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

Integrations

AgentStack is standalone. Its manifest, lock, trust, policy, dynamic delivery, and machine bootstrap work without a supervisor or custom UI.

An integration may launch a supported CLI, but it does not become a second configuration system.

T3 Code#

Stock T3 Code can launch and supervise Claude Code, Codex, Cursor, OpenCode, and other providers. Those processes read the same global gateway and managed provider configuration AgentStack already owns.

The practical setup is therefore only:

bash

$ agentstack init --connect
$ agentstack status     # inspect this worktree

When doctor detects T3 Code it checks provider guard coverage and reports provider-instance home overrides that could point a session away from the global configuration.

Several projects and machines#

T3 Code Connect can control work on remote machines. AgentStack handles the configuration on each machine independently:

T3 Code UI
  ├── machine A → project A → local AgentStack gateway
  └── machine B → project B → local AgentStack gateway
                          ↑
                same central library Git repo

Bootstrap each machine with the same library:

bash

$ agentstack up --library https://github.com/you/ai-setup.git
$ agentstack up --library https://github.com/you/ai-setup.git --write

Project manifests and locks travel with their repositories. Library content travels through its Git remote. Secrets, trust, policy, and audit evidence stay on the machine where the agent runs.

No T3 Code fork or AgentStack panel is required. The earlier experimental custom panel is not the supported setup path; its lasting lesson is that a UI should present AgentStack's JSON status and preview/consent contracts, and the CLI stays the authority.

For integrations, agentstack init --plan advertises the init-plan contract: it emits JSON with the detected setup and a plan_digest without changing the machine. There is no separate --json flag; --plan is already the JSON read. Pass that digest to the approved write step so a UI cannot apply a different plan from the one the human reviewed.

JSON contract discovery#

Read features in every JSON envelope before using an integration contract. This release advertises: init-plan, apply-setup, init-tool-managed-v1, trust-preview, trust-consent, status-v1, profiles-v1, diff-v1, restore-last, sessions-v1, profiles-edit-v1, diff-ownership-v1, toolset-create-v2, profiles-edit-batch-v1, toolset-rename-v1, toolset-delete-v1, library-remove-v1, manifest-remove-v1, trust-server-blockers-v1, trust-review-card-v1, trust-card-diff-v1, trust-card-groups-v1, activity-skill-load-v1, workflow-observe-v1, workflow-serial-roles-v1, doctor-advisories-v1, doctor-mode-v1, doctor-liveness-v1, doctor-probe-v1, diff-existence-v1, json-reads-v1, gitignore-opt-out-v1, doctor-cli-coverage-v1, status-honesty-v1, needs-your-yes-v1, update-offer-v1, package-members-v1, lease-status-v1, delivery-routing-v1, library-sources-v1, trust-gate-reading-v1, instruction-channels-v1, image-plan-v1, workflow-role-selection-v1, trust-content-drift-v1, and abandoned-render-v1.

Fresh worktrees#

A fresh worktree has the committed project manifest and lock. With zero-files delivery it does not need generated MCP configs or copied skill directories. It still needs local trust because consent is tied to the checkout path.

Run in the worktree:

bash

$ agentstack status

Then review the checkout:

bash

$ agentstack trust .

The next agent connection receives the default toolset automatically. A file- only provider may also need the rendered command printed by status.

Per-run evidence#

T3 Code sessions normally attribute to the machine's global audit. For a separate run identity, create a transparent AgentStack shim and configure that provider instance to launch it:

bash

$ agentstack more shim make claude

This is optional and does not change the provider's capabilities.

Other supervisors#

The same rule applies to any launcher: if it starts a supported CLI with its normal home and working directory, AgentStack works unchanged. If the launcher overrides either, run agentstack doctor on that machine and follow the reported adapter-specific fix.

Next: Another machine · Team setup · Adapter support

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