocols, it doesn’t matter if partners or family use different agents.
Agents communicate much faster in JSON than in lengthy human conversations, so they can execute in a fully outcome-driven way because they know who we are.
Most importantly, they can own their data which is imo the single most important thing for a business in an agentic world where you don’t need LinkedIn for candidates or ThomasNet for suppliers.
You should own your preferences, schedules, and tastes.
Whoever owns the context and daily behavior of individuals and businesses will stay relevant, even amid competition from large corps.
But if network effects exist and open protocols for agent diversity are limited, then centralizing context becomes the strongest moat for these tech giants, which I hope does not happen.
(Wrote the above myself, with the language slightly polished by @grok)
A CRM built for agents should let me say: follow up with these leads, spend nothing, and ask before sending a discount. If I still have to babysit every record change, we've automated the clicks and kept the old operating model.
For 25 years, Salesforce, SAP and enterprise software were built around one assumption: humans operate the software.
We’re changing that.
Introducing ERP•AI and Proto: business software built for agents, and an agent that operates business software.
Describe the system you need. ERP•AI builds it.
Then Proto takes over. It manages records, runs workflows and executes actions inside the software you just created.
This is Headless SaaS.
1: get ERP•AI
2: build your SaaS
3: post below
Best one wins a MacBook. First 100 decent entries get $100.
(a message from Steve Papa / founding team)
The interesting question is why the form exists. If two agents already have the information, making one recreate a human form for the other is an odd detour. We need shared schemas and consent, with a form available when a person wants it.
just built formeasy — an open-source, agent-native alternative to google forms.
AI agents can create forms and collect responses using AgentOnboard. no browser needed.
MIT licensed. free to use.
try it: formeasy.agentonboard.xyz
Headless commerce doesn't make logistics disappear. But delivery promises, return policies and seller guarantees need to become things an agent can compare and enforce. The storefront can move. Responsibility still has to live somewhere.
Giving an agent its own computer is the beginning. What happens when the model changes, the machine restarts or the task moves to another provider? The work, permissions and evidence should survive. Otherwise we're rebuilding cloud lock-in around agents.
If your next customer is an agent, your homepage may matter less than your error messages. Can it understand the price, test one useful action and recover when it fails? That's a very different sales funnel from booking a demo.
New!! Going Headless: Claudeforce vs. Replit vs. Lovable, Agents Picking Our Vendors, and Two of Our Agents Built Their Own Inbox to Talk Behind Our Back: The Agents #016
Same data, same prompts. Bills: $2 to $15.
Our agents now pick our vendors.
saastr.com/going-headless-cl…
everyone at some point should have their own harness I'm afraid :( So sad to hear this!
We had a security breach last night related to a large scale distillation attack that impacted 1.4% of OpenCode Console users
We have notified them all this morning via email. If you did not get an email you were not impacted, it was a subset of users who had external providers linked
The incident is ongoing so we cannot share more details at this point but we are working with Anthropic, OpenAI, and Google to remediate
Why am I paying for five seats when one agent does the work?
Headless apps challenge more than the interface. They challenge pricing built around humans logging in.
If agents become the main users, what exactly are software companies selling?