AI Config

Define your AI coding configuration once and keep Claude Code, Codex, GitHub Copilot, and OpenCode synchronized.

Every AI assistant reads its own files, from its own directory, in its own format. Keeping the same instructions, agents and commands in step across four of them by hand does not work for long, and the drift stays invisible until one assistant starts behaving differently from the others.

AI Config gives you a single directory to edit — .ai/ — and generates the provider files from it: .claude/, .codex/, .github/, .opencode/ and AGENTS.md. You write your intent once, in one format, and AI Config takes care of expressing it in each assistant's own conventions.

Everything runs locally in your editor. No LLM, no network access, no account, no telemetry.

Watch the introductory video on YouTube.

Supported assistants

• Claude Code
• OpenAI Codex
• GitHub Copilot
• OpenCode

Getting started

1. Install AI Config.
2. Open the repository you want to configure.
3. Run AI Config: Initialize Project from the Command Palette.
4. Select the assistants this repository should stay in sync with.
5. Add instructions, agents, skills or commands from the Add menu in the AI Config sidebar.
6. Add provider-specific settings where you need them.
7. Run AI Config: Synchronize.

The guided flows only ask for short structural details — a name, a path scope, which settings to include. They then scaffold valid files and open them in the normal VS Code editor, where you write the description, prompt or instructions. Long-form content is never requested through an input box, and you never have to learn a YAML schema before starting.

One source of truth

.ai/
├── agents/
├── commands/
├── instructions/
├── skills/
└── config.yaml

Agents — reusable specialists you can delegate to, such as a reviewer, a backend coder or a test writer.

Commands — explicit prompts you invoke yourself when you want them.

Instructions — rules the assistant should always follow, either repository-wide or scoped to specific paths.

Skills — reusable capabilities and workflows, with optional reference material, scripts and assets alongside them.

config.yaml records which assistants are enabled. Everything under .ai/ is yours to edit; AI Config only ever reads it and writes provider output.

Provider-specific overrides

Shared content lives in one file and reaches every enabled assistant. When one assistant supports a setting the others do not, you add it separately instead of duplicating the artifact.

.ai/agents/coder.md holds the agent's description and instructions, and is compiled for every enabled provider. Alongside it:

# .ai/providers/opencode/agents/coder.yaml
schema: 1
options:
  model: anthropic/claude-sonnet-4-20250514
  temperature: 0.1
  permission:
    edit: deny

The override adds OpenCode-only options. It never repeats the shared content and never changes which assistants the agent reaches — remove it and the agent is still generated everywhere.

AI Config only offers overrides that genuinely exist for that combination of provider and artifact, so you are never walked through a question that produces nothing. Claude Code and Codex agents accept model and reasoning settings, Copilot instructions accept path scoping, and so on; each scaffolded file lists the supported settings with a short explanation and a link to the provider's own documentation.

Guided creation

From the AI Config sidebar:

• Add Agent
• Add Command
• Add Instruction
• Add Skill
• Add Provider Override…, on any artifact that can still be customized

Every artifact row carries Edit and Delete…. Deleting removes the canonical source and every provider override written for it, then synchronizes, which removes the files generated from it. Deleting the file under .ai/ by hand works too: the generated files are removed on the next synchronization. Provider overrides written for it are kept, and AI Config notes that each one produces no output until an artifact of that name returns — the same way a disabled provider's overrides are kept. That is what makes switching to a branch without the artifact, and back, lossless. Delete them yourself when you want them gone, or use Delete… on the artifact, which takes both.

Existing overrides appear underneath the artifact they belong to, with actions to edit or remove them.

Synchronization

.ai/
  ↓
AI Config
  ↓
Claude Code · Codex · GitHub Copilot · OpenCode

AI Config: Synchronize compiles .ai/ and writes the provider files. AI Config also watches .ai/ and re-synchronizes automatically once a change is valid, so in day-to-day use the generated files simply stay current. Nothing is written while the configuration still has errors.

The sidebar lists each assistant with its current status — up to date, changes pending, drifted, conflicting or disabled — and expands to show the files that need attention. Each row carries the one action that applies to it. A disabled assistant offers +, which adds it to .ai/config.yaml and synchronizes, so its files exist by the time the row reports it as enabled. An enabled one offers Remove Provider…, which disables it, deletes the files generated for it — keeping any file another enabled assistant also produces — and sends its .ai/providers/<provider>/ directory to the system trash. The confirmation says which of those apply before anything is removed. Nothing else under .ai/ is touched, so enabling it again regenerates everything except the provider-specific settings you had written for it. The status bar reports the same at a glance, and validation problems appear in the Problems panel on the .ai/ file that caused them. AI Config: Show Status, AI Config: Validate and AI Config: Show Output Log give you the details on demand.

Exact and lossy mappings

Assistants do not support identical features, and AI Config says so rather than pretending otherwise. Each mapping is reported as exact, lossy, unsupported or unverified.

For example, an instruction scoped to src/api/** is expressed exactly in Claude Code and GitHub Copilot, which both enforce path scopes. Codex and OpenCode receive the same instruction through AGENTS.md, where the scope is recorded in prose but not enforced — a lossy mapping, reported as a warning naming exactly what was lost.

Drift protection

AI Config records every file it generates. If one of them is edited by hand, that is drift, and synchronization stops instead of overwriting your change. You can open the generated version side by side in VS Code's diff editor and, if the edit was a mistake, restore the generated file. If you meant to keep the change, AI Config tells you it is blocking and leaves the decision to you.

Provider files that AI Config did not generate are never touched. When you initialize a repository that already contains a .claude/ setup or an AGENTS.md you wrote yourself, AI Config lists those files and explains that they stay yours.

Safety

• Files AI Config never generated are never overwritten, deleted or adopted.
• Ownership is tracked per file, so generated output is always distinguishable from your own.
• Unsafe paths and symlinks escaping the repository are rejected.
• A generated file whose source you deleted is removed — unless you edited it, in which case it is preserved and reported.
• Commit .ai/.generated.json along with .ai/. It records which files AI Config owns, so a fresh clone can tell generated files from yours.

Current scope

AI Config manages instructions, agents, skills and commands, plus the provider-specific settings those artifacts support — which, depending on the assistant, can include per-agent model, reasoning, permission, MCP server and hook settings.

It does not manage repository-wide or global assistant configuration such as plugins, global model settings or standalone MCP and hook configuration files, and it does not import or merge configuration you already have. Those files stay under your control.

Configuration

aiconfig.showStatusBar — show or hide the AI Config status bar item.

Support

Found a bug or have a suggestion? Open an issue on GitHub.

License

MIT