Think "Burp Suite for developers".
Trayce stores requests as local files using the .bru file format. The UI is based on Flutter which means it offers a super-fast and modern desktop GUI with a total download size of 13MB (on Linux). I am still adding features to it so would love feedback. Currently the new features in the pipeline are: OAuth2, GRPC, and scripting. It is open source and free to use but a perpetual license must be purchased for continued use. The license model is similar to that of Sublime Text.
Thank you!
Browser agents aren’t new, but what is still hard is covering real-world flows reliably. The inspiration for Notte was to make a full-featured platform that bridges the agent reliability gap. We’ve packaged everything via a singe API for ease of use:
- Site Interactions - Observe website states, scrape data and execute actions
- Structured Output - Get data in your exact format with Pydantic models
- Stealth browser sessions - built-in CAPTCHA solving, proxies, and anti-detection
- Hybrid workflows - Combine scripting and AI agents to reduce costs and improve reliability
- Secrets vaults - Credential management to store emails, passwords, MFA tokens, SSO, etc.
- Digital personas - Digital identities with unique emails, phones for account creation workflows
With these tools, Notte allows you to automate difficult tasks like account creation, form filling, work on authenticated dashboards. Close compatibility with Playwright allows you to cut LLM costs and improve execution speed by mixing web automation primitives and include agents only for specific parts that require reasoning and adaptability.
Here’s a short YouTube demo: https://www.youtube.com/watch?v=b1CzmfpdzaQ
If any of this sounds interesting, you can run your first agent following our quickstart on GitHub https://github.com/nottelabs/notte. Or play around with our free plan through our Notte Console: https://console.notte.cc/
We’d love to hear if there’s anything else required before you’d try or trust it on your own workflows :)
I built it because I just wanted to share a password. Not spin up infra. Not register for some "secure" web app. Not trust Slack threads. Just send a secret.
Secrets are encrypted client-side with AES-256-GCM. You get a `uuid:key` token to share. Once someone reads it, it's gone. If they don't read it in 10 minutes, it expires and deleted.
Everything is verifiable. Every release is signed, SLSA-attested, SBOM-included, and logged in the Rekor transparency log. Every line of code is public.
There's also a browser-based companion: https://app.stasher.dev — works in a sandboxed popup using the same encrypted model. Share from the terminal, pick up in the browser.
No data stored unencrypted. No metadata. No logs. No surveillance.
---
GitHub (CLI): https://github.com/stasher-dev/stasher-cli GitHub (App): https://github.com/stasher-dev/stasher-app API (Cloudflare Worker): https://github.com/stasher-dev/stasher-api CI/CD (Open): https://github.com/stasher-dev/stasher-ci NPM: https://www.npmjs.com/package/stasher-cli Website: https://stasher.dev Browser App: https://app.stasher.dev (runs in sandbox from https://dev.stasher)
Built with Cloudflare Workers, KV, and Durable Objects. All code open, auditable, and signed.
Try it:
```bash npx enstash "vault code is 1234#" npx destash "uuid:base64key"
thanks for reading
I guess this is two fold, One: I’m doing another startup! Charlie is an agent for TypeScript teams focusing heavily on augmentation. :)
Two: Over the last week or so we put GPT-5 (through our Charlie Agent) head-to-head with Claude Code/Opus on 10 real TypeScript issues pulled from active OSS projects.
Our Results
GPT-5 beat Claude Code on all 10 case-by-case comparisons.
Pull requests generated by GPT-5 resolved 29% more issues than o3.
PR review quality rose 5% versus o3.
Head-to-head case study
We measured testability, description, and overall quality across 10 head-to-head PRs. Testability measures how thoroughly a code change is exercised by meaningful, behavior-focused tests. It considers whether tests are present and aligned with the diff, whether they explore edge cases and real-world scenarios, and whether they avoid vacuous, misleading, or implementation-dependent patterns common in code generated by LLMs.
Description evaluates how clearly and accurately a pull request’s title and summary convey the purpose, scope, and structure of the code change. It emphasizes technical correctness, relevance to the diff, and clarity for future readers — penalizing vague, verbose, or hallucinated explanations often produced by code-generating agents. Quality assesses the substance and craftsmanship of the code change itself — judging whether it is correct, minimal, idiomatic, and free from hallucinated constructs.
It emphasizes clarity, alignment with project norms, and logical integrity, while identifying agent-specific pitfalls like over-engineering, incoherent abstractions, or invented utilities.
Testability: Charlie (0.69) vs Claude (0.55) Description: Charlie (0.84) vs Claude (0.90) Overall Quality: Charlie (0.84) vs Claude (0.65)
Caveats
Single-shot runs; no human feedback loop. Quality score uses a secondary LLM reviewer—subjective but transparent.
Def looking for feedback on more evaluations we can do, also please do nit-pick the prompts, ideas, harness design etc etc. Tell us if this bar (CI + types) is the right one, or what you’d track instead.
On a personal note: I’ve spent my career working on tools to help creators create, I’m extremely passionate about enabling people to do more easily. I am still somewhat uneasy about Gen AI, however I do believe the future is bright, certainly things are going to change - I would encourage you all to stay optimistic builders.
Thanks for taking a look!
I’m a student with a strong interest in computer science and complexity theory. Recently, I worked on a manuscript attempting to prove that P ≠ NP.
I know how this sounds — it’s one of the hardest and most debated problems in CS, and many have tried and failed. I don’t claim to have the final answer, but I believe the approach I used might at least offer some fresh perspective or provoke useful critique.
The idea involves geometric separation between deterministic and nondeterministic computation, using high-dimensional lattice constructions and some physics-inspired intuition. The full argument is technical, but I tried to keep it logically structured.
The preprint is 93 pages long. It was originally written in Russian, and I created an English version via machine translation, so apologies in advance for awkward wording or formatting.
DOI and full paper (both languages): https://doi.org/10.5281/zenodo.16759468
I’m genuinely open to feedback — whether it’s pointing out flaws, questioning assumptions, or just explaining why this approach doesn’t work. Any form of critical input is welcome.
If there’s interest, I can also create a short video explanation in simple terms to walk through the core ideas — even though I don’t have a channel or any audience yet.
Thanks in advance for taking the time, even if it’s just to skim or tell me what to fix!
All data is stored locally. I built this is a proof of concept because I needed unlimited branching on tasks and a way to walk through the task tree step by step. I will build on this as I go, so all feedback is greatly appreciated.
What is a binary puzzle? There's a grid (n × n boar, commonly 6 × 6 or 8 × 8). Two symbols – traditionally 0/1, black/white, or, in LinkedIn’s case, moons and suns.
Rules Equal count – each row and each column contains the same number of each symbol (e.g. three 0s and three 1s in a 6 × 6 grid). No three in a row – you can’t have three identical symbols consecutively in any row or column These three simple constraints make the puzzle non‑trivial but always deterministic: a correctly generated board has a unique solution.
Some history: Early 2000s it appeared in Japanese puzzle magazines as Binairo. Later popularised in the West as Takuzu (“binary” in Japanese). LinkedIn rebranded the same mechanics as Tango, swapping 0/1 for moons and suns. The underlying logic is identical it's just the graphics that are a cosmetic layer.
How I generate an endless supply The hardest part is guaranteeing that every new board is both valid and uniquely solvable. My generator follows a standard constructive approach:
Back‑tracking placement – recursively fill cells, pruning any branch that would break a rule. Early symmetry breaking – enforce row/column uniqueness as soon as possible to cut the search space. Uniqueness verification – once a full grid is built, run a deterministic solver; if more than one solution exists, backtrack and try a different seed.
Live demo – https://taengo.vercel.app Source code – https://github.com/alexander-gekov/taengo (need to clean it up in the next days).
I have implemented the Reputation Circulation Standard (RCS), which uses exponential decay (like radioactive decay) to make harmful AI behavior economically irrational.
The key insight: instead of trying to control AI, create environments where beneficial behavior is the only stable strategy. Power accumulation becomes mathematically impossible.
Technical details: - Smart contracts for reputation circulation - Consensus algorithms for detecting divergence - <5% performance overhead - Scales with intelligence level
Nassim Taleb endorsed the approach. Already implemented a simpler version for onchain recording of reputation for Animoca Brands - Moca chain.
Now raising funds to deploy at AI labs before GPT-5: https://manifund.org/projects/preventing-ai-catastrophe-thro...
Paper: [SSRN coming] https://deepthinker.xyz has paper. Demo available under NDA.
Would love technical feedback from this community.
Any fellow BJJ folks around? o/
So, last week I registered for my first BJJ competition, and while reading the IBJJF rulebook, I realized I’m not as familiar with it as I thought. So, I decided to “vibe-code” this app to generate quizzes based on different rulesets (IBJJF, ADCC, AJP).
It’s very simple: 1. You choose the ruleset (example: IBJJF) 2. Pick the topics you want to focus on (Points Scoring, Advantages) 3. Select the number of questions 4. That’s it — it gives you a quick quiz with multiple-choice questions pulled from the official rulebook.
If you try it out, I’d love to hear some honest opinions. Just letting me know if you find it useful.
Some technical details: * All the questions are pre-generated by AI, not created on the fly. I’ll probably keep adding more and maybe focus on situational type of questions. * I completely vibe-coded this using vscode and copilot in agent mode with Claude Sonnet 4. It’s impressive how fast it could prototype and make it grow until the point I was almost lost in my own project!
[]s
Since its start in 2023, it’s now live in 200+ schools and colleges, offering cloud-based virtual lab solutions. With a vision for nationwide impact, that the planning of sivakumar kuppannan they are already conducted a hackathon on August 20 with 20+ colleges and giving awards for students. Now major hackathon by 2028 and aim to deliver virtual IT training across India.
This is a bit of an augmented concept. Adding vibration patters that "match" the audio frequencies –as much as possible– makes it a very different but pleasant exeperience.
I'd love to see what you think about it.