code-agnostic 0.3.10

Centralized hub for LLM coding config: MCP, skills, rules, and agents.

code-agnostic

One config, every AI editor. Keep MCP servers, rules, skills, and agents in a single hub and sync them into editor-specific layouts.

Why

AI coding tools each want config in a different place and format. When you use more than one, you end up copy-pasting MCP servers, duplicating rules, and manually keeping things in sync. code-agnostic removes that overhead: define once, sync everywhere.

How it works

~/.config/code-agnostic/          Your single source of truth
├── config/
│   └── mcp.base.json             MCP servers (editor-agnostic)
├── rules/
│   └── python-style/
│       ├── meta.yaml             Rule metadata
│       └── prompt.md             Rule instructions
├── skills/
│   └── code-reviewer/
│       ├── meta.yaml             Skill metadata
│       └── prompt.md             Skill instructions
└── agents/
    └── architect/
        ├── meta.yaml             Agent metadata
        └── prompt.md             Agent instructions

        ↓ plan / apply ↓

~/.config/opencode/               Compiled & synced for OpenCode
~/.cursor/                        Compiled & synced for Cursor
~/.codex/                         Compiled & synced for Codex
~/.claude.json and ~/.claude/      Compiled & synced for Claude Code

Each resource is cross-compiled to the target editor's native format. Rules become .mdc files for Cursor, AGENTS.md sections for OpenCode/Codex, and CLAUDE.local.md memory for Claude Code.

Legacy single-file rules, skills/<name>/SKILL.md, and markdown agents are still supported for migration, but bundle directories are the preferred source format for new config.

Today the implementation is still mixed: some assets are compiled and some are symlinked. The active migration plan is to move to generated outputs everywhere with a strict compiler contract instead of implicit per-app behavior.

Scope model

code-agnostic has two managed source scopes today:

• global source config under ~/.config/code-agnostic/, synced to enabled user-level app config;
• workspace source config under ~/.config/code-agnostic/workspaces/<name>/, propagated into repos inside a registered workspace.

Project-local skill folders that users create directly inside a repo, such as .agents/skills or .opencode/skills, are app-native inputs but are not managed as source by code-agnostic yet. First-class project-scoped installs are planned so a single registered project can have managed local source config without bypassing the hub.

Install

uv tool install code-agnostic

Or run without installing:

uvx code-agnostic

Or run the published Docker image to isolate filesystem access to mounted paths only:

docker run --rm -it \
  -v "$(pwd):/workspace" \
  -w /workspace \
  ghcr.io/dhvcc/code-agnostic:latest plan

By default, config stays inside the container at /root/.config unless you mount a host path.

Quick start

# Import existing config from an editor you already use
code-agnostic import plan -a codex
code-agnostic import apply -a codex

# Enable target editors
code-agnostic apps enable -a cursor
code-agnostic apps enable -a opencode
code-agnostic apps enable -a claude

# Preview and apply
code-agnostic validate
code-agnostic plan
code-agnostic apply

Editor compatibility

Feature	OpenCode	Cursor	Codex	Claude Code
MCP sync	yes	yes	yes	yes
Rules sync (cross-compiled)	yes	yes	yes	yes
Skills sync	yes	yes	yes	yes
Agents sync	yes	yes	yes	yes
Workspace root AGENTS.md link	yes	yes	yes	yes
Native repo config include for workspace AGENTS.md	yes	--	--	--
Repo/subdir gets shared workspace instructions today	yes	--	yes	yes
Nested AGENTS.md discovery	--	yes	yes	--
Workspace propagation	yes	yes	yes	yes
Import from	yes	yes	yes	yes
Interactive import (TUI)	yes	yes	yes	yes

yes means the resource type is synced for that editor. Some metadata is still target-specific or lossy; run code-agnostic explain-lossiness to see fields that are omitted or rejected for a selected target.

Cursor workspace propagation writes repo-local MCP, skills, and agents when those resources exist in the workspace source config.

OpenCode workspace configs include the shared workspace AGENTS.md natively via instructions, so repos under the workspace get both repo-local and shared workspace instructions. Codex repos receive workspace instructions through a generated AGENTS.override.md, which is added to each repo's .git/info/exclude. Claude Code receives workspace instructions through generated CLAUDE.local.md files, never by editing committed CLAUDE.md.

Cursor documents AGENTS.md support in project roots and subdirectories. code-agnostic does not copy or link the shared workspace AGENTS.md into child repos; Cursor will load repo-local or nested AGENTS.md files that already exist in the opened project. Codex documents nested AGENTS.md discovery, but not a native config include for an extra workspace file.

Features

Sync engine

Plan-then-apply workflow. Preview every change before it touches disk.

code-agnostic validate              # check canonical source files
code-agnostic plan -a cursor        # dry-run for one editor
code-agnostic plan                   # dry-run for all
code-agnostic apply                  # apply changes for all enabled editors
code-agnostic status                 # check drift and disabled app states
code-agnostic explain-lossiness      # show fields omitted or rejected per editor

Bare plan and apply target every enabled editor; bare status also shows disabled app states. Use -a codex, -a cursor, -a opencode, or -a claude when you want one editor at a time.

If managed outputs need repair after an apply, restore the active synced revision:

code-agnostic restore
code-agnostic restore -w myproject

MCP management

Add, remove, and list MCP servers without editing JSON by hand.

code-agnostic mcp add github --command npx --args @modelcontextprotocol/server-github --env GITHUB_TOKEN
code-agnostic mcp list
code-agnostic mcp remove github

Env vars without a value (--env GITHUB_TOKEN) are stored as ${GITHUB_TOKEN} references.

Rules with metadata

New rules should use bundle directories with schema-validated metadata and a separate prompt body:

rules/python-style/
├── meta.yaml
└── prompt.md

# rules/python-style/meta.yaml
spec_version: v1
kind: rule
description: "Python coding standards"
globs: ["*.py"]
always_apply: false

<!-- rules/python-style/prompt.md -->
Always use type hints. Prefer dataclasses over dicts.

Cross-compiled per editor: Cursor gets .mdc files with native frontmatter, OpenCode/Codex get AGENTS.md sections. Legacy single-file rule markdown with YAML frontmatter remains supported for migration.

code-agnostic rules list
code-agnostic rules remove --name python-style

Skills and agents

Use bundle directories for new skills and agents, then let code-agnostic cross-compile them per editor. Install or edit skills in the code-agnostic source of truth, then run plan / apply; do not hand-copy generated skills into .codex, .cursor, .agents, or OpenCode directories.

code-agnostic skills list
code-agnostic agents list

Manual skill install today:

mkdir -p ~/.config/code-agnostic/skills
cp -R ./my-skill ~/.config/code-agnostic/skills/my-skill
code-agnostic plan
code-agnostic apply

Global skills live under ~/.config/code-agnostic/skills. Workspace-local skills live under ~/.config/code-agnostic/workspaces/<name>/skills and can be inspected with code-agnostic skills list -w <name>. Codex generated skill outputs are written to ~/.agents/skills, while Codex agents and config remain under ~/.codex. Claude Code generated skills and agents are written under ~/.claude/skills and ~/.claude/agents, with workspace copies under repo-local .claude/skills and .claude/agents.

Project-local skills are not first-class source inputs in code-agnostic yet. If a target app discovers repo-local skill folders such as .agents/skills, .opencode/skills, or user-created .claude/skills, treat those as unmanaged app inputs. Workspace sync writes only the exact generated paths recorded in .sync-state.json.

Planned convenience command:

code-agnostic skills install ./my-skill --apply

That command should copy the skill into the source of truth and then run the normal compiler/apply flow.

Workspaces

Register workspace directories. Workspace rules are compiled into a canonical AGENTS.md at the workspace root. Repos keep their own repo-specific AGENTS.md; Codex receives the workspace rules through generated, git-excluded AGENTS.override.md files, while OpenCode workspace configs reference the shared workspace file through instructions. Claude receives generated CLAUDE.local.md files and project MCP entries in ~/.claude.json["projects"][absolute_repo_path]["mcpServers"]. Workspace source config, skills, and agents are propagated into repo-local generated paths for OpenCode, Cursor, Codex, and Claude; user-created project-local skill folders remain unmanaged until project-scoped installs are supported.

Cursor propagation intentionally stays to repo-local MCP, skills, and agents; it does not copy the shared workspace AGENTS.md into child repos.

code-agnostic workspaces add --name myproject --path ~/code/myproject
code-agnostic workspaces list

Git exclude

Prevent synced paths from showing up in git status. Managed per-workspace with customizable patterns.

code-agnostic workspaces git-exclude                            # all workspaces
code-agnostic workspaces git-exclude -w myproject               # one workspace
code-agnostic workspaces exclude-add --pattern "*.generated" -w myproject
code-agnostic workspaces exclude-list -w myproject

Import

Migrate existing config from any supported editor into the hub.

code-agnostic import plan -a codex
code-agnostic import apply -a codex
code-agnostic import plan -a claude
code-agnostic import apply -a cursor --include mcp --on-conflict overwrite
code-agnostic import plan -a codex -i    # interactive TUI picker

import plan previews what will be copied into the hub; import apply writes only the selected sections. Conflicts are skipped by default, so rerun with --on-conflict overwrite only after reviewing the preview. Use --include, --exclude, --source-root, and --follow-symlinks to narrow what gets imported.

CLI conventions

All commands use named flags (-a, -w, -v). Singular aliases work too: app = apps, workspace = workspaces.

Compiler docs

The compiler migration is documented in:

• docs/compiler/overview.md
• docs/compiler/skills.md
• docs/compiler/agents.md
• docs/compiler/rules.md
• docs/compiler/mcp.md
• docs/compiler/lossiness.md

Roadmap

• Plan/apply/status sync engine
• MCP server sync across editors
• Skills and agents sync across editors
• Workspace propagation into git repos
• Import from existing editor configs
• Consistent CLI with named flags and aliases
• MCP add/remove/list commands
• Rules system with YAML frontmatter and per-editor compilation
• Cross-compilation for skills and agents
• Per-workspace git-exclude customization
• Interactive TUI for import selection
• Claude Code support
• Project-scoped skill installs and sync
• rules add / skills add / agents add commands (open $EDITOR with template)
• Planner integration for cross-compiled skills and agents
• Shell auto-complete
• Full TUI mode (command palette + menus)

Testing

uv sync --dev
uv run pytest

Before pushing release-prep work, run the supported Python matrix locally. Tox delegates each environment to uv run --python, so uv can provide the requested interpreter when it is not already installed:

uvx tox run -p auto
uvx tox run -e uv310 -- tests/test_version.py -q

For a hermetic Linux matrix that does not depend on locally installed Python versions, run the Docker matrix:

./scripts/run-docker-matrix.sh
./scripts/run-docker-matrix.sh tests/test_version.py -q

Limit the Docker matrix while iterating:

PYTHON_VERSIONS="3.10 3.14" ./scripts/run-docker-matrix.sh tests/test_version.py -q

The release gate still requires GitHub Actions to pass because the published workflow is the source of truth for OS coverage across Ubuntu, macOS, and Windows on every supported Python version.

Real app-ingestion E2E is gated because it requires installed target CLIs and uses each tool's own introspection surface:

CODE_AGNOSTIC_REAL_APP_E2E=1 uv run pytest tests/e2e/test_real_app_ingestion_e2e.py -q
CODE_AGNOSTIC_REAL_APP_E2E=1 CODE_AGNOSTIC_REAL_APP_TARGETS=codex,opencode,claude uv run pytest tests/e2e/test_real_app_ingestion_e2e.py -q