contextgraph

Steward helps Claude Code define and activate repository-native stewards: narrow, rubric-centric agents that understand a specific codebase domain. It provides:

1. Repository-grounded steward definition: Inspect a codebase, draft a narrow ownership zone, and preview a rubric before writing anything
2. Agent-owned activation: Reconcile inventory and initialize the steward from the same Claude Code session
3. Persistent action management: Create, update, and track Steward actions across sessions

The plugin connects Claude Code to the Steward MCP server so steward setup happens inside the repository where the context lives.

Installation

Via the /plugin menu in Claude Code:

1. Select "Manage marketplaces"
2. Add marketplace: contextgraph/claude-code-plugin
3. Select "Browse and install plugins"
4. Install steward

Then reload plugins:

/reload-plugins

To update an existing install:

/plugin update steward

If update does not pick up the new skill, reinstall:

/plugin uninstall steward
/plugin install steward

Verify Installation

Check that the Steward MCP server is connected and authenticated:

/mcp

You should see steward listed as a configured MCP server. Complete the browser authentication handoff if Claude Code prompts for it.

Check that the steward definition skill is installed:

/steward:define-steward

Check that the backlog execution skill is installed:

/steward:work-backlog

To smoke-test the skill and MCP tool together, first confirm /mcp shows the steward server is authenticated, then open Claude Code in the repository you want the steward to watch and run:

/steward:define-steward

Run the steward onboarding preflight for this repository, then inspect this
repository and call configure_steward with action="validate" for a draft
steward. Do not create or update anything yet.

To test the full steward creation path, use:

/steward:define-steward

Inspect this repository, draft a narrow steward, preview it, and ask before
creating it. If I approve creation, continue through inventory reconciliation
and initialization preview before asking for final approval to save
initialization artifacts.

Features

• •Persistent Action Management: Create, update, and track actions across sessions
• •Repository-Native Steward Definition: Use /steward:define-steward to inspect every repository the steward will cover — including repos beyond the current checkout when the steward is scoped to several repos or the whole workspace — create a rubric-centric steward with configure_steward, reconcile inventory across all covered repos, and initialize the steward from Claude Code
• •Backlog Execution Loop: Use /steward:work-backlog to make progress on a Steward backlog. With no argument it runs a conversational session (shows the backlog, asks what you want, does the work, and hands back a command to schedule nightly); with an argument it runs directly — a specific item, a steward's top item, or a bounded batch (--max-items N --autonomy pr) suited to an unattended nightly Claude Code Routine. A run tends in-flight backlog PRs (resolving conflicts, fixing CI, dismissing already-done work), claims and implements queued items through linked PRs, and reports what it did plus the decisions left for you
• •Plan Review With Stewards: Use /steward:plan-review to consult the repository's stewards on a proposed plan before implementation, surface their attributed concerns, and propose a new steward when the plan introduces an unowned domain such as a new integration
• •Hierarchical Planning: Organize work in parent-child relationships
• •Dependency Tracking: Manage action dependencies and execution order
• •Semantic Search: Find actions using natural language queries
• •Completion Context: Rich completion stories with technical changes and outcomes
• •MCP Integration: Full Model Context Protocol support for seamless Claude Code integration

MCP Tools Available

Once installed, Claude Code can use these tools:

• •create - Create new actions with parent and dependency relationships
• •update - Update action properties including agentReady, dependencies, and completion context
• •complete - Mark actions as done with completion context
• •search - Search actions using semantic similarity and keywords
• •fetch - Get full details for specific actions
• •list_agent_runs - List agent runs for an action with filtering options
• •fetch_agent_run - Get detailed agent run information including event logs
• •move - Reorganize action hierarchy
• •suggest_parent - Get AI-powered parent suggestions
• •parse_plan - Convert unstructured text into structured actions
• •fetch_tree - View hierarchical action trees
• •prepare_steward_onboarding - Check MCP auth, workspace resolution, and GitHub App repository access before defining a steward
• •integration - Inspect workspace integration status and metric measurement capabilities
• •configure_steward - Validate, preview, create, or update rubric-centric stewards
• •manage_backlog_work - Consolidated Steward backlog execution lifecycle tool: peek, claim, release, or dismiss item/group work
• •list_steward_backlog_items - List backlog items for a steward with optional state filtering
• •create_steward_backlog_item - Create a new backlog item for a steward
• •update_steward_backlog_item - Update a steward backlog item title, objective, rationale, or priority

Skills Available

Plugin skills are namespaced by Claude Code as /steward:<skill-name>.

• •/steward:define-steward - Inspect every repository the steward will cover (not just the current checkout — a steward can be scoped to several repos, or to the whole workspace), draft a narrow steward spec, validate and preview it with configure_steward, create or update it after user approval, then reconcile inventory and initialize the steward from the coding agent. For repos beyond the current checkout the skill resolves a local checkout, offers a read-only clone, or asks where it lives before grounding anything.
• •/steward:work-backlog - Use the Steward MCP server to claim a backlog item — the repository-wide top item by default, a specific item you name (by id or reference), or the top item for a named steward — set up the registered branch/worktree, implement the work, open a PR linked by that branch, and continue through checks and review comments until the PR is merge-ready.
• •/steward:plan-review - Before implementing a plan or design, consult the repository's stewards via consult and return their attributed feedback. If the plan crosses a domain (often a new integration) that no existing steward owns, propose a narrow new steward for that domain.

Agent-Ready Status

The agentReady field allows programmatic control over whether an action subtree is ready for autonomous agent execution. This field can be toggled via the MCP update tool without requiring web UI access or direct database modifications.

Purpose

Use agentReady to control which action subtrees are available for autonomous agent execution. When set to false, the action and its entire subtree are hidden from agent selection, even if they meet other readiness criteria.

Usage Patterns

Temporarily blocking agent execution:

// Mark an action as not ready for agents
update({
  action_id: "action-uuid",
  agentReady: false
})

Re-enabling agent execution:

// Make the action available to agents again
update({
  action_id: "action-uuid",
  agentReady: true
})

Common Use Cases

• •Work in progress: Block agent execution while you refine an action's description or requirements
• •Human review required: Mark actions that need human decision-making before agent work can proceed
• •Dependency resolution: Keep actions blocked until prerequisite work or decisions are complete
• •Quality gates: Control when autonomous work can begin on specific subtrees

How It Works

When an action has agentReady: false:

• •The action won't appear in agent selection interfaces
• •Child actions are also blocked from agent execution (inherited blocking)
• •The action remains visible in the web UI and API responses
• •Manual agent execution can still be triggered via the web UI if needed

This field complements other readiness signals (dependencies, prepared status, etc.) to give you fine-grained control over autonomous execution flow.

Getting Started

After installation, Claude Code can help you:

1. Create your first action:

`` Create an action to implement user authentication ``

1. Search existing work:

`` Find actions related to database migrations ``

1. Complete work and capture context:

`` Mark action [id] as complete with the changes we just made ``

1. Create a repository-native steward:

``text /steward:define-steward `` Then ask Claude to inspect the current repo and preview a steward.

Web Dashboard

Visit steward.foo to view and manage your actions in a visual interface.

Links

• •Website: steward.foo
• •MCP Server: mcp.steward.foo

Version

Current version: 0.10.0

License

MIT