now live for every build · engine v0.5.1

Kiln

The harness that fires your prompts into working apps. We replaced the general-purpose coding agent inside every playground sandbox with an engine built for exactly one job — and everything got faster.

before / after

openclaudekiln
agent size in your sandbox21 MB33 KB
agent bootup to 60 s19 ms
agent memory~600 MB~28 MB
fresh sandbox setup~90 sseconds
processes per sandbox21
per-edit approval round-trip1 per file changenone

measured on our eval fleet, 2026-07. every golden-prompt build passed on both production models before kiln served a single real user.

what changed for you

sandboxes start in seconds

Setting up a fresh sandbox used to clone and compile a 21 MB agent (~90 seconds). Kiln ships as a single 33 KB file — new and returning projects come up almost instantly.

faster, steadier builds

Every file change used to make a network round-trip for approval before it could apply. Kiln executes tools in-process, retries flaky provider moments automatically, and a built-in watchdog means a hung model can't freeze your build.

bounded and measured

Every build runs with hard turn and token ceilings, and reports real usage — including cached tokens — so behavior is predictable and improvements are driven by data, not guesses.

your keys stay yours

Bring-your-own-key setups no longer store credentials inside the sandbox at all. Keys travel per request from our servers, so switching providers applies on your very next build.

what stayed the same

Everything you built, and how you build. Same models, same chat, same live preview, same published apps. Kiln speaks the exact same protocol the playground always used — your projects didn't notice the engine swap, they just got faster.

changelog

v0.5.1current

2026-11-13

  • long sessions run faster and cheaper — when a chat gets long, kiln now sets aside older history in one step instead of a little on every turn, so the model can reuse what it has already read

v0.5.0

2026-11-05

  • build check — before kiln calls a build done, your app is actually built. If it doesn't compile, kiln reads the errors and fixes them (up to two rounds) instead of handing you a broken app. Rolling out to a share of builds first
  • "checking the build" shows in the chat while the check runs
  • your own model key (BYOK) never enters the build sandbox — kiln's model calls go through Twigpine, which adds the key on our side

v0.4.0

2026-11-04

  • every build is traced turn by turn — each step, file, error and timing — so failed builds can be found and fixed (your prompts and code are not recorded)
  • "did this build work?" — one tap after a build tells kiln what to improve next
  • fixed: very large file writes could go missing from your chat history and from project restores

v0.3.4

2026-07-23

  • failed edits now show the closest matching file content in the error — the model corrects in one turn instead of re-reading and retrying (edit mismatches were the #1 tool-error class)
  • build telemetry survives platform deploys (interrupted builds are recorded instead of vanishing)

v0.3.3

2026-07-23

  • adaptive stall detection — providers that stream text, then go quiet for minutes while composing a large file (seen with Grok) are no longer cut off at 90 seconds; the strict timeout applies only before output starts
  • stall errors report the limit that actually fired

v0.3.2

2026-07-23

  • huge single-file writes survive the per-turn output limit — the engine repairs the cut call and the model continues in smaller pieces, instead of the build dying with a provider error
  • provider rejections fail fast with the real error instead of ~90 seconds of silent retries
  • richer failure telemetry: error messages and per-tool error categories recorded for every build

v0.3.1

2026-07-21

  • live thinking stream — watch the model reason while it plans your app
  • build ticker: current file, turn count, and elapsed time while the kiln cooks

v0.3.0

2026-07-21

  • big-spec survival: 60-turn ceiling (up from 24) with a cost-weighted token budget — 50+ turn builds now complete
  • wrap-up and finish phases so long builds land a working app instead of falling off the turn cliff
  • visible-output watchdog — a model reasoning silently for minutes can no longer wedge a build

v0.2.0

2026-07-20

  • kiln takes over every playground build
  • per-build metrics: turns, tools, tokens (including cached), stop cause — the data driving every release above

where it comes from

Kiln began as a fork of OpenClaude, our open coding agent. We transplanted the battle-tested parts — tool semantics, editing rules, retry behavior — into a purpose-built engine a fraction of a percent of the size, then hardened it against real-world provider failures before cutover. OpenClaude remains our general-purpose agent; Kiln is what it fired into shape.