Towhid Khan (@yourtowhid) header
Towhid Khan (@yourtowhid) profile picture

Building wigolo — open-source, local-first web search & research for AI agents. SWE @ pgEdge (distributed Postgres). Go/TS. Sharing what I learn shipping OSS.

Cloud and AI world
Joined April 2024
Photo by Towhid Khan (@yourtowhid)Photo by Towhid Khan (@yourtowhid)Photo by Towhid Khan (@yourtowhid)Photo by Towhid Khan (@yourtowhid)Photo by Towhid Khan (@yourtowhid)Photo by Towhid Khan (@yourtowhid)Photo by Towhid Khan (@yourtowhid)Photo by Towhid Khan (@yourtowhid)Photo by Towhid Khan (@yourtowhid)Photo by Towhid Khan (@yourtowhid)
"Local-first" tools usually mean an afternoon of setup. So I recorded the whole thing with no cuts hiding anything. Frame 1: rm -rf ~/.wigolo — genuinely fresh machine, browser engines and models wiped. One command: npx wigolo init --agents=claude-code. It downloads the browser engine and on-device ML models, runs health checks, wires the MCP config into your agent, and prints a per-component summary. 90 seconds of downloads, once — you can watch the real progress bars. Then the payoff: I ask Claude Code what's new in the latest TypeScript release — something it can't know from training. Live answer in 25 seconds, cited to devblogs.microsoft.com, through wigolo's own search stack. No API keys. No accounts. No bill that grows with how much your agent thinks. Works with Claude Code, Codex, n8n — or any agent that speaks MCP, REST, or SDK. github.com/KnockOutEZ/wigolo
2
4
1
15
3,821
Put your coding agents behind your own approval gates. Tencent LoopForge (MIT, self-hosted): - drives Claude Code, Codex + 8 more CLIs - agent, script and human-gate steps - send work back to any step - Docker sandbox per run, opens the PR - ./start.sh, demo needs no API key
1
2
Make Python require a type on every variable, not just functions. gawain (MIT): - flags untyped vars, params, returns, attrs - then runs ty or pyrefly on them - can ban Any - 0.46 s on CPython's stdlib, their own bench - uv tool install "gawain[ty]" github.com/matheuzgomes/gawa…
8
A voice model that listens and talks at once, on your own GPU. Haru-4B (Apache 2.0): - 5.27B params, 10.5 GB bf16 - 133 ms frames, hears and speaks in parallel - ~12 GB NVIDIA GPU or MLX on a Mac - pip install -e ./haru, then localhost:8998 - 9 languages, voice from a 3 s clip
1
4
Test if an agent skill helps before you ship it. NVIDIA SkillEvaluator (Apache-2.0): - pip install skillevaluator - static score: no API key or GPU - live runs with vs without the skill - flags overlapping guidance - GitHub/GitLab CI templates github.com/NVIDIA/SkillEvalu…
12
Embed a spreadsheet with formulas in your app, or run it headless. Univer (Apache-2.0, 25.5K stars): - formulas, sort, filter, validation - canvas render, React or Vue - same engine on Node 18.17+ - pnpm preset, 1 container div - xlsx import/export is paid Pro
1
10
Detection, OCR and masks on a phone, from one 22 MB model. Mirage (Cactus, Apache 2.0): - 51.3M params, 21.6 MB at 2-bit - boxes, text, masks, relations as JSON - 74 ms on a Galaxy S26 NPU, their own bench - pip install -U cactus-compute huggingface.co/Cactus-Comput…
1
13
Alibaba's in-house AI code reviewer is now open source. Open Code Review (Apache-2.0): - ~1/9 the tokens of Claude Code, same model (their bench, 200 PRs) - ocr review for diffs, ocr scan for repos - plugins for Claude Code, Codex, Cursor - delegate mode: no extra API key
2
84
Peer-to-peer WireGuard tunnels, no account, no root. Tailcat (Tailscale, BSD-3): - netcat-style CLI + Go library - a ~140-byte address is the whole config - NAT hole-punching, relay fallback - forward ports, SSH, SFTP drop box - free relays or your own github.com/tailscale/tailcat
1
42
Voice-cloning TTS now runs in your browser, with no server. ZONOS3-Mini (Zyphra, Apache 2.0): - 800M params, about 100M active (MoE) - 663 MB GGUF - WebGPU in browser, MLX on Mac, ggml on CPU - 44.1 kHz, up to 10 min per generation - 27 stock voices, or clone from a clip
1
184
Coding agents now get web advice from browser teams. Modern Web Guidance (Chrome, Edge, Firefox, MDN): - 250 features: 140 CSS, 36 HTML, 74 JS - Claude Code 57% to 88%, their own bench - Codex 54% to 82% - npx modern-web-guidance install - Apache 2.0 github.com/GoogleChrome/mode…
1
3
83
You can now run a Brother Skitch without its subscription app. ZigZag (Apache 2.0): - SVG or PNG logo in, 11 stitch types - exports PES, DST, EXP, JEF, VP3 - sends to the Skitch PP1 over Bluetooth - MCP server so Claude Code or Codex can draw it - signed Mac DMG, Windows zip
1
1
115
One photo in, a 3D object out, on a 10 GB GPU. Idol (speridlabs): - 3.1 GB weights, 7.2 GB VRAM peak - up to 1M Gaussians, standard .ply - RTX 4090 19s, RTX 3080 46s (their own bench) - python -m idol.sample --image photo.jpg - Apache 2.0, NVIDIA only huggingface.co/speridlabs/id…
1
194
Claude Code can now cut your product demo video locally. skill-video-studio: - Playwright CLI captures real clicks in your app - HyperFrames adds title cards, chapters, CTA - FFmpeg masters it, plus thumbnail - 6 local music styles - MIT, v1.0.1
2
1
164
Every Linux syscall number for 27 architectures, no kernel source needed. syscall.sh: - Linux 3.3 to 7.3 - x64 369 syscalls, arm64 326, riscv64 327 - args and registers per arch - JSON with since/removed_in per entry - free JSON API with per-release diffs syscall.sh
62
A tool-calling agent model that fits in about 1GB on your phone. LiquidAI LFM3-1.7B: - Q4_K_M GGUF 1.10GB, QAD Q4_0 1.05GB - 256K context - post-trained for Pi, OpenClaw, Hermes Agent - BFCL V4 71.4 vs Qwen3.5-9B 62.8 (their own bench) - llama.cpp b11960+, Ollama, MLX
1
2
258
Run a full game studio of Claude Code subagents from one repo. Claude Code Game Studios: - 49 agents, 74 skills, 15 hooks - Godot 4, Unity and UE5 sets - clone, run claude, type /start - their own test: the 30-doc build ranked last - MIT github.com/Donchitos/Claude-…
87
Test APIs and inspect webhooks in your terminal, no Postman account. Ghostman: - pip install ghostman, Python 3.10+ - ~30MB RAM, ~150ms cold boot (their own numbers) - webhook listener on :9090 with Ctrl+W - collections saved as plain JSON/YAML, commit them to Git - MIT
1
1
111
Who-said-what transcripts of an hour of audio, offline on one RTX 5090. NVIDIA Nemotron 3.5 Transcribe: - 34B MoE, 4.3B active, NVFP4 ~21.6 GB - 40 languages, up to 1h per file - speaker tags + word timestamps - vLLM, or GGML on native Windows huggingface.co/nvidia/NVIDIA…
1
1
67
Your coding agents can now hand work to each other and hear back. Herdr, free, Apache 2.0: - hc run --bg codex "review my diff" - helper's report lands in your agent's input - sidebar: working / done / blocked - quit the app, agents keep running - SSH boxes in the same sidebar
4
1
2
221
Find every panic path in your Rust binary, std included. Call Graph Analyzer: - reads monomorphized MIR incl. deps - cargo cga check: backtrace per panic - can ban std::fs / std::process calls - ~2-3x cargo check time - MIT, Tweede golf + Microsoft github.com/OpenDevicePartner…
1
60