@origozero/zeromind
@origozero/zeromind is two things: a declaration of the remote ZeroMind MCP server (https://origozero.ai/mcp) written into your agent harness's own config, and the skills that teach the agent to check ZeroMind before it builds anything. Nothing runs on your machine — the CLI links the machine to your ZeroMind account, writes the server entry, and exits. Every tool the agent then calls (ZeroMind content discovery, worlds, and the engine of the world you have open) is served by ZeroMind over HTTPS, so a new tool is available the moment ZeroMind ships it, with nothing to upgrade here.
Each supported harness (Claude Code, Cursor, Codex CLI, Gemini CLI, OpenCode, Cline, Continue, Windsurf, Zed, openClaw, JetBrains Junie, Sourcegraph Amp, GitHub Copilot, Block Goose, Aider) gets its onboarding through that harness's own native channel — skills, AGENTS.md, GEMINI.md, .cursor/rules/*.mdc, .clinerules, CONVENTIONS.md, .goosehints — plus the MCP entry in the config file that harness reads. A harness with no custom-crafted integration uses the generic streamable-HTTP wiring in ide/README.md.
What the agent gets, once linked: the ZeroMind library (search, inspect, preview, install, engage), worlds (list, create, fork, launch, connect, and the reversible delete / trash / restore), and — for the world it is connected to — the engine itself (execute, guides, capture, the VFS, bash). tools/list on https://origozero.ai/mcp is the current surface; it is served rather than shipped, so this README does not enumerate it.
Prerequisites
Node.js 18 or newer, on your PATH, to run zeromind link, zeromind install, zeromind upload and zeromind open. Your harness then talks to https://origozero.ai/mcp over HTTPS and spawns nothing locally.
- macOS:
brew install node(or download from https://nodejs.org) - Linux: your distro's package manager, or https://nodejs.org / nvm
- Windows: https://nodejs.org (LTS installer) — restart your IDE after install so it picks up the new PATH
Verify with node --version.
Install
Claude Code
/plugin marketplace add OrigoZero/zeromind-plugin
/plugin install zeromind
npx -y @origozero/zeromind link
The bundle carries the skills and an .mcp.json whose Authorization header expands ${ZEROMIND_INSTALL_SECRET}; zeromind link writes that variable into ~/.claude/settings.json env. Restart Claude Code afterwards so the new setting is read.
Every other harness
npx -y @origozero/zeromind install <harness>
One command per harness — it links the machine first if it isn't linked yet, writes the harness's native instruction file, and writes the /mcp entry into that harness's own MCP config:
npx -y @origozero/zeromind install claude # .claude/skills/zeromind-{getting-started,library}/SKILL.md
npx -y @origozero/zeromind install cursor # .cursor/rules/zeromind.mdc
npx -y @origozero/zeromind install codex # ~/.codex/AGENTS.md (or ./AGENTS.md)
npx -y @origozero/zeromind install gemini # ~/.gemini/GEMINI.md (or ./GEMINI.md)
npx -y @origozero/zeromind install opencode # .opencode/skills/zeromind/SKILL.md
npx -y @origozero/zeromind install cline # .clinerules/zeromind.md
npx -y @origozero/zeromind install continue # .continue/rules/zeromind.md
npx -y @origozero/zeromind install windsurf # ./AGENTS.md
npx -y @origozero/zeromind install zed # .claude/skills/zeromind/SKILL.md
npx -y @origozero/zeromind install openclaw # skills/zeromind/SKILL.md
npx -y @origozero/zeromind install junie # ./AGENTS.md
npx -y @origozero/zeromind install amp # ./AGENT.md
npx -y @origozero/zeromind install copilot # .github/copilot-instructions.md
npx -y @origozero/zeromind install goose # ~/.config/goose/.goosehints
npx -y @origozero/zeromind install aider # ./CONVENTIONS.md
npx -y @origozero/zeromind install hermes # ~/.hermes/config.yaml + plugin bundle
npx -y @origozero/zeromind install --list # enumerate
Every command is idempotent: shared files (AGENTS.md, GEMINI.md, CONVENTIONS.md, .goosehints, copilot-instructions.md) carry a delimited `` block that is replaced in place; owned files (skills, rule files) need --force to overwrite. Two harnesses have no MCP config path to write into — openClaw's is unconfirmed, and Aider takes servers on the command line — so the install prints their entry for you to paste. ide/README.md has the full table, with a README per harness.
Linking
A machine links once, and the link is the same one the Zero engine uses: both read and write session.json in the OS config dir (~/.config/zero/session.json; %APPDATA%\zero\session.json on Windows; mode 0600 off Windows). Linking from either side links both, and unlinking revokes every harness on the machine at once.
npx -y @origozero/zeromind link # device code, approved in the browser
npx -y @origozero/zeromind link --username <handle> # pre-fill the approval page's agent name
npx -y @origozero/zeromind status # what this machine is linked as
npx -y @origozero/zeromind unlink # revoke it
link registers this machine as its own ZeroMind principal, prints a URL and a code, and waits for you to approve it on that page. The approved account is the agent's identity — its own ZeroMind account, distinct from yours.
On a machine that already holds an approved install — linked by an earlier run, or by the Zero engine, which writes the same file — link prints who the machine acts as, asks for no new code, and still hands the secret to Claude Code. One linked install per machine is the whole rule; unlink, then link, is how a machine changes which account it acts as.
unlink revokes the install at ZeroMind and removes it from this machine's cache — the install fields only, so a session the Zero engine signed in with, and the issuer it signed in against, stay where they are. If ZeroMind answers that it holds no such install, the local clear happens anyway; if the revoke cannot be made at all, nothing is cleared and the command tells you to run it again once ZeroMind is reachable.
unlink also takes ZEROMIND_INSTALL_SECRET back out of ~/.claude/settings.json when it holds the secret being revoked — before it revokes anything, so a settings file it cannot read is reported and the unlink still completes. (That one line is then yours to delete.) The entries install <harness> wrote into other harnesses keep the revoked secret and answer 401 until you link again and re-run the install for that harness — the command says so when it runs.
The credential a harness presents is that install's secret, sent as Authorization: Bearer ins_sec_…. A 401 / invalid_token from any tool means the machine is not linked (or the entry holds a stale secret): run link, re-run install <harness> so the entry is rewritten, and restart the harness.
Uploading local files into a world
npx -y @origozero/zeromind upload <path> --world <name-or-guid> [--to <vfs dir>]
Copies a file — or a folder, its relative layout preserved — from this machine into the world's engine VFS, under /source unless --to says otherwise. The bytes go from disk to the world's write_file as base64, so an asset pack never passes through an agent's context window. Ceilings of 256 MiB and 10 000 files are checked before anything is written (--max-bytes / --max-files raise them), and a world with no engine running is reported rather than half-written.
Opening a world on this machine
npx -y @origozero/zeromind open <world guid | name | https url | zero:// url> [--play | --edit] [--native | --browser] [--dry-run]
Every tool an agent calls is served by ZeroMind, which has no browser tab and no engine process on your machine: world.open and world.launch answer where a world opens, and this command is what opens it here. It opens the world's editor by default and the player's face with --play; it hands the URL to the desktop engine when this machine has a zero:// handler registered and to a browser tab when it does not. --native / --browser decide it outright and are remembered in open.json beside session.json, so later calls with neither flag follow what you chose. --dry-run prints the target and the URL and changes nothing.
A guid or a URL is opened without a single network call. A name is resolved against your own worlds over /mcp, with this machine's install secret — which is never printed. No engine is contacted at all.
Pointing at a local / self-hosted ZeroMind
Two environment variables move the whole CLI to another backend:
For the local ZeroMind dev stack, point the issuer at the front door (http://127.0.0.1:3003), not the bare API on :3001 — the front door serves the /v1 REST surface, /mcp, and the web app (including /link and the /edit/<guid> engine pages) on one origin, which is what the CLI assumes.
ZEROMIND_ISSUER=http://127.0.0.1:3003 \
ZEROMIND_CONFIG_DIR=$HOME/.config/zero-local \
npx -y @origozero/zeromind link
ZEROMIND_ISSUER=http://127.0.0.1:3003 \
ZEROMIND_CONFIG_DIR=$HOME/.config/zero-local \
npx -y @origozero/zeromind install claude
The entry that lands is the production shape at the local address:
// .mcp.json
{
"mcpServers": {
"zeromind": {
"type": "http",
"url": "http://127.0.0.1:3003/mcp",
"headers": {
"Authorization": "Bearer ins_sec_...",
"X-ZM-Harness": "claude-code"
}
}
}
}
Or with the Claude Code CLI:
claude mcp add --transport http zeromind http://127.0.0.1:3003/mcp \
--header "Authorization: Bearer ins_sec_..." \
--header "X-ZM-Harness: claude-code"
Plain http:// issuers work end to end: auth is a Bearer header on every request, with no cookies and no TLS-only assumption.
Updating
- The tools are served from
https://origozero.ai/mcp. Every harness reaches the running version; there is nothing to upgrade on your machine. - The artifacts
zeromind install <harness>wrote — instruction blocks, skills, the MCP entry — refresh by re-running the install command. - The Claude Code bundle (skills +
.mcp.json) updates through/plugin.
Maintainers: package.json version is the source of truth. tests/directory-manifests.test.ts fails when a manifest it names carries a different one.
Status
0.7.1 — the plugin runs no MCP server of its own. It links a machine, writes the remote /mcp entry into sixteen harnesses, uploads local files into a world, opens a world on this machine, and ships the two skills. The tool surface is ZeroMind's, served live at https://origozero.ai/mcp; the tests drive the CLI end to end against the bundled mock ZeroMind server in tools/mock-zeromind/.
This release needs a Zero engine that keeps install_secret in the shared session.json it writes: an engine whose cached-session struct has no such field re-serialises the file without it, and the machine reads as unlinked the next time the CLI runs. 0.7.1 therefore ships after that engine release.
Development
npm install
npm test # builds, then runs the suite against the bundled mock ZeroMind server
npm run build # compile TypeScript
npm run lint
The canonical agent operating manual lives in templates/manual.md; every per-harness installer wraps that one file with the harness's expected frontmatter. Adding a harness is one entry in src/cli-install.ts plus an ide/<harness>/README.md.
Directory manifests
Four plugin directories read a manifest from the root of this repository:
Nothing generates these files or the bundles under dist-publishing/. tests/directory-manifests.test.ts runs in CI and fails when a root manifest and its bundle disagree, when a path one of them names does not resolve, or when a bundle's copy of skills/ differs from the original.
Releasing
- Bump
versioninpackage.jsonand in every manifesttests/directory-manifests.test.tschecks against it. git tag vX.Y.Z && git push --tags— the publish workflow runs.
Requires the NPM_TOKEN secret in repo settings, scoped to the @origozero npm org.
Security
SECURITY.md says how to report a vulnerability.
License
MIT. See LICENSE.