Examples

Two runnable projects. Both work with agentlift plan offline (no key); deploying and running needs ANTHROPIC_API_KEY with Managed Agents beta access.

quickstart/ — one agent, one skill, knowledge

The smallest real agent: a system prompt with an identity sign-off, a tool allowlist (read/glob/grep), one uploaded skill (receipt-stamp), and a knowledge file folded into the prompt.

agentlift plan   ./quickstart
agentlift deploy ./quickstart -y
agentlift run knowledge-agent --project ./quickstart --task "What is a North Star metric?"
agentlift run knowledge-agent --project ./quickstart --task "What is RICE?" --local   # same folder, locally
agentlift destroy ./quickstart -y

The RECEIPT: line in the output is the uploaded skill firing inside the runtime.

team/ — multi-agent, shared resources, a coordinator

Shows everything agentlift wires:

• shared skill cite-sources used by two agents → uploaded once
• shared MCP server docs (remote URL) with a specific-tool allowlist
• a coordinator (lead) with a subagents roster (bug-finder, researcher)
• a per-tool permission: bug-finder declares bash:ask, so the hosted agent pauses for caller approval before each bash call

agentlift plan ./team        # see the dedup, the coordinator ordering, the MCP wiring
agentlift deploy ./team -y
agentlift run lead --project ./team --task "Find the bug in utils.py and explain RICE."

The docs MCP server points at https://example.com/mcp (a placeholder). Swap in a real remote MCP URL before relying on its tools; the agent deploys and runs fine without ever calling it.

in-a-project/ — .managed-agents/ embedded in a real codebase

A stand-in for an actual project: a repo-level CLAUDE.md, some src/ code, and a local .claude/agents/pr-reviewer.md subagent — alongside a .managed-agents/ folder with an orchestrator coordinator over three subagents sharing two skills.

agentlift plan ./in-a-project    # only the 4 managed agents appear; the repo
                               # CLAUDE.md, app code, and local subagent never do

Demonstrates context isolation (nothing outside .managed-agents/ is read or uploaded) and shared-skill dedup across a roster.

deploy-workflow/ci-deploy.yml — git-push-to-deploy

A GitHub Actions template. Copy it to your repo's .github/workflows/, add an ANTHROPIC_API_KEY secret, and every push that touches .managed-agents/ validates, deploys, and commits the updated lockfile. See docs/deploying.md.

claude-code-skill/deploy-managed-agents/ — deploy from Claude Code

Drop this skill into your repo's .claude/skills/. Then in Claude Code you can say "deploy my managed agents" or "run the researcher with …" and it maps your words to the right agentlift command.