Yujie Zha

Building an AI agent that remembers, learns, and grows with you. CowAgent, open source.

[Yujie Zha @zhayujie]

CowAgent Desktop is here, on macOS and Windows 💻

The open-source agent now runs as a native app on your own machine. It drives your browser, terminal and file system directly, and holds long-term memory, a personal knowledge base, and self-evolution — all running locally.

You can also reach it from your phone to put your computer to work. 🧵

[Yujie Zha @zhayujie]

Hey, I work on CowAgent. Thanks for giving it a shot, and sorry we didn't see this sooner.
fair point on the default access. Since 2.1.7 you can set per-session permissions (read-only / workspace-write / full-access) and pin a session to one workspace, so it doesn't get free rein over your machine. Proper sandboxing is next on our list, and we'd love your input on it.

[Yujie Zha @zhayujie]

Also in 2.1.7:

• 🗂 Multiple project workspaces, grouped & pinnable sessions
• 🔔 Task-done / failure notifications, click to jump back
• 🎙 Voice input on desktop
• 🤖 New models: glm-5.3, qwen3.8-max, gemini-3.7/3.6-flash

Release → https://t.co/MAnDdx5ICl

[Yujie Zha @zhayujie]

CowAgent 2.1.7 is out 🚀

Your local AI agent now works across multiple workspaces, bind each session to its own project space, so its files and outputs stay neatly separated.

Pair that with per-session permissions (read-only, workspace-write, or full-access), for more flexible and secure control over what the agent can touch. 🧵

[Yujie Zha @zhayujie]

Available on both web and desktop. 🖥

Permissions work per session: set a default for new ones, and tune any single session on the fly. When a tool call gets blocked, the prompt is clickable, adjust access in one tap.

[Yujie Zha @zhayujie]

A few more implementation details from the source that stood out:

1. tool-cordis — self-evolution: if the agent needs a capability it lacks, it can write a plugin on the fly and load it into its own running instance, extending itself at runtime

2. Code Mode: beyond the standard loop of calling tools one by one, it can write code and run it in a sandbox to batch-call tools in one go

3. heterogeneous multi-agent orchestration: one sub-agent backend can drive Claude Code / Codex via ACP/SDK, so the harness acts as a coordinator orchestrating multiple agents

[Yujie Zha @zhayujie]

Tested & read through DeepSeek Harness (dsh) source. Two things stood out:

① Everything's a plugin: Agent Loop, tools, storage, even the UI are swappable. Very extensible.

② Sessions are an append-only event log (file/SQLite); model context is reassembled from it each time, so forking & crash recovery are clean. There's even a "Trajectory" view (a plugin itself) that visualizes the log like a browser dev console request trace.

search, browser, sub-agents, skills, sandbox, all there. Runs fast, cache hit rate 98%+ in a few cases (stable context prefix)

that said, 0.1.0 is a minimum-viable target, for non-dev users, usability still has a way to go

[Yujie Zha @zhayujie]

Open source, MIT.

⭐ https://t.co/K0reeGKFYa
💻 Desktop download: https://t.co/LTGz5nE9K6

[Yujie Zha @zhayujie]

CowAgent Desktop is here, on macOS and Windows 💻

The open-source agent now runs as a native app on your own machine. It drives your browser, terminal and file system directly, and holds long-term memory, a personal knowledge base, and self-evolution — all running locally.

You can also reach it from your phone to put your computer to work. 🧵

[Yujie Zha @zhayujie]

Self-evolution: it also reviews past conversations on its own: fixing skills that misfired, finishing tasks it left undone, tidying up memory. The more you use it, the better it fits you.

The full five-layer design 👇
https://t.co/7GrS3ztlYv

[Yujie Zha @zhayujie]

https://t.co/hHKqUBtCtR

[Yujie Zha @zhayujie]

Knowledge: everything it learns gets curated into a Markdown wiki with a visual knowledge graph.

[Yujie Zha @zhayujie]

memory works in three layers: context, daily notes, and a core memory file, each distilled into the next over time. All plain files on your disk

[Yujie Zha @zhayujie]

Message it from Telegram, Slack, Discord and more, and it acts on your computer at home — pulls a file off your desktop, runs a task, sends the result back. Your machine, on call 24/7

[Yujie Zha @zhayujie]

the desktop app bundles the backend, no setup, no terminal. Download, open, and it's running on macOS or Windows. (One-line installer's still there if you prefer the CLI or need to run it on a server.)

[Yujie Zha @zhayujie]

@zeropsio exactly. a real shell is the whole precondition, once the agent has one, the CLI becomes the natural path and MCP turns optional. sounds like Zerops is a great fit for this

[Yujie Zha @zhayujie]

we didn't build an MCP server for our product. we shipped a CLI instead

MCP is great, but for a lot of products a CLI is the lower-friction way to reach an agent:

- almost every agent already has a Bash tool (Codex, Claude Code, Cursor, CowAgent, etc). the product doesn't need to stand up an MCP server, and the client side doesn't need to support an MCP client either

- the CLI wraps auth, params, pagination and errors into the command, the agent doesn't construct HTTP calls

- it doesn't need the full API reference in context. you integrate it as a skill — a short command description is enough to pick the right command, which saves tokens and cuts mistakes

not a replacement for MCP, just a cheaper path when the agent already has a shell.

open-sourced the CLI project here: https://t.co/ct6HfrXdwd

[Yujie Zha @zhayujie]

https://t.co/jRgw4HWpkx

[Yujie Zha @zhayujie]

@paulo_eduardosp good point, real gap tbh. right now we only log the authorized side (scopes granted, exit codes in --json) not what actually happened. signed receipt per call, inputs + scopes used + output hash, that's a solid idea. gonna look into it

[Yujie Zha @zhayujie]

https://t.co/jRgw4HWpkx

[Yujie Zha @zhayujie]

As agents mature, most products will end up serving two kinds of users: humans and agents

and a CLI turns out to be one of the easiest ways for an agent to use a product — it already has a shell, no MCP server to stand up, no full API docs to load into context

so in the near future a lot of products may want to ship their own CLI, to expose their core capabilities to agents

I recently open-sourced a CLI for a AI platform product, and wrote up the design decisions behind it: language, login/auth, flags, skills, distribution, security.

https://t.co/UQBJnsLlF4

[Yujie Zha @zhayujie]

https://t.co/jRgw4HWpkx

[Yujie Zha @zhayujie]

ok this one still feels a little crazy to me
the agent ran into a limit in its own code so it edited the code restarted itself and the limit was just gone next time
https://t.co/7GrS3ztlYv

[Yujie Zha @zhayujie]

https://t.co/hHKqUBtCtR

[Yujie Zha @zhayujie]

How do you make an agent actually get better the more you use it?

I added a five-layer self-evolution mechanism to CowAgent (open source). It goes from basic memory upkeep up to the agent editing its own source code and restarting itself.

👇 Full write-up here:
https://t.co/Nx1WoZhNEe

[Yujie Zha @zhayujie]

https://t.co/hHKqUBtCtR