using-agentstacklisted

# Using agentstack Use this skill when a task involves an agent CLI's setup: adding/removing MCP servers or skills, switching capability profiles, missing servers/skills in a project, secrets that don't resolve, a proxied tool call being refused, or auditing what an agent can access. ## The mental model (one paragraph) agentstack is a **compiler and a runtime**. Compiler: intent lives in a commit-safe manifest (`.agentstack/agentstack.toml`, or legacy root `agentstack.toml`) and is rendered into each CLI's native config. Runtime: the same manifest can be served live through a **gateway** (`agentstack mcp`) that proxies the project's MCP servers to any harness — trust-gated per repo, firewalled by two policy layers, every call logged. Skills and server definitions can live in a machine-wide **central library** (`~/.agentstack/lib/`) and be referenced **by name** (pinned by digest in `agentstack.lock`; the gateway refuses a library definition that drifted from its pin). Secrets are `${REF}` placeholders resolved per machine (env → varlock → OS keychain → `.env`); an unresolved secret **blocks** the write — and blocks the call, at the gateway. Nothing touches disk without `--write`. ## The adoption ladder (meet the user where they are) agentstack is adopted in six steps; most projects sit partway up. Before proposing anything, detect the current step — bare `agentstack` reports the directory's state and next step, `agentstack doctor` names what's unwired, `agentstack guard