What this repository needs
Toolsets, project policy, capability references and the lockfile. Commit these with the repository.
travels with the repo
Standalone · open source · zero-files by default
Your AI coding setup, on every machine.
Keep your MCP servers, skills and instructions in your own Git repo. AgentStack gives the right setup to Codex, Claude Code, OpenCode and other coding CLIs—without copying capability files into every project.
One setup, delivered where the work runs
Project manifest + lockToolsets, policy, exact reviewed pins
Personal capability libraryReusable skills and MCP server definitions, synced with your Git remote
↓
Laptop
Desktop
Remote
↓
Codex · Claude Code · OpenCode · moreDirectly, or launched by a supervisor such as stock T3 Code
Zero-files: MCP-capable tools receive servers and skills live. File-only features are rendered automatically into the format their CLI understands.
The whole model
Three layers. That is all you need to remember.
The portable parts travel. The sensitive and machine-specific parts stay on the machine that runs the agent.
What you reuse everywhere
Skills, MCP server definitions and personal instructions. Sync this folder through your own Git remote.
travels through your GitWhat must stay local
Secret values, trust decisions, installed CLIs, policy ceilings and audit history. Every machine supplies its own.
never synced by AgentStackYour central library
A normal Git repo. Every skill and MCP in one place.
Use GitHub, GitLab or any Git remote you already trust. AgentStack maintains the folders, index, checksums and provenance—no hosted account or special server required.
One simple ruleCommands with
--write apply a change. Remove it to preview first.01 · LINK ONCE
Choose your repo
Make this Git checkout the first place AgentStack reads from and writes new reusable capabilities to.
$agentstack lib link ~/GitHub/ai-setup --name central --first --write
02 · ADD
Keep what you reuse
Add a skill folder or an MCP server definition. AgentStack updates the library index for you.
$agentstack lib add ./my-skill --write$agentstack lib add-server my-tools --file ./my-tools.toml --write
03 · SYNC
Send it to your machines
Check the Git state, then let AgentStack commit, pull and push the portable library safely.
$agentstack lib sync --status$agentstack lib sync
What travels: skills, MCP definitions, instructions and extensions. What never travels: secret values, trust decisions and audit history. Server files use ${REF}; each machine supplies its own value.
Two short journeys
Start once. Recreate it anywhere.
AgentStack detects the coding CLIs on each machine and connects them to the same portable project contract.
First machine
Review and import existing MCP entries. Native skills and app-owned tools stay where they are unless you explicitly adopt them.
$cargo install --git https://github.com/Tarekkharsa/agentstack --locked agentstack$agentstack init$agentstack status
The wizard shows every write before it happens. Status ends with one next action.
Another machine
Point AgentStack at the same library repo. This machine then asks only for its own secrets and trust review.
$agentstack up --library https://github.com/you/ai-setup.git$agentstack up --library https://github.com/you/ai-setup.git --write$agentstack status
Preview, apply, done. Later refreshes are simply
agentstack up and agentstack up --write.Daily use
Most days, you only need status.
A trusted project with one default toolset opens it automatically when an agent connects.
Automatic
The trusted default toolset opens live for each agent connection.
Understandable
agentstack status shows what is locked, trusted, live and what to do next.
Portable
The repository plus your library recreate the same setup on another machine.
Recoverable
agentstack undo reverses recorded writes; uninstall removes managed empty folders too.
Governed workflows
Parallel agents, one reviewed plan.
Run several narrow reviewers at once, then let an independent verifier challenge their findings. Every child uses an approved toolset and gets its own recorded grant.
$agentstack workflow explain review-diff$agentstack workflow run review-diff$agentstack workflow report w-...
Works where you work
Standalone first. Supervisor-friendly too.
AgentStack does not require T3 Code, a desktop UI or a hosted service. It stays on the machine where the coding CLI runs.
✓Direct CLI use: Codex, Claude Code, OpenCode and the other supported adapters.
✓Stock T3 Code: local, T3 Connect and SSH environments use that environment's AgentStack setup.
✓Other supervisors: ordinary provider launches work without an AgentStack-specific UI.
✓File-only tools: capabilities that cannot travel live use the compatible rendered lane.
Make this machine the first one.
Install AgentStack, run the guided import, and let status tell you the only remaining action.