docs

kiln is a native macOS app for codex and opencode. it keeps sessions, files, tools and model controls together, and adds a small account layer for public profiles, measured usage and shared chats.

install

you need macOS 14 or newer, and codex or opencode installed separately.

download the latest build from releases
open kiln, pick a folder, pick a model, start a conversation
sign in to your model provider through its own CLI: codex login or opencode auth login

sessions

each conversation keeps its own model, reasoning level, permissions and fast-mode choice. the folder in the header is where the agent starts. drafts are saved per chat, and interrupted queued requests come back as drafts instead of sending themselves after a restart.

models

codex models come from the CLI kiln is running, including the reasoning levels and fast-mode tiers it advertises. opencode keeps its provider/model ids. being in the list does not mean your account has access to a model.

account

kiln accounts live here, at kiln.raya.ac. you can register and sign in from the app or a browser. keep the recovery code shown at registration — there is no email recovery. provider login is separate and still belongs to codex or opencode.

public profile

your profile is public: handle, display name, bio, location, website and join month. handles are permanent. you can upload a profile photo from the app's own kiln avatar and choose an accent colour, presets or custom.

measured usage

usage reporting is off until you enable it on a device and account. it sends measured token counts, provider and model identifiers, an opaque session id, and timestamps — never messages, memories, file paths, attachments or credentials. counts the provider didn't report stay unknown.

showing usage on your public profile is a separate switch. it renders aggregate and per-model counts only, and turning it off never deletes history. estimated spend is a list-price estimate, split into cached and non-cached, and models without published pricing are excluded.

shared readmes

you can share a conversation as a public README. only user and assistant text is published; tool calls, reasoning, attachments and file paths are stripped. sharing is explicit and revocable: disable or delete any readme from account settings. by sharing you grant kiln access to that chat data for Ash's training.

the "share chats from the start" preference only defaults the choice on; every chat still needs an explicit share.

memory

memory connects to your installed engram and mythic runtimes over local pipes. project context only includes explicitly scoped active memories and checkpoints. whole-store search is a separate action. nothing is pasted into every turn automatically.

remote

open the authenticated remote link from settings to check a run or send something from another device. keep that link private — its token grants access to kiln. file editing, git, the terminal and local cognitive services stay native-only.

build

git clone https://github.com/raya-ac/kiln.git
cd kiln
make bundle ARCH=arm64
open dist/arm64/Kiln.app