Coworking for Punks
Your personal AI coworker, minus the velvet rope
There’s a lot of excitement about “cowork” – a product from from Anthropic, but also a concept: using a local agent for non-coding, “knowledge” work. The stuff you do on your computer, now with dedicated assistance from an intelligent agent.
If you use Claude Cowork and like it, that’s great. Despite some limitations and instability, it is a good gateway to the world of agents for non-coders. But if you’re ready for something even better, I’ve got an alternative recipe you might prefer. It requires a bit more work to set up, but if you’re up to it, it’s worth the hassle.
OpenCode Desktop
OpenCode, the premier free, open, flexible, and beautiful local agent, has a desktop app that is a pleasure to use. I’m using it as my main driver for coworking, and I love the polished UI, ease of use, and relative flexibility and configurability.
In comparison to tightly controlled products like Claude Cowork, OpenCode desktop offers a choice of AI models and providers, permission modes, and local configuration. I love that I can store skills in the standard ~/.agents/skills directory and install them conveniently with npx skills, and that I can set the agent to default to “YOLO” mode, where it doesn’t ask me for permissions (the latest models are good enough to be trusted with most operations without constant babysitting).
GPT-5.4 (and other models)
There’s no question, OpenAI’s latest model, GPT-5.4, is far ahead of anything we’ve ever seen for agentic operation. Anthropic’s Claude models reigned supreme for much of the last year, but with GPT-5.4 we’ve got a model that is more intelligent, precise, and capable. Being able to switch to the latest best model is a huge advantage we lose when using the products of one of the model vendors. Using a flexible open-source agent sets us up for making the most of the fast developments and fierce competition in the model market.
I have access to GPT-5.4 via my $20/m ChatGPT Plus subscription - OpenAI are providing a generous allowance to subscribers - which is officially and legally supported in OpenCode. And using my $49/m GitHub Copilot Pro+ subscription I have access not just to GPT-5.4 but also to Anthropic’s Claude and Google’s Gemini models. Like OpenAI, GitHub chose to officially and legally support the use of the Copilot subscription in OpenCode and other open-source products. OpenCode’s own $10/m Zen Go subscription provides access to the best open-weights models, or pay-as-you-go access to any of the popular models.
…
Looking to learn how to use AI agents effectively for software development? The spring cohort of Elite AI-Assisted Coding, the #1 comprehensive course on making the most of agentic software development, is open for signups, and we’re offerring a 25% early bird discount until March 23rd. Join us!
…
Connectors
For many tasks, it’s helpful to have access to the various apps and stores of knowledge you rely on. Developments in MCP servers, CLI helper programs and, most importantly, agent skills, give us lots of choice.
Here’s what I’m using:
Google Workspace CLI + Skills – works great with my Gmail, Calendar, Drive, Sheets, Docs, Contacts, and other facilities in Google’s suite. The initial setup is a bit involved, but the CLI does its best to guide you through the process, and works well with both individual Gmail accounts and work accounts.
Alternatives: if you’re using Microsoft 365 you may be able to use their Work IQ MCP server similarly. Notion has an MCP server you can connect to if you’re using it as your main workspace.
Todoist CLI + Skill – I use Todoist for task and project management, and my agent works great with the CLI tool released recently and a skill I have created.
Alternatives: most task and project management apps have released MCP servers or CLI helpers for connecting to them from an agent, or if not and they have a usable API, someone will surely have released a solution. Find the relevant connector and corresponding skills and install them.
Agent Browser + Skill – this is the best solution I found yet for letting my agent browse the internet and do things. It’s fast and unintrusive and very capable.
Obsidian + CLI + Skills – I use Obsidian for editing and managing a collection of knowledge, including drafts, project plans, daily journaling, meeting notes, contacts. The Obsidian CLI and skills let me drive a lot of that work agentically, run complex searches, synthesise, edit, records, rearrange.
QMD + Skill – companion utility for working agentically with Obsidian or any collection of text files. It efficiently indexes the files and runs semantic or keyword searches on them.
Note that while some connectors are available as MCP servers, I don’t actually use any. All of my connectors are CLI utilities and/or agent skills. That’s a matter of personal preference, not an ideology – in many cases MCP servers are a great solution too, and they are easy to install and configure in OpenCode.
…
You don’t need to be a hardcore techie to set up your own personalised agentic AI environment. If you’re looking to get advice, work through some issues, and see how others work with agents, join an upcoming free Sunday School session – drop in, vibe on.
…
Skills
The primary extension and specialisation mechanism for agentic coworkers is Agent Skills. I have amassed quite the collection, both downloaded and home-cooked. Here are a few I find especially useful for coworking:
playground and artifacts-builder let me spin up ad-hoc web interfaces for exploration, learning, review, and planning. I’m creating them at a rate of a few dozen a day.
remotion-best-practices helps me create nice-looking videos agentically. I am completely useless when it comes to visual work, so being able to rely on an agent for this kind of work is significant.
nano-banana-2 and gpt-image-1-5 for creating and modifying images.
slidev helps me create presentation slides programmatically. See my demo of using this skill alongside nano banana to create a presentation.
tavily performs web searches. I prefer this to any of the built-in or extension web searches for agents I have found so far.
anki-connect and flashcard-generator are how I manage and continuously extend my growing collection of flashcards for spaced repetition (I currently have over 20,000 cards in my collection and I review an average of 100 a day).
raindrop-api connects to my bookmarks collection on raindrop.io for searches, saving new bookmarks, and managing the growing collection.
upstash-redis-kv serves as my central, cross-project key:value store in Upstash Redis. For most projects a local file is the easiest durable store to manage, but in some cases I just want to put something in memory and be able to recall it later.
docx, pdf, pptx, xlsx, markdown-converter for creating, reading, and managing all these files.
youtube-transcript because so much interesting knowledge comes in the form of YouTube videos and podcasts these days.
…
To learn more about Agent Skills, the most important extension and specialisation mechanism for AI agents, join an upcoming Mastering Agent Skills workshop, or follow the free Introduction to Agent Skills series.
to the world of agents for non-coders. But if you’re ready for something even better, I’ve got an alternative recipe you might prefer. It requires a bit more work to set up, but if you’re up to it, it’s worth the hassle.
OpenCode Desktop
OpenCode, the premier free, open, flexible, and beautiful local agent, has a desktop app that is a pleasure to use. I’m using it as my main driver for coworking, and I love the polished UI, ease of use, and relative flexibility and configurability.
In comparison to tightly controlled products like Claude Cowork, OpenCode desktop offers a choice of AI models and providers, permission modes, and local configuration. I love that I can store skills in the standard ~/.agents/skills directory and install them conveniently with npx skills, and that I can set the agent to default to “YOLO” mode, where it doesn’t ask me for permissions (the latest models are good enough to be trusted with most operations without constant babysitting).
GPT-5.4 (and other models)
There’s no question, OpenAI’s latest model, GPT-5.4, is far ahead of anything we’ve ever seen for agentic operation. Anthropic’s Claude models reigned supreme for much of the last year, but with GPT-5.4 we’ve got a model that is more intelligent, precise, and capable. Being able to switch to the latest best model is a huge advantage we lose when using the products of one of the model vendors. Using a flexible open-source agent sets us up for making the most of the fast developments and fierce competition in the model market.
I have access to GPT-5.4 via my $20/m ChatGPT Plus subscription - OpenAI are providing a generous allowance to subscribers - which is officially and legally supported in OpenCode. And using my $49/m GitHub Copilot Pro+ subscription I have access not just to GPT-5.4 but also to Anthropic’s Claude and Google’s Gemini models. Like OpenAI, GitHub chose to officially and legally support the use of the Copilot subscription in OpenCode and other open-source products. OpenCode’s own $10/m Zen Go subscription provides access to the best open-weights models, or pay-as-you-go access to any of the popular models.
…
Looking to learn how to use AI agents effectively for software development? The spring cohort of Elite AI-Assisted Coding, the #1 comprehensive course on making the most of agentic software development, is open for signups, and we’re offerring a 25% early bird discount until March 23rd. Join us!
…
Connectors
For many tasks, it’s helpful to have access to the various apps and stores of knowledge you rely on. Developments in MCP servers, CLI helper programs and, most importantly, agent skills, give us lots of choice.
Here’s what I’m using:
Google Workspace CLI + Skills – works great with my Gmail, Calendar, Drive, Sheets, Docs, Contacts, and other facilities in Google’s suite. The initial setup is a bit involved, but the CLI does its best to guide you through the process, and works well with both individual Gmail accounts and work accounts.
Alternatives: if you’re using Microsoft 365 you may be able to use their Work IQ MCP server similarly. Notion has an MCP server you can connect to if you’re using it as your main workspace.
Todoist CLI + Skill – I use Todoist for task and project management, and my agent works great with the CLI tool released recently and a skill I have created.
Alternatives: most task and project management apps have released MCP servers or CLI helpers for connecting to them from an agent, or if not and they have a usable API, someone will surely have released a solution. Find the relevant connector and corresponding skills and install them.
Agent Browser + Skill – this is the best solution I found yet for letting my agent browse the internet and do things. It’s fast and unintrusive and very capable.
Obsidian + CLI + Skills – I use Obsidian for editing and managing a collection of knowledge, including drafts, project plans, daily journaling, meeting notes, contacts. The Obsidian CLI and skills let me drive a lot of that work agentically, run complex searches, synthesise, edit, records, rearrange.
QMD + Skill – companion utility for working agentically with Obsidian or any collection of text files. It efficiently indexes the files and runs semantic or keyword searches on them.
Note that while some connectors are available as MCP servers, I don’t actually use any. All of my connectors are CLI utilities and/or agent skills. That’s a matter of personal preference, not an ideology – in many cases MCP servers are a great solution too, and they are easy to install and configure in OpenCode.
…
You don’t need to be a hardcore techie to set up your own personalised agentic AI environment. If you’re looking to get advice, work through some issues, and see how others work with agents, join an upcoming free Sunday School session – drop in, vibe on.
…
Skills
The primary extension and specialisation mechanism for agentic coworkers is Agent Skills. I have amassed quite the collection, both downloaded and home-cooked. Here are a few I find especially useful for coworking:
playground and artifacts-builder let me spin up ad-hoc web interfaces for exploration, learning, review, and planning. I’m creating them at a rate of a few dozen a day.
remotion-best-practices helps me create nice-looking videos agentically. I am completely useless when it comes to visual work, so being able to rely on an agent for this kind of work is significant.
nano-banana-2 and gpt-image-1-5 for creating and modifying images.
slidev helps me create presentation slides programmatically. See my demo of using this skill alongside nano banana to create a presentation.
tavily performs web searches. I prefer this to any of the built-in or extension web searches for agents I have found so far.
anki-connect and flashcard-generator are how I manage and continuously extend my growing collection of flashcards for spaced repetition (I currently have over 20,000 cards in my collection and I review an average of 100 a day).
raindrop-api connects to my bookmarks collection on raindrop.io for searches, saving new bookmarks, and managing the growing collection.
upstash-redis-kv serves as my central, cross-project key:value store in Upstash Redis. For most projects a local file is the easiest durable store to manage, but in some cases I just want to put something in memory and be able to recall it later.
docx, pdf, pptx, xlsx, markdown-converter for creating, reading, and managing all these files.
youtube-transcript because so much interesting knowledge comes in the form of YouTube videos and podcasts these days.
…
To learn more about Agent Skills, the most important extension and specialisation mechanism for AI agents, join an upcoming Mastering Agent Skills workshop, or follow the free Introduction to Agent Skills series.