Verkstead

Build intentional software with AI agents, on your terms

Features

Commitrate-limiting

Refuse requests over the limit with 429 and Retry-After

tidewatchb7e41c23 files+58−4

Message

Requests past a key's limit are refused with 429 rather than queued. The limiter sets Retry-After from the window's remaining seconds, so a well-behaved client backs off without guessing.

Diff

src/limit.rs+31 −2

@@ -41,6 +41,8 @@ pub fn check(&self, key: &ApiKey) -> Verdict

41 let window = self.window(key); 42− if window.count > key.limit { return Verdict::Queue; } 42+ if window.count >= key.limit { 43+ return Verdict::Refuse { retry_after: window.remaining() }; 44 } 45 window.count += 1; 46 Verdict::Allow

Experienced developers know AI output needs a human eye, and that reading every line of a fast-moving codebase is impossible. Verkstead presents each change as a written summary and a diagram above the diff, so you can judge the decisions that matter at a glance. Then a fresh-eyed review session reads the whole pull request and puts its findings to you as questions, each with a fix you approve before it is made.

Waiting on an agent drains your momentum, and juggling terminal tabs only splits your attention. Verkstead runs each piece of work as its own conversation and asks for you only when a decision needs you, with a push to wherever you are. When a run stops you get a notice, and can step in to resolve it whenever you have time. Everything else keeps going.

limit.rsTerminal

$ pwd
/work/tidewatch
$ ls ~/Documents
ls: cannot access '/home/you/Documents': No such file or directory
$ cargo test
running 14 tests ... ok
$

Every agent session runs inside a sandbox with its worktree, its git directory, and nothing else of your machine. Bubblewrap on Linux, Apple's sandbox on macOS, a dedicated local account on Windows. Where no boundary is available, Verkstead refuses to run the session rather than running it bare. Let your agent run unattended, knowing it can never reach the wrong files.

Process

New conversation

Add rate limiting to the public API. Last night one client sent 40k requests in a minute and took the service down. Limit per API key, with a sensible default that can be changed per key.

Repotidewatch ProcessDevelop Agent

Opus 5

+ Save as draft Start work

Verkstead starts with a clean text box. Write what you want in your own words, as rough as you like. There is no prompt engineering to do. The development process is built in through Verkstead's agent skills, so you are free to focus entirely on your problem.

Rate limiting for the public API

Preface

POST /v1/keys has no limit today. A limiter can land this week; where the counter lives is the one call that is yours.

Questions

Q1Where should the counter live?

1In-process counter★ 2Shared Redis counter Q1 — Your thoughts

Q2Default limit per key?

1600 requests a minute★ 260 requests a minute

Grilling is the single biggest change that separates intentional development from vibe coding. Your agent reads the code you have, then asks you about every technical detail that matters, with a recommended answer for each, and goes on only once everything is settled. Related questions arrive together as one question set you consider and answer as a whole, from your phone if you like.

RunningImplementing

Task listrate-limiting2 of 5 done

☑Count requests per key in a sliding window01
☑Refuse requests over the limit with 429 and Retry-After02
☐Make the limit configurable per key03
☐An admin endpoint to read a key's usage04
☐Document the limit05

Committidewatcha19f3d7

Count requests per key in a sliding window

4 files+121−2

Committidewatchb7e41c2

Refuse requests over the limit with 429 and Retry-After

3 files+58−4

Agent run14 turns

Your settled design becomes a handoff, a task list, or a staged roadmap, depending on its size, and each step runs in a fresh session with just the context it needs. Every commit arrives with a written summary and a diagram of what changed, so you can follow along without scanning every line.

Waiting on checksWrapping

Pull request#128tidewatch✓

Rate limiting for the public API

Question setReview of rate-limiting

Q1Retry-After rounds down to 0 on the window's last second—
Q2The admin endpoint is unauthenticated—
Q3Limit config is read on every request—

A job is rarely done just because every task is ticked. Verkstead opens the pull request, waits on CI, fixes what fails, resolves merge conflicts, and runs a fresh session to review the whole thing. Its findings come to you as one question set, so you choose what gets fixed and how.

The record of every question and decision is yours to share with colleagues, as a single file or a link. And when there is more to do, steer the conversation into a follow-up and extend the same pull request.

Not everything is a feature.

TinkerMake a series of simple one-off changes with minimal fuss
InvestigateAsk questions about a code base without committing any code
ReviewReview an existing PR or branch, and fix any issues found
Fix merge issuesFix up a PR which is unable to merge

Platforms

Code

tidewatch ▾src/ main.rs limit.rsM keys.rs ▸tests/

limit.rsTerminal

1// One key's requests in the current window. 2pub struct Window { 3 count: u32, 4 opened: Instant, 5} 6 7impl Limiter { 8 pub fn allow(&mut self, key: &str) -> Decision { 9 let window = self.window(key); 10 if window.count >= self.limit(key) { 11 log::debug!("refused {key}"); 12 return Decision::Refuse; 13 } 14 window.count += 1; 15 Decision::Allow 16 } 17}

Verkstead Desktop runs with full functionality on all three. The workbench includes a code editor and a terminal that run inside the session's sandbox, so you can try the agent's work and finish it by hand without leaving the window.

← Conversations

rate-limitingtidewatch

Waiting on youGrilling

Question set

Rate limiting for the public API

Q1Where should the counter live?—
Q2Default limit per key?—
Q3What does a refused client get back?—

Brief

Add rate limiting to the public API. Last night one client sent 40k requests in a minute and took the service down. Limit per API key, with a sensible default that can be changed per key.

The workbench is fully responsive, and Verkstead's client-server design means the work carries on at home while you answer from the train. Switch on remote access over Tailscale, scan the code, and the next question finds you wherever you are.

Install Verkstead on more than one machine and they form a cluster. Open any one of them and drive them all. Move a conversation to the Mac to build the Mac side, branch and working changes included, and bring it back when it is done. Cross-platform development, from one place.

New conversation

Add rate limiting to the public API. Last night one client sent 40k requests in a minute and took the service down. Limit per API key, with a sensible default that can be changed per key.

Repotidewatch ProcessDevelop Agent

Opus 5

Verkstead is MIT-licensed and works with Claude Code, Codex, Grok Build, and the open-source OpenCode, so you get the latest frontier models without being tied to one provider. Yes, it works with your Claude subscription. Your process lives in Verkstead, not in any one vendor's tool.

Downloads

Platform	File	Note
Linux	`Verkstead-x86_64.AppImage`	Needs bubblewrap and FUSE on the host
macOS	`Verkstead-universal.dmg`	Signed and notarized: drag into Applications
Windows	`Verkstead-x86_64.exe`	Unsigned: SmartScreen, More info, then Run anyway
Command line and NixOS		The releases page, and the README