baker

Open source · AI-agnostic · v0.1

Bake projects every AI agent understands on day one.

One command sets up the instructions, memory, skills and MCP config that Claude Code, Codex, Gemini, Cursor, Copilot and friends all read. When you hit a usage limit, continue in another agent without losing the thread.

go install github.com/knightlybuilds-tech/baker/cmd/baker@latest
Read the docsView on GitHub
~/projects
$ baker new shop --recipe laravel
→ Running the laravel installer…
→ Baking shop (recipe: laravel)
  + AGENTS.md          Laravel 13 · PHP 8.4 · Pest
  + CLAUDE.md          @AGENTS.md
  + GEMINI.md          @AGENTS.md
  + .agents/skills/    session-close, consolidate…
  + .mcp.json …        laravel-boost, every agent
  + memory/            MEMORY.md, hot.md, notes
  + baker.toml

✓ shop is ready.
$ cd shop && claude

The chores you redo in every project

If you build with AI agents across many folders, you know these. baker makes them one command.

Every new project needs the AI setup again

One command bakes it in: AGENTS.md, bridges, skills, memory, MCP and hooks.

Every agent wants its own file

Write AGENTS.md once. baker generates CLAUDE.md, GEMINI.md and every MCP config from it.

One giant CLAUDE.md nobody trusts

Memory is small, dated, linked notes. Agents open only what they need.

Usage limits kill your flow

baker resume codex picks up the task in another agent or account.

One file. Every agent.

AGENTS.md is the open standard most coding agents already read. baker makes it your single source of truth and generates tiny bridges for the tools that need them. It never writes legacy files like.cursorrules that would hide it.

baker only edits text between its own markers, and a lock file means it never overwrites your edits.

AGENTS.md
  • Claude CodeCLAUDE.md → @AGENTS.md
  • Codexnative
  • Gemini CLIGEMINI.md → @AGENTS.md
  • Cursornative
  • Copilotnative
  • OpenCodenative
  • Aiderread: AGENTS.md
  • Windsurfnative
  • Ampnative
  • Zednative
  • Clinenative
shop: Claude hit its limit
$ baker resume codex
→ starting OpenAI Codex with the resume brief…

# Continuing work on shop
## Handoff note (hot.md)
Goal: monthly invoices per Tenant
[x] Migration
[ ] Finish InvoiceTenants job (skip free plan)
## Repo state (live)
Branch feat/billing · 2 files uncommitted
## Last exchange (transcript)
User: skip tenants on the free plan
Agent: Adding the filter in handle() now…

Continuity

Hit a usage limit? Keep going.

Switch accounts, or switch from Claude to Codex, and the new session starts with a brief built from three layers. Any one of them is enough to continue:

  1. hot.md: the agent writes Goal, Done and Next after every finished step.
  2. Checkpoints: a hook saves git state, the todo list and the last messages after every turn.
  3. Transcript recovery: baker reads the previous agent's local transcript, even from another account.

Local only. Secrets are redacted. No LLM needed. How it works →

Memory

Small notes, not a giant file.

One idea per file, the title is the claim, under 200 words, linked with [[wikilinks]]. Agents read a one-line-per-note index and open just what they need. Every note carries a date and a source, andbaker doctor flags anything stale.

Linked notes already form a graph, so open memory/ in Obsidian and see it.

Memory rules →

memory/lessons/queue-workers-must-restart-after-deploy.md
---
type: lesson
as_of: 2026-10-18
last_verified: 2026-10-18
source: session 2026-10-18
---
# Queue workers must restart after every deploy

Workers keep old code in memory, so jobs ran
stale logic after the 10/12 deploy.

Apply: deploy runs queue:restart.
See [[deploys-use-envoyer]].

Recipes for your stack

A recipe knows how to install a stack, detect its versions and tools, and what agents should know about it.

laravel

Runs composer create-project, detects the Laravel and PHP versions and Pest vs PHPUnit, registers the Laravel Boost MCP server, and adds a migration skill.

generic

Any project. Adds the full AI layer and starter Commands and Conventions sections for you to fill in.

yours

Drop a folder in ~/.baker/recipes/ to add a stack or override a built-in one with your own taste. Write a recipe →

The whole toolbox

baker new <name> -r <recipe>Create a project, run the stack installer, add the AI layer
baker initAdd the AI layer to an existing project (plan → confirm → apply)
baker syncRegenerate bridges, skills, MCP and managed blocks; your edits are protected
baker doctorCheck bridges, sizes, legacy files, hooks and memory health
baker resume <agent>Continue the current task in another agent
baker global syncPush your personal rules into every agent's user-level file

Plain markdown. No database. No lock-in.

Everything baker writes is a file you can read, grep, commit and change. Delete baker tomorrow and your project still works.

go install github.com/knightlybuilds-tech/baker/cmd/baker@latest