These pages describe v0.19.0, the current release. That is the build the installer gives you. agentstack --version says which build you have.

Documentation

AgentStack keeps your reusable capabilities — MCP servers, skills, and instructions — in one central library, lets each project select a small toolset by name, pins exactly what it selected, and serves that to every supported agent CLI without copying configuration into every repo.

your library repo  →  each project's manifest + lock  →  every agent CLI

Everything published here is linked from this page: the guides, the how-tos, the reference, and the enforcement record.

Start here#

These pages take you from an empty machine to a working, reviewed setup. You do not need the full reference to begin.

• Get started — install, set up the first machine, add project files, lock and trust them, and repeat on a second machine.
• Central library — the folder layout, several linked libraries, precedence, shared instructions, and syncing.
• Tutorial — the step-by-step map of the walkthrough, and where each step is taught.
• Concepts — the three layers, delivery modes, dynamic skills, tool discovery, trust, and policy.
• Which protection do I need? — how capabilities are routed for you, and how much enforcement to ask for.

Most-used commands#

Every command, flag, and exit code is in the feature reference.

How-to guides#

One task per page, start to finish.

• Add an MCP server — store one reusable ${REF}-safe definition and select it by name.
• Add a skill — write a description that retrieves well, validate it, and make it loadable.
• Name a toolset — bundle the smallest set of servers and skills a task needs, and pick the default.
• Run a multi-agent workflow — fan work out to narrow roles, verify the result, and read one evidence tree.
• Trust a project — why a fresh clone is inert, and what the consent ceremony shows you before you accept it.
• Lock down a run — confine an agent run to the files and destinations you approved.
• See what your agents did — inspect what was run, called, and spent, from recorded evidence.
• Undo AgentStack changes — the timeline, and how to reverse managed writes byte for byte.
• Share one setup with a team — what travels in Git, what stays local, and how each teammate starts.
• Use it in CI — gate a repository's agent setup in continuous integration.

When something is wrong#

• Troubleshooting — search for the text your terminal printed, and run the repair command it names.
• FAQ — the questions people ask before and after adopting it.

Common ones:

• What does the lock mean?
• How does the agent see my skills if they are not copied into the repo?
• Can I link my GitHub library and keep my old local library?
• How do CLI- and model-specific instructions work?

Bring the setup you already have#

• Migration recipes — import existing MCP entries and native skills, each recipe starting from a read-only plan.

Reference#

• Feature reference — every command, flag, and behaviour that is implemented and tested.
• Adapter support matrix — which CLIs are supported, and how far each one is verified.
• Integrations — how AgentStack sits under t3code and other callers, and what stays standalone.
• Automation contract — the stable JSON surface, so a program can read everything a person can.

Deeper#

• How it works — crates, boundaries, and the runtime path a call takes.
• Governed workflows — the workflow engine, its evidence, and its limits.
• Enforcement matrix — every enforcement claim, what backs it, and what it deliberately does not cover.
• Demos — scripts that check the claims on this site and fail loudly.

Source of truth: docs/docs.md — this page is generated from it.