LOCAL_PLUGINS.md

Local Plugins

A Skill remains a first-class DeepCode resource. It can live independently in .agents/skills, ~/.agents/skills, or a supported compatibility root. A Plugin is an optional package that can contribute additional Skills to the same catalog; it does not replace the standalone Skill lifecycle.

Standalone Skill ─┐
                  ├── SkillCatalog → SkillRuntime
Plugin Skill ─────┘

Local Skills retain precedence when a Plugin contributes a Skill with the same name. The Plugin copy remains visible as shadowed metadata instead of changing the active local workflow.

Supported package format

DeepCode recognizes Agent Plugins 1.0.0 and its fixed Skill location:

review-tools/
├── plugin.json
├── skills/
│   ├── review/SKILL.md
│   └── verify/SKILL.md
└── mcp.json                 # optional Agent Plugins 1.0 MCP component

The smallest manifest is:

{
  "$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json",
  "name": "review-tools"
}

version and description are optional metadata. No experimental DeepCode manifest fallback exists: $schema must identify Agent Plugins 1.0.0. Skills are discovered only as immediate child directories of the fixed skills/ directory; the manifest does not list Skill paths. Unknown manifest fields are reported and ignored. Unknown extensions namespaces remain opaque and are not validated by DeepCode.

Plugin Skills use the Agent Skills contract: the frontmatter name is required, must be a lowercase slug, and must exactly match its parent directory. An invalid Skill is skipped with a diagnostic while the remaining Skills and components continue loading. Standalone DeepCode Skills retain their existing compatibility parser; this stricter package boundary does not migrate or break user Skill directories.

DeepCode validates the Agent Plugins 1.0 mcp.json contract independently of the manifest and Skills. A valid component contributes stdio, SSE, or Streamable HTTP servers to the same session-scoped MCP runtime used by standalone configuration. Invalid server entries are skipped independently; an invalid MCP component does not disable valid Plugin Skills.

Discovery remains inert: listing or registering a Plugin never imports Python code or starts a process. When an Agent Session first uses the package, DeepCode supplies immutable PLUGIN_ROOT and installation-specific PLUGIN_DATA paths, applies tool filters and permissions, then starts the MCP server. Startup failure is non-fatal unless a user policy explicitly marks the server required.

The package cannot embed DeepCode credential references. User configuration may bind credentials and narrow an active Plugin server without replacing its command, URL, arguments, or working directory:

{
  "pluginMcpServers": {
    "review-tools/analyzer": {
      "enabledTools": ["inspect"],
      "approvalMode": "writes",
      "credentialEnv": {
        "SERVICE_API_KEY": {
          "credentialRef": "provider:service"
        }
      }
    }
  }
}

Local lifecycle

deepcode plugin add ./review-tools
deepcode plugin list
deepcode plugin disable review-tools
deepcode plugin enable review-tools
deepcode plugin remove review-tools --yes

The registry lives under $DEEPCODE_HOME/plugins/registry.json (normally ~/.deepcode/plugins/registry.json). add assigns an opaque plg_... installation ID and links the canonical source directory; it does not copy it. remove unregisters the installation; it does not delete the source. Desktop uses installation IDs and exposes the same operations in the Plugins workspace.

Plugin Skills use the ordinary /skill, $name, --skill, and Desktop Composer flows. Plugin MCP tools use the ordinary mcp__server__tool catalog, approval, timeout, and cancellation flow. Disabling a Plugin invalidates idle Sessions; an already running Turn keeps its immutable Skill and tool snapshot.

Security boundary

DeepCode rejects oversized or malformed manifests, duplicate JSON keys, unsupported schemas, external symbolic-link targets, and authority/package mismatches. An invalid package contributes no Skills; invalid fixed components and individual Skills are reported independently; standalone and bundled Skills continue to work in every case.

There is no Marketplace, Git download, remote update, signing, rollback, Hook, App, or arbitrary Plugin-code hook execution in this phase. An MCP subprocess declared by a valid mcp.json is executable code and therefore runs only as a session capability, with a minimal environment and explicit user policy.