Amit (@amit5214) profile picture

Senior full-stack engineer · React, TypeScript, Node · Building tools for AI agents (Grenz, Treeship) · Founder of JobJam · Berlin

Berlin, Germany
Joined January 2011
Photo by Amit (@amit5214)Photo by Amit (@amit5214)Photo by Amit (@amit5214)Photo by Amit (@amit5214)Photo by Amit (@amit5214)Photo by Amit (@amit5214)Photo by Amit (@amit5214)Photo by Amit (@amit5214)Photo by Amit (@amit5214)Photo by Amit (@amit5214)
Pinned Tweet
Too many tasks, and dinner's ready. Queue them in Vakr before you leave. Each one runs as its own coding agent, in its own worktree, on your Macs overnight. In the morning: one calm review. Mac app, waitlist open: vakr.app
3
1
4
128
You queued five tasks and went to bed. Overnight: – obeyed hidden orders in a README – installed "reacct" (not react) – weakened a test you approved – hid an error – tried to send your live API key out Vakr caught all five. Happy Halloween 🎃 vakr.app
1
1
19
Coding agents say "tests pass" a lot. Vakr reads what the run actually did: every command, its exit code, every file it touched. "Tests pass" with exit 1 gets flagged, not believed.
2
1
57
Here's what that looks like 👇
19
Every agent app is a chat now. Every chat ends in "Done ✓". Run 24 agents and you're reading 24 chats. Vakr checks what each agent said against what it actually ran, and leaves you the 3 decisions that need you. vakr.app · early access
1
1
20
What my morning looks like now: ☕ 6 changes waiting ✅ tests already run on each ❌ 1 failed, sent back to the agent with the error 🔀 5 merged in 10 minutes Queued the night before. Ran on my own Macs. Nothing touched main. vakr.app
1
19
Works with Claude Code, Codex and ~20 more agents. Want to try it? Reply and I'll DM you the build."
13
Works with Claude Code, Codex, OpenCode and about 22 more agents. Your code stays on your Mac. Agents never touch main: every run gets its own branch, and you merge locally.
17
Most agent UIs stop at chat. Once an agent can run commands and edit files, you need other screens: approve one action, see every tool call, review the diff, stop it. I built an open-source React kit for those screens. A thread:
3
1
72
It isn't tied to an agent framework: components take plain props. There's a working example with the Vercel AI SDK. A ToolLoopAgent with tool approval, plus one adapter file, and the approvals and kill switch answer through the SDK.
1
7
3am: accepted into Claude Startups for JobJam 🎉 7am: offer paused while they re-review applications 😅 Hundreds of thousands of founders applied in days. That says a lot about what people are building on Claude. Fingers crossed for the re-review 🤞 @AnthropicAI @claudeai
1
61
Interesting Claude Code behavior: Sonnet 5 refused to commit/push, even after I asked again. Switched to Opus 5.5 → it pushed and created the PR. Same task. Same instructions. Different model → different behavior.
1
1
56
An agent refunded $40,000. Three weeks later: who approved it? Open the chat log. Change $40,000 to $400. Change alice to bob. Nothing objects. Nothing can. 80 seconds on what changes when the record is signed 👇
1
1
2
80
I wired WebMCP into JobJam and let ChatGPT drive it. The agent never got a token. Not one.
1
1
1
70
Tools run in the page and call same-origin fetch. The browser attaches the HttpOnly session cookie. Postgres RLS scopes every row to whoever is logged in. No key to leak, no OAuth dance, no scope list to get wrong. The agent cannot exceed the human sitting in front of it.
2
1
50
Then it went to spend one of my credits, and the page stopped it to ask. An agent that spends your money and leaves the screen unchanged is a bug. So the tool navigates to the result it just paid for. Built for the WebMCP Challenge: devpost.com/software/jobjam-…
19