I made llx-desk: a 2.6 MB browser window Claude Code can drive: test your app, find flights, fill forms
Comment on any element → the agent gets it with a screenshot and fixes it
No bundled Chromium · Rust · MIT · macOS + Linux
https://t.co/7BRhGCCtND
I made llx-desk: a 2.6 MB browser window Claude Code can drive: test your app, find flights, fill forms
Comment on any element → the agent gets it with a screenshot and fixes it
No bundled Chromium · Rust · MIT · macOS + Linux
https://t.co/7BRhGCCtND
I built a Claude Code plugin that shows where your tokens actually go.
First run on my own sessions: 68% of the spend came from messages sent when the context was already over 400K tokens.
Not cache misses (1.5%). Compaction kicking in too late.
Second finding: 44% went to subagents, 37% to subagents on Opus. They inherit the main model unless you say otherwise.
The plugin suggested two settings:
"autoCompactWindow": 400000
CLAUDE_CODE_SUBAGENT_MODEL=sonnet
Day one after: spend above 400K context fell from 72–80% to 16%. Weekly numbers soon.
tokendoctor:
• finds where tokens went, from local transcripts
• proposes fixes, applies only what you approve
• tracks by day whether it helped
• status line warns before limits run out: week 35% → 100% in 2d17h
Local, MIT.
https://t.co/eLEeB9PNlA
I built a Claude Code plugin that shows where your tokens actually go.
First run on my own sessions: 68% of the spend came from messages sent when the context was already over 400K tokens.
Not cache misses (1.5%). Compaction kicking in too late.
Second finding: 44% went to subagents, 37% to subagents on Opus. They inherit the main model unless you say otherwise.
The plugin suggested two settings:
"autoCompactWindow": 400000
CLAUDE_CODE_SUBAGENT_MODEL=sonnet
Day one after: spend above 400K context fell from 72–80% to 16%. Weekly numbers soon.
tokendoctor:
• finds where tokens went, from local transcripts
• proposes fixes, applies only what you approve
• tracks by day whether it helped
• status line warns before limits run out: week 35% → 100% in 2d17h
Local, MIT.
https://t.co/eLEeB9PNlA
Decision models (Jev from TypeSafe) are a relatively new class of models: they don't write text, they make decisions. I built them into my pet project in a few places. Here's what they are and how it went.
A regular LLM picks an action by calling a tool: it reasons first, then writes the name and arguments token by token. Every step takes seconds. A decision model doesn't generate text at all. You give it a state (a page, a conversation, a game board) and typed questions:
- yes/no → a probability
- pick from options → a probability for each option
- rate on a scale
The answer comes back in a single pass, and you pay only for input tokens. For Jev that's $0.042 per million, with output free. A decision takes about 0.3 seconds.
Where I used it:
1. Browser automation. An LLM sets the goal, and Jev picks the next action and the page element at each step. The big model doesn't need to be called on every click; it only has to verify the result.
2. Game agents. The model picks a move from the legal ones for the current board state. No text generation.
3. Agent orchestration. In a group chat, the model decides which agent replies next, even when the message doesn't mention anyone. On a labeled set of dialogues, this raised the share of correct picks from ~74% (@mentions only) to ~92%.
Where I see more potential:
→ Routing between models: which LLM, and at what reasoning level, to use for a task. TypeSafe already does this with Jev Router.
→ Trading. A decision model doesn't predict price; it answers narrow questions. Does this news affect an open position? Does the trade fit the risk profile? Is the signal even worth processing? Profit is a coin flip, but it's fun to watch.
→ DevOps: classify incidents and pick a response playbook.
→ CI/CD: decide which tests to run after a change.
→ UI that adapts to the user on the fly.
And much more
Localix is a desktop development environment built around agents. The editor, terminals, browser, databases, plugins, and coding agents like Claude Code and Codex are parts of one app, so anything you open can become context for an agent, and anything an agent does shows up where you work
https://t.co/9HBZTB7Ds5
The provider streams to my local server, not the browser. the tool marks move as sealed, and the server swaps it for [sealed] before anything hits the db or the UI. the plugin keeps a salted sha256, you only see the hash, and your browser checks it on reveal. the one leak left is the model's own thinking))
Localix is a desktop development environment built around agents. The editor, terminals, browser, databases, plugins, and coding agents like Claude Code and Codex are parts of one app, so anything you open can become context for an agent, and anything an agent does shows up where you work
https://t.co/9HBZTB7Ds5
After months of building, Localix is out 🚀
An open-source agentic development environment: a team of AI agents sharing your editor, terminals, browser, databases, and more—all in one workspace
https://t.co/LH1kAnsegT