One source.
Every AI coding tool.

Write your project instructions and agents once, in your repository. agnostic-ai sync writes the file each tool reads, in its own format.

Works with 32 coding tools

Install agnostic-aiAll install options

curl -fsSL https://raw.githubusercontent.com/Chemaclass/agnostic-ai/main/scripts/install.sh | bash

Set up with an agent Try it in the browser

With agnostic-aiWithout

.agnostic-ai/agents/reviewer.md

model: fast

agnostic-ai syncEdit each file by hand

.claude/agents/reviewer.mdClaude Codemodel: haikumodel: haiku
.codex/agents/reviewer.tomlCodexmodel = "gpt-6-luna"model = "gpt-5.4-mini"
.gemini/agents/reviewer.mdGemini CLImodel: gemini-3.5-flashNo agent yet.
+ 29 more tools

1 spec
fast is a tier in agnostic-ai.yaml

3 files by hand
Models go stale

Your setup outlives the tool you picked.

Every AI coding tool ships its own config file. Adopt three and your instructions live in three places, drifting apart. agnostic-ai keeps one source in your repository and writes the rest.

You own the source

Markdown and YAML in your repo. No account, no service.

Generated files are outputs

Same input, same bytes. Catch drift in CI with sync --check.

Keep native tool files

Each tool reads its own files. Sync flags what a tool cannot use.

Stop syncing at any time

The generated files keep working.

Why not just symlink one file?

Each tool reads a different format at a different path. A symlink shares bytes. It cannot translate them.

See how the formats diverge

Start with one note. Then change it once.

Choose your tools

agnostic-ai init

Run this in your project. Pick Claude Code and Codex, or the tools you use. Accept the import offer to keep existing instructions.

Write one project note

.agnostic-ai/AGNOSTIC_AI.md

Add a short instruction, such as the test command for your project. This is the source you maintain.

Generate and read the files

agnostic-ai sync

Preview with agnostic-ai sync --plan, then sync. With Claude Code and Codex, your note is in AGENTS.md, and CLAUDE.md points at it.

Make your next edit

agnostic-ai sync

Change the note and sync again. Every tool you picked reads the new text.

All install options Follow the setup guide Bring an existing setup

Share more when you need it. Explore the source files.

Everything lives in .agnostic-ai/, next to agnostic-ai.yaml, which picks your tools and model tiers. Pick an entry to see its source and what sync writes for each tool.

.agnostic-ai/

.agnostic-ai/ also holds reviews/, environments/, and ignore/. See every spec kind.

Project instructions, written into each tool's root file. Select rules/ to see where rules land.

.agnostic-ai/AGNOSTIC_AI.md

# Billing API

- Run `pnpm test` before every commit.
- Never edit `prisma/migrations/` by hand.

sync

CLAUDE.md (excerpt)

# Billing API

- Run `pnpm test` before every commit.
- Never edit `prisma/migrations/` by hand.

AGENTS.md (excerpt)

# Billing API

- Run `pnpm test` before every commit.
- Never edit `prisma/migrations/` by hand.

GEMINI.md (excerpt)

# Billing API

- Run `pnpm test` before every commit.
- Never edit `prisma/migrations/` by hand.

Docs: Entry-point files

Subagents with a role, what they may do, and a model. model: fast names a tier from agnostic-ai.yaml, and sync writes each tool's model and effort.

.agnostic-ai/agents/test-writer.md

---
name: test-writer
description: Writes a failing test that reproduces a bug. Use before fixing it.
can: [read, edit]
model: fast
---

Reproduce the bug in a failing test. Do not touch production code.

sync

.claude/agents/test-writer.md

---
name: test-writer
description: Writes a failing test that reproduces a bug. Use before fixing it.
tools:
  - Read
  - Edit
model: haiku
effort: low
---

Reproduce the bug in a failing test. Do not touch production code.

.codex/agents/test-writer.toml

name = "test-writer"
description = "Writes a failing test that reproduces a bug. Use before fixing it."
model = "gpt-6-luna"
model_reasoning_effort = "low"
developer_instructions = """
Reproduce the bug in a failing test. Do not touch production code.
"""

.gemini/agents/test-writer.md

---
name: test-writer
description: Writes a failing test that reproduces a bug. Use before fixing it.
model: gemini-3.5-flash
tools:
  - read_file
  - replace
---

Reproduce the bug in a failing test. Do not touch production code.

Saved prompts you run by name, such as /review. Codex receives a skill you invoke yourself.

.agnostic-ai/commands/review.md

---
name: review
description: Review a change against main.
---

Review the current branch against main and report bugs with file and line references.

sync

.claude/commands/review.md

---
name: review
description: Review a change against main.
---

Review the current branch against main and report bugs with file and line references.

.agents/skills/review/SKILL.md

---
name: review
description: Review a change against main.
---

This skill is a command. Run it only when the user calls it by name, and treat the text they type after the name as its arguments.

Review the current branch against main and report bugs with file and line references.

.gemini/commands/review.toml

description = "Review a change against main."
prompt = """
Review the current branch against main and report bugs with file and line references.
"""

Commands that run on tool events. Write on: after-edit once, and sync writes the event and edit tool names each tool uses.

.agnostic-ai/hooks/format.yaml

name: format
on: after-edit
command: .agnostic-ai/scripts/format.sh

sync

.claude/settings.json (excerpt)

{
  "env": {
    "AGNOSTIC_AI_TARGET": "claude"
  },
  "hooks": {
    "PostToolUse": [
      {
        "hooks": [
          {
            "command": ".claude/hooks/format.sh",
            "type": "command"
          }
        ],
        "matcher": "Edit|MultiEdit|Write|NotebookEdit"
      }
    ]
  }
}

.codex/hooks.json

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Edit|Write",
        "hooks": [
          {
            "type": "command",
            "command": "export AGNOSTIC_AI_TARGET=codex; .codex/hooks/format.sh",
            "commandWindows": ".codex/hooks/format.sh"
          }
        ]
      }
    ]
  }
}

.gemini/settings.json (excerpt)

{
  "hooks": {
    "AfterTool": [
      {
        "hooks": [
          {
            "command": ".gemini/hooks/format.sh",
            "env": {
              "AGNOSTIC_AI_TARGET": "gemini"
            },
            "type": "command"
          }
        ],
        "matcher": "^(write_file|replace)$"
      }
    ]
  }
}

Your own specs on this machine, ignored by Git. Sync layers them over the shared specs.

.agnostic-ai/local/rules/my-style.md

---
name: my-style
alwaysApply: true
---

Answer in short sentences.

sync

.claude/rules/my-style.md

---
name: my-style
---

Answer in short sentences.

AGENTS.md (excerpt)

### my-style

Answer in short sentences.

GEMINI.md (excerpt)

### my-style

Answer in short sentences.

Docs: Local overrides

MCP server connections, one YAML file per server. Secrets stay in environment variables.

.agnostic-ai/mcps/github.yaml

name: github
type: http
url: https://api.githubcopilot.com/mcp/
headers:
  Authorization: Bearer ${GITHUB_TOKEN}

sync

.mcp.json

{
  "mcpServers": {
    "github": {
      "headers": {
        "Authorization": "Bearer ${GITHUB_TOKEN}"
      },
      "type": "http",
      "url": "https://api.githubcopilot.com/mcp/"
    }
  }
}

.codex/config.toml (excerpt)

[mcp_servers]
  [mcp_servers.github]
    bearer_token_env_var = "GITHUB_TOKEN"
    url = "https://api.githubcopilot.com/mcp/"

.gemini/settings.json (excerpt)

{
  "mcpServers": {
    "github": {
      "headers": {
        "Authorization": "Bearer ${GITHUB_TOKEN}"
      },
      "httpUrl": "https://api.githubcopilot.com/mcp/"
    }
  }
}

Docs: MCP servers

Native keys no spec covers, one file per tool: the Claude Code status line, Codex notification settings. Sync merges each into that tool's own file.

.agnostic-ai/overlays/claude.settings.json

{
  "statusLine": {
    "type": "command",
    "command": "~/.claude/statusline.sh"
  }
}

.agnostic-ai/overlays/codex.config.toml

[tui]
notification_condition = "always"

sync

.claude/settings.json (excerpt)

{
  "statusLine": {
    "command": "~/.claude/statusline.sh",
    "type": "command"
  }
}

.codex/config.toml (excerpt)

[tui]
  notification_condition = "always"

Gemini CLI has no overlay file. Put its native keys under x-gemini in a settings/ spec, and sync merges them into .gemini/settings.json.

Conventions, always on or tied to file patterns and folders.

.agnostic-ai/rules/testing.md

---
name: testing
globs: ["**/*.test.ts"]
---

One behavior per test. Mock HTTP with `msw`.

sync

.claude/rules/testing.md

---
name: testing
paths:
  - "**/*.test.ts"
---

One behavior per test. Mock HTTP with `msw`.

AGENTS.md (excerpt)

### testing

Applies to: `**/*.test.ts`

One behavior per test. Mock HTTP with `msw`.

GEMINI.md (excerpt)

### testing

Applies to: `**/*.test.ts`

One behavior per test. Mock HTTP with `msw`.

Hook scripts. Sync copies each one into each tool's hooks folder and points the hook there.

.agnostic-ai/scripts/format.sh

#!/bin/sh
files=$(agnostic-ai hook paths) || exit 1
printf '%s\n' "$files" | while IFS= read -r f; do
  [ -z "$f" ] || npx prettier --write "$f"
done

sync

.claude/hooks/format.sh

#!/bin/sh
files=$(agnostic-ai hook paths) || exit 1
printf '%s\n' "$files" | while IFS= read -r f; do
  [ -z "$f" ] || npx prettier --write "$f"
done

.codex/hooks/format.sh

#!/bin/sh
files=$(agnostic-ai hook paths) || exit 1
printf '%s\n' "$files" | while IFS= read -r f; do
  [ -z "$f" ] || npx prettier --write "$f"
done

.gemini/hooks/format.sh

#!/bin/sh
files=$(agnostic-ai hook paths) || exit 1
printf '%s\n' "$files" | while IFS= read -r f; do
  [ -z "$f" ] || npx prettier --write "$f"
done

Docs: Shared hook scripts

The default model and effort per tool, and permissions.

.agnostic-ai/settings/model.yaml

model:
  claude: opus
  codex: sol
  gemini: gemini-3.1-pro-preview
effort:
  claude: xhigh
  codex: high

.agnostic-ai/settings/permissions.yaml

permissions:
  allow: [shell(pnpm test:*)]
  ask: [shell(git push:*)]
  deny: [read(.env)]

sync

.claude/settings.json (excerpt)

{
  "effortLevel": "xhigh",
  "model": "opus",
  "permissions": {
    "allow": [
      "Bash(pnpm test:*)"
    ],
    "ask": [
      "Bash(git push:*)"
    ],
    "deny": [
      "Read(.env)"
    ]
  }
}

.codex/config.toml (excerpt)

model = "gpt-6.1-sol"
model_reasoning_effort = "high"

.gemini/settings.json (excerpt)

{
  "model": {
    "name": "gemini-3.1-pro-preview"
  }
}

Codex and Gemini CLI have no portable permission list, so sync prints a note instead of guessing. Sync writes no effort for Gemini CLI; set its thinking level under x-gemini.

Reusable workflows. Files beside SKILL.md, such as check.sh, copy with it. disable-model-invocation: true runs it only when you ask; Codex reads that from agents/openai.yaml.

.agnostic-ai/skills/release/SKILL.md

---
name: release
description: Cut a release and verify it. Use only when asked to release.
disable-model-invocation: true
---

Run `./check.sh`, bump the version, and push the tag.

sync

.claude/skills/release/SKILL.md

---
name: release
description: Cut a release and verify it. Use only when asked to release.
disable-model-invocation: true
---

Run `./check.sh`, bump the version, and push the tag.

.agents/skills/release/agents/openai.yaml

policy:
    allow_implicit_invocation: false

.gemini/skills/release/SKILL.md

---
name: release
description: Cut a release and verify it. Use only when asked to release.
---

Run `./check.sh`, bump the version, and push the tag.

Gemini CLI has no setting that keeps a skill for when you ask, so sync drops disable-model-invocation there and prints a note.

Project settings, beside .agnostic-ai/. targets picks the tools sync writes for. A tier names a model role once: an agent with model: fast gets the model the tier names for each tool, and a Codex alias becomes the pinned id.

agnostic-ai.yaml

version: 1
targets:
  - claude
  - codex
  - gemini
models:
  fast:
    claude: haiku
    codex: luna
    gemini: gemini-3.5-flash
    effort: low

sync

.claude/agents/test-writer.md (excerpt)

model: haiku
effort: low

.codex/agents/test-writer.toml (excerpt)

model = "gpt-6-luna"
model_reasoning_effort = "low"

.gemini/agents/test-writer.md (excerpt)

model: gemini-3.5-flash

Gemini CLI agents have no effort setting, so sync leaves out the tier's effort there and prints a note.

Docs: Model tiers

Keep the setup current. Know what changed.

What shipped in agnostic-ai, and what changed upstream in the tools it targets.

Read the latest briefing Browse all updates Subscribe by RSS

Latest briefings

29 October 2026

agnostic-ai v0.90.0: per-tool settings and safer sync

27 October 2026

agnostic-ai v0.89.0: review context and edit settings in your browser

25 October 2026

agnostic-ai v0.88.0: inspect and change agent setup in your browser

Keep your instructions together.
Make the next edit once.

agnostic-ai sync

Read getting started Set up with an agent