Agnostic AI

Your AI coding harness, ported to every client you use.

You spent months tuning Claude Code: a working agreement in CLAUDE.md, guard hooks that block secrets and destructive commands, skills, custom subagents, slash commands, MCP servers. Then you open Codex, or Cursor, or Gemini CLI, and none of it is there. Agnostic AI captures that harness once and applies it, identically, to every other client on the machine. Switch models and tools without switching how the agent behaves.

git clone https://github.com/ucsandman/Agnostic-AI.git && cd Agnostic-AI
npm run port

That is the whole install. No npm dependencies, no Python, nothing phones home. Node 18 or newer.

What gets ported

Codex CLI works as the source too: the same eight components are read back from ~/.codex and written into Claude Code and the rest. The live matrix for your machine is npm run status; the generated per-client table is docs/targets.md.

Supported clients: Claude Code, Codex CLI, Gemini CLI, Antigravity CLI, Cursor, Windsurf, GitHub Copilot, Cline, Aider, OpenHands, Goose, Continue, Zed, OpenCode, Trae, Amazon Q, Sourcegraph Cody, OpenClaw, Hermes, and a generic system prompt for any local or API model.

How it works

capture  ~/.claude  ──▶  harness/  ──▶  apply  ~/.codex  ~/.gemini  ~/.cursor  ...
         (the client        (client-neutral        (each client's own dialect,
          you use)           bundle)                same guard scripts, same skills)

1. Capture reads the client you use into a client-neutral bundle: markdown rules with every @import inlined, hooks in one canonical dialect, the list of skills, agents, commands, MCP servers and permissions. A credential never enters the bundle: an env value that looks like a token is replaced with ${NAME} and you are told which variable to export.
2. Apply renders the bundle into every other installed client, in that client's dialect. Hooks are not copied; each client is pointed at the same scripts, wrapped in a shim where the client speaks another payload format. Codex gets its hook trust hashes pre-computed so nothing asks for a /hooks review. Skills are linked, not copied, so an edit is live everywhere at once.
3. Nothing is destroyed. Generated files carry a header that marks them as the port's; user-owned files (config.toml, settings.json, mcp.json) get a marked region or per-key ownership and are otherwise preserved byte for byte. Every overwrite is backed up first. --check exits 1 on drift.
4. Every drop is explained. A hook that only makes sense in one client, a skill that needs a tool only one client has, an MCP server bound to one OAuth grant: each is listed by npm run explain with its reason, and the list lives in core/port.json where you can change it.

Details, dialect tables and the not-ported list: docs/porting.md.

Commands

Flags: --from claude|codex, --to codex,gemini, --check, --dry-run, --force, --home <dir>, --json. Full list: docs/configuration.md.

Also in the box

• Safety policy. core/safety/guards.json is one file read by the shipped guards (secret-guard, dashclaw-guard) and the dashboard simulator. Secret paths are always blocked; hard-stop commands need a human; a missing or unreachable policy fails closed. npm run setup:default installs the guards into your primary client, and the port carries them everywhere else.
• A learning loop. npm run harvest collects errors and corrections from local agent logs into candidate rules; npm run distill runs a promotion ladder (observation, fact, rule) and writes a proposal; you approve from the dashboard and the rule lands in your working agreement.
• Authoring mode. Prefer to keep the working agreement in this repo? Write core/rules/global-rules.md, run npm run sync to compile it into your primary client, then npm run port.
• Human surfaces. npm run dashboard (command center: candidates, rules, skills matrix, project recommendations, DashClaw settings, guard simulator), npm run recall (search rules and memory), npm run parity (port status). All bind 127.0.0.1; mutating routes need a per-process token and a loopback origin.
• Governed autonomy (optional). Point DASHCLAW_BASE_URL at a DashClaw instance and risky calls are held for remote approval.

Documentation

Repository layout

engine/harness/   capture.cjs, apply.cjs, status.cjs, cli.cjs, bundle.cjs, toml.cjs, common.cjs
                  sources/ (claude, codex)   targets/ (codex, claude, gemini, agy, cursor, generic)
engine/hooks/     shim.cjs (dialect translation), universal-adapter.cjs, secret-guard, dashclaw-guard,
                  fable-delegate-guard, capability-graph-guard, correction-tracker
engine/           sync/ (authoring mode), harvest/, distill/, ingest/, skills/, audit/, setup/, docs/, tests/
core/             port.json (policy), templates/targets.json (registry), safety/guards.json,
                  rules/ + traits/ (authoring mode), examples/
tools/            Local web UIs: sync/ (port status), dashboard/, recall/
harness/          Your captured bundle (gitignored; holds machine paths, never secrets)
storage/          Runtime state (gitignored): ownership, reports, backups of every overwritten file
jobs/             PowerShell wrappers for a scheduled port and the nightly distill

Development

npm test               # engine suite + sync, hook, wire-protocol and port regressions
npm run docs:check     # generated docs are current

Every test builds a throwaway home directory; nothing in the suite touches yours. CI runs on Ubuntu (Node 18 and 22) and Windows (Node 22). See CONTRIBUTING.md.

Using this repository as a template

Click Use this template on GitHub, clone your copy, run npm run port. Edit core/port.json to change the source client, restrict the targets, or exclude a hook, skill or MCP server with a reason. To version your harness, remove the harness/ line from .gitignore; the bundle is plain markdown and JSON and the save refuses anything that looks like a secret.

Related

• agnostic-agent: the terminal coding agent that used to live in this repo. Local or hosted models, subagents, the same safety policy.
• DashClaw: remote approvals and execution evidence for unattended agents.
• agent-capsule: move a whole Claude Code harness onto a fresh Linux box.

License

MIT. See LICENSE.