@contextgraph/agent
v0.6.3
Published
Autonomous agent for contextgraph action execution
Maintainers
Readme
steward
Local CLI for steward.foo.
Install (one command)
npx -y @contextgraph/agent@latest installsteward install is an interactive wizard: it authenticates you, detects which
coding agents you have (Claude Code, Cursor, Codex CLI, VS Code, Gemini CLI,
Windsurf, …), and configures the steward MCP server plus the steward
skills for each one.
- It asks whether to install globally (
~/, follows you across every project) or for this project (committable, shared with your team). - Skills are written to the Agent Skills
open-standard location (
.agents/skills) and mirrored into.claude/skills, so they work across the whole ecosystem. - For Claude Code it installs the plugin (MCP + skills +
/steward:commands). - For UI-only destinations (claude.ai, ChatGPT) it prints the connector steps.
install sets up the connection (MCP server) and skills — not a steward
itself. After it finishes, restart your agent (or reload MCP servers), authorize
the steward server when prompted (/mcp in Claude Code), then create your first
steward from inside the agent (/steward:define-steward in Claude Code) or the
dashboard.
Useful flags:
steward install --print # show copy-paste config, write nothing
steward install --scope project # skip the scope prompt
steward install --client cursor codex # target specific agents
steward install --no-mirror # write skills only to .agents/skillsFor non-interactive environments, set STEWARD_API_TOKEN and pass --scope/--client.
Other commands
steward --help
steward backlog --help
steward backlog top
steward review <repository> <sha>
steward consult <repository> --message "..."steward auth opens a browser login and stores credentials locally in ~/.steward/credentials.json.
Tell Your Agent
Append this to CLAUDE.md or AGENTS.md in your project:
echo "Use the \`steward\` CLI for steward.foo work. Learn the workflow from \`steward --help\` and \`steward backlog --help\`." >> CLAUDE.mdReleases
Versioning and npm publishing are automated by release-please. Use Conventional Commits on main so the right semver bump and changelog entries are picked up:
fix: ...→ patch bumpfeat: ...→ minor bumpfeat!: ...or aBREAKING CHANGE:footer → major bump (minor while pre-1.0.0)chore:,docs:,refactor:,test:,ci:→ no release on their own
When releasable commits land on main, release-please opens a rolling "Release PR" with the proposed version bump and changelog. Review and edit that PR's notes if you want to adjust user-facing wording — merging it tags the release and publishes to npm via Trusted Publishers (OIDC, no token).
