Skip to content

v0.36.0

Choose a tag to compare

@github-actions github-actions released this 16 Nov 05:16
· 157 commits to main since this release

Extensions

Extensions replace plugins. They connect 0 to your tools and services: GitHub, MCP servers and rules like scope. Each is pinned to a commit, with one decision point before every command.

Highlights

  • Extensions: Add an extension pinned to a commit with 0 skills add @. A prompt shows what it can do and the diff on update, and the new Extensions page replaces Explore.
  • Safer MCP servers and scope: MCP servers live inside extensions and start on demand, every tool is pinned and blocked until you approve a change, and scope is on by default whenever a scope is set.
  • Plugins are gone: The old plugin system, its registry and self_extend are removed. Installed plugins are kept disabled and imported once, and GitHub is now a built-in connector.

Added

  • Extensions connect 0 to your tools and services: GitHub, MCP servers and rules like scope. Each is pinned to a commit and never updates on its own, and an extension can come with a skill that explains how to use it. Install one with 0 skills add <repo>@<sha>. A prompt shows what it can do and, on update, the diff since the last approved version.
  • 0 skills extensions list|info|enable|disable|update|remove manage extensions from the CLI. A folder that no longer matches its pinned commit refuses to load, and scripts get a scrubbed environment with only the names you approved.
  • The Extensions page (/extensions) lists installed and built-in extensions, what each can do, its source at a commit and an on/off switch. Add shows the install prompt and Update shows the diff.
  • MCP servers now live inside extensions and start on demand. A thread's default tool list carries one line per server and no MCP schema; tool schemas load only when the model asks for a server. Add one from the Extensions page (Add MCP server) or 0 skills extensions mcp.
  • Skills inside an enabled extension appear on the Skills page, in list_skills and in load_skill, labelled with the extension's name. Disabled, refused or tampered extensions provide none.
  • Connectors are extensions. A connector declares how it connects, what it may read and write, the hosts it may reach and a health check, and the Extensions page shows one Connect card each. GitHub is a built-in connector. See Connect a service.
  • A one-time importer carries earlier skills and custom MCP servers into extensions, reports what it could not migrate and never deletes anything.

Changed

  • Every command now goes through one decision point (beforeCommand) covering bash, network targets and the Ask before running commands setting. Any deny wins, then any ask, and a scope denial beats a prompt.
  • Scope is a built-in extension that is on by default whenever a scope is set (--scope, a thread, Agent or automation scope, a pull request checkout). With no scope set everything is allowed. This reverses the 0.30 to 0.33 rule that a scope file alone never activates checks.
  • If the scope extension is off or errors while a scope is set, commands are denied instead of allowed. Projects that already enabled scope keep it enabled.
  • Every MCP tool is pinned by a hash of its name, description, schema and annotations. If a server changes, adds or removes a tool, the server stays blocked until you approve it again, and 0 skills extensions info <name> shows Changed, New and Removed.
  • MCP tool results are treated as untrusted: they are labelled, fenced and capped at 32 KB. Launchers such as npx, uvx and pip must pin an exact version, and remote hosts that resolve to private, loopback or metadata addresses are refused.
  • 0 skills add owner/repo goes through the extension prompt (pin it with @<full sha> to install). Catalog ids and owner/repo/path keep the earlier skill flow.

Removed

  • The old plugin system is gone. Installed plugins are kept on disk, imported once as disabled entries marked "could not migrate" and never deleted. Re-express them as scripts run through bash or as MCP servers.
  • Executable plugins: the plugin.js runtime, loader, wire protocol and the per-tool approvals for them (first shipped in 0.16.3, review UI in 0.34.0).
  • self_extend, the live harness (agent drivers, plugin views and harness trust) and the allowModelSelfExtension setting (0.14.0 to 0.16.3). The model writes a skill or script folder with bash and you install it through the install prompt.
  • The plugin registry, the hackstore index, ZERO_REGISTRY_URL, plugin submit (0.14.0), and ZERO_REGISTRY_PUBLIC_KEY_FILE with Ed25519 registry signatures (0.30.0). Signatures and publisher identity are not offered.
  • The Plugins page, plugin detail page and Explore plugin entries (0.29.0 to 0.32.0), the browser MCP sheets and service plugin adapters (0.29.0), and their /api/console routes. /plugins redirects to /extensions.
  • The Explore page and its Plugins and Themes tabs. Themes and backgrounds live in Settings > Appearance, and old Explore links open Extensions.
  • 0 skills run and the 0 plugin commands. plugin list|search|info|enable|disable|install and plugin mcp add|list|remove stay as hidden aliases that print moved: use ...; the rest print removed: ....

Fixed

  • A malformed glob in a command no longer fails the protected-paths check.
  • Extension folder checks, update diffs, imports and git fetches are stricter, and built-in enablement has one store that is locked and audited.

Security

  • An agent can no longer approve its own extensions by flag. Approval commands refuse from an agent session and without a terminal, and commands from an agent that touch approval, enablement or settings files are denied. These are layers.
  • Every executing tool (monitor, pty_session, python_exec, run_command and file tools) now goes through beforeCommand, and extension approvals are sealed.
  • A connector's stored token reaches only the scripts and MCP servers of the extension that declared and was approved for it, and is masked in what comes back.
  • MCP approval prompts show sanitised plain text with an old versus new description and schema diff. Extension server ids are namespaced (x-), and a dead process or hung call blocks the server.

Install

curl -fsSL https://raw.githubusercontent.com/0sec-labs/0security/main/install.sh | bash
0 --version

Windows (experimental): download 0-windows-x64.exe below. Docker: docker run --rm ghcr.io/0sec-labs/0security:latest --help.

Verify downloads against the attached checksums.txt (SHA-256), for example sha256sum -c checksums.txt --ignore-missing.

Full changelog: https://github.com/0sec-labs/0security/blob/v0.36.0/docs/CHANGELOG.md