What Cards does not do
Cards is a work surface and context manager for building with agents. It orchestrates and controls agent runs—spawning your harness as a subprocess, injecting context and hooks, and owning the session lifecycle—but it is not a governance or policy control system, not an autonomous execution engine (the Factory/Jules/Devin style that owns the agent end to end), not the agent harness itself, and not a deployment platform. Establishing clear boundaries prevents integration conflicts and keeps you in control of your development environment.
Architectural Coordinator
Cards operates as a coordinator between you, your local Git repository, and your chosen agent harness. It prepares the context, manages the workspace isolation, spawns the harness as a subprocess, and logs the history, but it leaves the agent's tool loop and model inference to that harness and its provider.
Responsibility Matrix
| Feature | Handled By | What Cards Does NOT Do |
|---|---|---|
| Agent Execution | Agent Harness (e.g., Claude Code, Codex) | Cards does not run the agent loop, manage tool-calling logic, or handle agent planning. |
| Inference & Models | Model Provider (e.g., Anthropic, OpenAI, Ollama) | Cards does not host models, manage model weights, or provide LLM APIs. |
| Data Storage | Local Filesystem (~/.cards/) | Cards does not store your codebase or card content in the cloud. |
| Workflow Enforcement | Git / User Decisions | Cards does not enforce custom git workflows, branch naming rules, or code reviews outside the merge gate. |
| App Building | Compiler / Bundler / IDE | Cards is not a low-code app builder or code-generation engine. |
| Deployment | CI/CD (e.g., GitHub Actions, Vercel) | Cards does not deploy code, run production builds, or manage hosting. |
What Cards does NOT do
1. Host models or manage inference
Cards does not run LLMs. You must configure your own API keys or run a local model server (such as Ollama). The agent harness you select makes all inference calls directly.
2. Control the agent loop
Cards does not decide how the agent solves a task. It does not select which tools the agent calls or how it handles errors. The extension simply launches the agent CLI as a subprocess and streams its output.
3. Store card content in the cloud
We do not host your card descriptions, plans, or transcripts. If you delete the local ~/.cards/ directory without a backup, your card history is permanently lost.
4. Enforce gates beyond its two optional approval gates
Cards has exactly two gates, both optional and off by default:
- The plan gate (
planRequired/planApproved) blocks a card from moving fromtodotoactiveuntil the plan is approved. - The merge gate (
mergeRequestRequired/mergeApproved) blocks a card from reachingdone—and prevents the agent's worktree branch from merging into your active branch—until the merge is approved.
Both gates default to off, so cards progress without enforcement unless you opt in. Beyond these:
- Cards does not block you from editing files, committing code manually, or running commands outside the extension.
- Cards does not enforce branch protection rules, linting checks, or test suites on your main repository. You can configure your own Git hooks or CI/CD pipelines to handle these checks.
5. Live-pairing or collaboration
Cards is not a collaborative editing tool. It does not support real-time multi-user editing of the same card or worktree. Each card represents a local job run by a single local user and their agent subprocess.
Privacy and data
Review our local-first privacy architecture: how card data stays on your local disk, how licensing is managed, and how token traffic flows directly to your model provider.
Troubleshooting and support
Diagnose and resolve common issues: recovering crashed sessions, moving cards to Done, cleaning orphaned worktrees, and fixing registration problems.