Capabilities

Generated chains can reach beyond the LLM via tools and AgentSkills. MAESTRO surfaces everything installed through the Catalog (Ctrl+K, or care catalog).

Tools

Tools are Python functions a chain’s tool steps can call.

Registry — register functions with @carl_tool; the tools config controls the registry (CARE_TOOLS__*).
Bundled builtins — MAESTRO ships ready-made tools (e.g. web_search).
On-the-fly synthesis — when a chain needs a tool that isn’t registered, MAESTRO can synthesise one from a description, plan its inputs, and wire it in.

See CARL tool steps for how chains invoke tools.

AgentSkills

AgentSkills are portable skill folders (SKILL.md + scripts/assets). MAESTRO’s catalog discovers installed skills (from ~/.agents/skills/, ./.claude/skills/, and the agent-skills library) and chains run them via AgentSkill steps. The full model — URIs, execution modes, resolvers — is documented in CARL → AgentSkills.

Sandbox

Skill scripts run inside a sandbox so untrusted code stays contained. Backends: local, docker, e2b, firejail (configure via CARE_SANDBOX__*). Trusted skills are tracked in a SHA-pinned trust store you can audit and revoke from the Sandbox Trust screen (/sandbox).

Context injection

MAESTRO can inject extra context into generation (CARE_CONTEXT__*):

CARE.md — a project context file, picked up like a system brief.
Long-term-memory digest — a summary of saved memory folded into the prompt.

Telemetry

Opt into an event-stream sink (e.g. Langfuse) via CARE_TELEMETRY__* to trace generation + execution in a dashboard.