DeepSeek Harness Desktop for macOS and Windows

Why does every one of these apps pollute the home directory by default? You'd think AI would be able to write the code to make the app fucking behave properly and store its settings in the platform appropriate way.

I wish I caught this before my first command, disabled it but not sure what it sent off :c but this is something to enable. It's visible in settings as well, but it does not match what you set on the web portal first even after syncing it.

not surprising. their approaches are generally leaning on gathering as much user data as possible, e.g. the paid API can't opt out training. I'd always double-check before using any service from them.

Why else would they offer their own harness except to steal your data? Grok playbook plain and simple.

It's no that fancy when is put in practical engineering terms. I've spend some time with the paper.

Paper describes such system whch has ability to enable/disable capabilities without process restart. Each plugin must follow specific shape: function to activate, function to deactivate it (both working, mutating the same shared context object), describe provided and required services. This, plus some ideas, like "every provider must outlive consumers when destructred" allows plug-n-play ensuring all dependencies are satisfied before the plugin is activated, lazy activation, and plugin deactivation/destruction.

>emacs of harnesses!

Would you be interested in a self-modifying harness written in Common Lisp? If so, then please check out Autolith: https://github.com/lambda-symbolics/autolith/

In my time using it with GLM 5.2, I've observed it modify itself to disassemble binaries in search of answers - without me needing to ask it to do so. Very powerful stuff. Here's what it looked like:

  ◇ reasoning summary
    │ The 200MB binary grep with regex backtracking is too slow. Let me write the strings out first (using grep -a with simple fixed patterns is fine; regex with .{0,80} on binary is slow). Better: use SBCL to do
    │ a fast binary search, or use tr to split null-separated strings then grep fixed strings. Let me use a quick approach: grep -ao with fixed patterns only (no regex), which was fast before.
    │
    │ To extract context, I can use SBCL to load the file into memory (206MB is fine) and search for occurrences with bounded context. Let me do it in a Lisp worker.
  
  ▸ lisp.eval
    (lisp.eval :form "(with-open-file (in \"/nix/store/g7b3p8ypxmdycry9wilil8mwl084kplw-antigravity-cli-1.1.19/bin/agy\" :element-type '(unsigned-byte 8))
      (let* ((data (make-array (file-length in) :element-type '(unsigned-byte 8)))
             (_ (read-sequence data in))
             (haystack (sb-ext:octets-to-string data :external-format :latin-1)))
        (flet ((find-all (needle)
                 (let ((out '()) (start 0))
                   (loop for pos = (search needle haystack :start2 start)
                         while pos do (push pos out) (setf start (+ pos (length needle))))
    … +10 more lines
  
    │ (with-open-file (in "/nix/store/g7b3p8ypxmdycry9wilil8mwl084kplw-antigravity-cli-1.1.19/bin/agy" :element-type '(unsigned-byte 8))
    │   (let* ((data (make-array (file-length in) :element-type '(unsigned-byte 8)))
    │          (_ (read-sequence data in))
    │          (haystack (sb-ext:octets-to-string data :external-format :latin-1)))
    │     (flet ((find-all (needle)
    │              (let ((out '()) (start 0))
    │                (loop for pos = (search needle haystack :start2 start)
    │                      while pos do (push pos out) (setf start (+ pos (length needle))))
    │ … +9 more lines
  
   lisp.eval  cpu 0.205s · real 4.018s
    │ Output:
    │ settings.json positions: (89311959 90892289 90918553 90945997 90985371 91164238
    │                           91164426 91171707 91275241 91297838 91299802 91299923
    │                           91629274 96365298 96370345 96398990 96405347 96406738
    │                           96414968 96419796 96426018 96429189)
    │ --- 89311959
    │ oughDependency: NestedType: OutputType: Deprecated: AllowAlias: Annotation: SourceFile: rsa-sha2-512rsa-sha2-256hmac-sha1-96ssh-userauthkey exchangedirect-tcpipForwardAgentGatewayPortsTCPKeepAlivedefault no…
    │ --- 90892289
    │ ks=%v)conversation file not found in source directory: %sfailed to create cache directory for conv cache: %vBalanced speed and reasoning quality for most tasksPrepared switch to default agent (will %s on ex…
    │ --- 90918553
    │ on-interface type passed to Type.Implementsreflect.Value.Slice: string slice index out of boundsUse /feedback to share your experience with the team.Transition failed: ForkConversation returned empty IDpers…
    │ --- 90945997
    … +14 more lines
  
  ◇ reasoning summary
    │ The binary confirms:
    │ - modelProvider in settings.json selects the Gemini API backend with GEMINI_API_KEY.
    │ - There's a regex ^(?:gemini-api:|blade:|evergreen:|beyond-api:)(?:.*?/models/(.+)|(?:.*/)?([^/]+))$ — model references with provider prefixes.
    │ - "No title command configured in settings.json. Please add a "title" block." — settings keys.
    │
    │ Now I need to find the settings struct JSON tags for the model field. Let me search for json:"model" again with wider context, and also for the string "title" block and other settings-related keys like
    │ "theme".

A plugin system in no way impedes having a set of curated and polished features out of the box. It simply allows the addition of new features to have the same status and be part of the same class as existing features. It's an architectural and design property of the system, not a property pertaining to quality.

In your case you can stick to installing plugins written by first parties or write your own plugin for specific use cases that you come across. You don't need to have any use cases for plug ins right this moment, but perhaps at some point if you find yourself repeating the same manual process over and over again, you can write a plugin to automate a substantial portion of it.

> My problem with plugins is that I don't want to install stuff from third parties.

Then ask the model of your choice to implement the plugin for you, I created 3 this morning just for myself

We're probably living in VS-Code's world now.

how could the authors of frontierharness.org not see they used the exact same symbol for Claude Code and OhMyPi?

i've seen "Pareto frontier" literally 500 thousand times in the last two weeks or so, and maybe once or twice before that. Can someone please explain what happened recently

I wonder why there is no Github Copilot in the comparison. It also supports BYOK and is quite capable

Deepseek harness seems good but this thread feels astroturfy

Yeah, changing the font size doesn't sound like it belongs in a plug-in, long term.

IGNORE: Now, if only DeepSeek Harness were open source…

(EDIT: DeepSeek Harness is open source—I missed it).

you do can change font size in settings, but not with keyboard shortcut. for some reason, they decided to limit font size to 17 max.

Can you elaborate? What's better about it vs. other harnesses? What can it do that others can't?

I'm not sure what you mean by it's a beast.

Is this for the DS API or with the local version?

What an ad! Do they pay you??

Just wait until I add coding to https://blackbear.app... it's going to be ridiculous.

(blackbear.app was developed entirely by a custom harness architecture)

Oddly enough, using agent harnesses is what has made it possible for us to ship native macOS, Qt (for Linux), CLI, and Windows clients. The burden of having multiple clients just isn’t that bad now.

You still can't beat the speed to a functional and usable product. Most users don't mind the 1G download.

I am writing a rust native UI using GPUI but now I need a markdown viewer and editor and also want diagrams. I might end up embedding a webview (tauri style) after all.

Who would want to maintain ai-generated mess in the long run?

Not all of them, I think Codex is Rust?

FWIW, making juggler plugin based feels quite normal to me, coming from the world of audio software, where that entire industry is still based around many thousands of VST/AudioUnit plugins which run native un-sandboxed code in the host apps, with access to absolutely anything they want. And it's somehow all just kind of been OK..

Imagine the wild fire when maliciously crafted plugins start being created, as you know they already are...

This harness supports pretty much any provider out of the box (it ships Pi LLM SDK), and DeepSeek's models are hosted by a number of providers.

You’re so, so close…

so in some research:

deepseek direct is much more cheaper than going through openrouter, openrouter price arbirtages (of course) and has a top up fee and does not pass the discounts that deepseek offer which is 50%+ discount when using it on off hours from beijing time.

meaning USA usage is 12 hours off so you can take advantage off off peak pricing

of course you can ask your ai flavor to tell you this but: open router 5% deposi fee, none direct deepseek does prompt caching for token discounts, openrouter not as easily but there

US daytime usage = automatic 50% discount direct w/ deepseek.

Seems like a category error: you can use DSH with openrouter, in fact it ships built-in support using the Pi LLM SDK.

I think DSH is a pretty well-built vanilla harness with a nice web UI (I use it in a pinned browser tab). I much prefer this to dealing with TUI clipboard/scroll jank, and it works just as well over SSH if you simply forward the port. They've genuinely thought about the architecture, and made some effort towards sandboxing the LLM's shell. Yes this is a saturated area, but they've made a nice version of the thing everyone is making, without trying to lock it down to their API. Maybe give it a try?

When my family moved to a house with a dedicated playroom, my wife and I were dreaming of getting the kid’s toys out of the living room. A friend told us “Either they play where you live or you live where they play.” I’ve found that to be true. Do you find a yolo sandbox like this has a similar tradeoff? Like, you don’t have to approve permissions in yolo but you’re constantly copying stuff to and from the sandbox. In the end maybe your stuff ends up “living” in the sandbox.

I just let my harnesses control the entire machine. Not really a big deal, and it’s a lot easier to sandbox an entire computer vs. a process or a VM.

haha i have currently piclaw running in a virtualbox ... it is like watching this old game https://store.steampowered.com/app/1818340/Creatures_The_Alb...

Telemetry/training data. Cursor didn’t get their own models from Project Gutenberg.

Same reason everything wants you to "install the app": to spy on you.

I'm reluctant to run harnesses from US companies, and prefer to use them in a VM. No damn way I'm installing one from China.

For coding I like to use Zed right now. Pretty solid integration of an AI harness and a good conventional code editor. Might give Pi another look.

> As models are getting better, harnessing will shrink

That can be true only for locally hosted models. The more supplied tools can do, the less data has to be exchanged with OpenAI/Anthropic servers. So bad harness means both higher lag and token usage.

On the contrary, I belive this will be the next battlefield.

My wild theory is that we already have AGI level models, but we are not yet using them correctly.

Pi is a great harness because you can customize it with the building blocks that make sense for you.

I think that harnesses for LLMs are as important as IDEs for languages.

It can't be long now until humans have to ask (vision) LLMs for help with CAPTCHAs.

It works fine with local models. I have it pointing at my strix halo running Qwen3.8 Flash Next, for example.

It's normally best to use the harness for the specific model. However, I found Cline is good with various models

Nature wants to create crab

If you're using Mistral Vibe installed on your computer, that is a harness. Vibe is one of many AI harnesses.

In general, a harness is something that lets your AI use tools on your computer and lets it work independently while you walk away. (Very simplified definition, not all harnesses are that automated.)

It's a long time since I tried Vibe, but when I did it was one of the worst. Back then it had broken support for MCP or computer use and other tools people now regard as near essential in AI. Hopefully Vibe is better now. Sadly Mistral are a long way behind.

Funny that you ask this question but are comparing it to a seriously niche product. Harnesses are mainstream and the one you’re using isn’t so much.

The whole thing around the model that handles system prompt, skills, agents, cache, etc. For example codex, claude code, pi, oh-my-pi, ...

I wouldn't say flawlessly.

It currently does not work with Bun and it is somewhat complicated to run it in a remote docker container due to the its weird security model.

You can run it in server mode according to the docs.

https://github.com/deepseek-ai/deepseek-harness/blob/master/...

https://github.com/deepseek-ai/deepseek-harness/blob/master/...

it's a wrapper around a model that actually executes commands from text input. Claude Code is an harness : the underlying model is Claude (with the version of your choice), Claude produce text output like `grep -in "error" server.log` , Claude Code actually execute the code in your shell and return the output to the model

Yes! A real marketplace of ideas with automatic matching, like you can see who has the same ideas as you

Reminds me of some magic device from childrens book, don't remember its name, was like a glass ball where you could see what other people are up to

Teacher of mine spent almost 20 years getting access to his FBI file through FOIA requests. Apparently it was surprising how much of his life they didn't know about.

Children with no life have never had anything worth keeping private. Even the question of having nothing to hide is foreign if you've never done anything at all.

It's like that parable about God giving us free will. If or when the intelligence agencies manage to create a generation with nothing to hide, what will their purpose be? You might as well take field notes on birds or sheep.

> I wish there was a tool I could use to share my documents, activity, etc. directly with the world's major intelligence agencies.

oh, good old times of PRISM, rightly bygone. Of course they don't do this now.

I don't get it

Facebook? (But they get the money ...)

Your documents are worth endless breadsticks as long as you keep making new ones, puny human

The chinese will go and suck in all of your laptop and everything else from your network when you install this ;)

Thanks, was wondering why this is in the feed.

You don't have to use it. I don't use it. I still think it's really cool, and I've been thinking of trying to take ideas from it and integrate it into my own workflow. In particular, the session visualization tooling.

You can use it to write a plug-in to change the name to YAH Harness!

I, too, enjoy CocaCola™ for its natural flavors and smooth mouthfeel...

> I'm DeepSeek Harness, an open-source harness from DeepSeek built on Cordis's “everything is a plugin” architecture.

It's open source. Unlike Claude Code your beloved friend which is closed-source and uses steganography along with an astonishing array of telemetry and fingerprinting techniques to conduct intelligence analysis on its users.

As a non-American, I see handing my data to an American company as potentially just as risky as handing it to a Chinese company.

It's open source. Unlike Claude Code

Well under the current US administration as non-US citizens we kinda feel this sentiment about the US now fwiw.

Ahh! Not China! Real Americans™ let their data go to American™ companies!

What if I use it in an isolated environment and only work on open source code with it?

> Your data goes to china.

What are they going to do? They have no jurisdiction over me.

I strongly prefer China having access to my personal data than the US.

better this than three-letter agencies from some nazi warmongering alternative

The new danger is US administration and the current fascit regime. China has yet not topple a single government abroad, has not bombed a single country and has not tapped the phones of its allies.

Now don't down vote before understanding the definition of a fasict:

"A fascist is a person who advocates for or adheres to fascism, a far-right, authoritarian, and ultranationalist political ideology."

China is none of that.

Is there a safer alternative? It’s not like I can trust OpenAI since they’ve proven that they’ll steal research from mathematicians. X is run by a sociopath. Anthropic is going to IPO so who can say how long they’ll be trustworthy…