Building wigolo — open-source, local-first web search & research for AI agents. SWE @ pgEdge (distributed Postgres). Go/TS. Sharing what I learn shipping OSS.
"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 https://t.co/EMZSAYLKrP, 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.
https://t.co/2QACIIebve
@opencode Can the GitHub Actions trust be pinned to the token's sub claim, like one repo plus a branch or environment? Otherwise any workflow in the repo, including one on an unreviewed branch, can mint a model token.
Your coding agent can now draw diagrams a designer won't redo.
diagram-design (MIT agent skill):
- 45 diagram types, light/dark/editorial each
- one HTML + SVG file, no build step
- redraws Mermaid, https://t.co/LNOfZKreH6, Excalidraw files
- pulls palette and fonts from your site
- npx skills add cathrynlavery/diagram-design
@thinkymachines config.json has only 4 of 24 layers on global attention (5, 11, 17, 23), the rest are 1024-token sliding window. Do 11 and 17 keep a separate KV cache for each of their 3 passes? That's roughly 16GB vs 32GB of bf16 KV at the full 1M context.
A CAD kernel you can ship in any app, MIT and pure Rust.
springrep 1.0:
- booleans, fillets, shells, sweeps
- reads and writes STEP files
- safe Rust, zero dependencies
- cargo add springrep
- their own bench: 3,487/3,487 cases, OpenCascade 81%
https://t.co/4B72xwaNvy
@jullerino What happens when 5432 is already taken on the desktop by a local Postgres? Does the forward fail, or remap to a free port and show you which one it picked?
Clean noisy voice recordings on your own machine, free.
Inworld Voice Isolator (Apache 2.0):
- 256 MB offline model, 233 MB realtime
- 48 kHz full-band, 20 ms look-ahead for live audio
- CPU, CUDA or Apple Silicon (mps)
- pip install inworld-voice-isolation
- saves the removed background too
@0x0SojalSec Fine print from the README: that 125B (Qwen3.8-Flash-Next) on a 12 GB card also wants 64 GB of system RAM, since PolyStrata keeps only the busiest experts on the GPU and pages the rest from RAM or SSD.
@csaba_kissi Saved me once: recent cronie (Fedora, Arch) copies your old crontab to ~/.cache/crontab/crontab.bak on every edit or removal, so crontab -r is recoverable there. Adding -i makes -r ask y/n first.
Your coding agent can now drive GIMP, Blender and LibreOffice.
CLI-Anything (Apache 2.0):
- 30 app CLIs, 2,464 passing tests
- --json on every command
- renders with the real app
- /cli-anything ./repo builds new ones
- Claude Code, Codex, Cursor, Pi
https://t.co/RYRkxPi43B
@GergelyOrosz If the DB is D1, he already has a safety net: Time Travel restores to any minute in the last 7 days on the free plan (30 on Workers Paid) with wrangler d1 time-travel restore. Worth knowing before the first bad migration.
Dialogs, menus and tabs that work with JavaScript off.
Alloy CSS (MIT):
- one stylesheet, ~24 KB gzipped
- dialogs, drawers, menus, tabs, toasts
- built on popover, :has() and commandfor
- one link tag, no build step
- dialogs need Chrome 135, Firefox 144, Safari 26.2
@ithilgore Detail I didn't expect: deny globs like **/*.env are expanded to concrete paths at launch, so a .env that a long-running command creates mid-run isn't denied. Any plan to watch for new matches, or is that gap left to the proxy blocking exfil?
@_philschmid Codex just went the other way: codex mcp-server has been removed, and hosting Codex now means its app server, a separate JSON-RPC protocol with approvals and streamed events instead of MCP tool calls.
A 319 MB LoRA that makes local Qwen a far better coding agent.
Qwen3.8-27B pi adapter (Apache 2.0):
- SWE-bench Pro hard 51: 29.4% to 62.7%
- Terminal-Bench 2.1: 39.5% to 55.8%
- 55-62% fewer tool calls
- their own evals
- llama.cpp --lora over Q8_0
https://t.co/Dnw28keqrb
@RoundtableSpace The neat trick in the README: the big matmuls skip the 68000's MULS and use T-MAC bit-plane lookup tables, still bit-exact. Three assembly rounds cut "capital of peru" from 130 s to 77 s on a 367K-param int8 model.
@ClaudeDevs If you have a PreToolUse hook gating git push, test it before turning this on: from Code Mode a push can arrive as cd 'app' && git push. To vet the script itself, match the REPL tool and read tool_input.code.
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
@sydneyrunkle On expiry: does the dreaming pass stamp a valid-until on things like launch dates when it writes them, or does staleness only get caught when a later run's traces contradict the memory?