blair.dev

blair.dev

Frontend, mostly. Still adjusting that one pixel.

Joined 10/30/2026
2
Projects
0
Collaboration requests
0
Iterated
0
Likes
0
Saves
0
Views
0
Comments
AI Apps

PyBot

A chatbot with no model at all — 640,000 lines of plain Python rules, written by a swarm of agents.

PyBot is a chatbot with no model inside it at all. It answers everyday questions using about 640,000 lines of plain Python — if-else branches, regex patterns and lookup tables — with no weights, no training and nothing to download. You can try it in your browser or in your terminal with Python 3.8 or newer. It exists to answer one question: how far can a rule-based chatbot go if agents do the writing instead of people? Hugging Face set a swarm of over 1,000 agents to build it. The agents wrote and evaluated more than 5,000 versions, and the chatbot's win rate on the SmolTalkBench test climbed from 0% to about 26% — roughly on par with SmolLM-360M-Instruct, a small language model. Under the hood it is 528 Python files split into four parts: an understanding engine that works out what you typed, a memory store that remembers the conversation, a knowledge base of facts kept as data, and a response composer that phrases the reply. Every reply also shows which module produced it, so you can see the rules at work. The agents wrote 550,000 lines of tests alongside the code. More than 350 developers starred it within five days of launch. It is a research demo, not a daily assistant: because every answer comes from fixed rules, it can miss the question or answer wrongly, and it takes a few seconds to start while Python loads all 640,000 lines.

Developer Toolsclaude-code

Agent HUD: Quotas, Tokens and Approvals for All Your Coding Agents

One glance at the edge of your screen tells you which coding agents are working, which are waiting on you, and how much of your quota is left.

Run two or three coding agents at once and you spend half your attention just keeping track of them: which one is still working, which one is stuck waiting for your approval, and how much of your quota each one has already burned. Agent HUD puts a small strip at the edge of your screen that answers all three questions at a glance. More than 490 people have starred it on GitHub since it appeared in late October 2026. Each agent gets a coloured glow — green while its quota is under 70% used, yellow from 70%, red from 90% — and hovering shows its quotas with reset times or burn rates, its balance, the last 24 hours of tokens, and its sessions, with the ones waiting on you first. You can approve, deny or answer an agent's questions right on the HUD, and one click jumps you back to that session's thread, tab or terminal pane. It follows 27 agents, including Claude Code, Codex, Cursor, Antigravity and DeepSeek Harness, installs with one Homebrew or npx command, and keeps going on your iPhone and Apple Watch when you walk away from your desk. What sets it apart is that the agents can use it too. A built-in MCP server lets twelve of them message each other — Claude Code can ask a Codex session to review a change and get the answer back — and lets an agent rearrange your HUD to show only the numbers you care about, previewed first and kept only when you press Save. Token history stays on your own computer in a local database, and the phone link runs through your own iCloud with end-to-end encryption. Two honest notes. macOS is where it is at home today: Windows and Linux builds are still in preview, and the Linux build has not been tried on a real desktop yet. And while the core HUD is free, some of the on-the-go Remote features — conversations, decisions and alerts away from home — need a subscription. The code is Apache-2.0, the app is light on your machine at about 240 MB of memory, and new code landed on 15 November 2026.