Delegate the coding,
not the understanding.

Cards is a local-first VS Code extension. It groups your plans, transcripts, and commits into project-scoped Git repositories on your disk. When you launch Claude Code or Codex, Cards feeds them this curated context, improving your agentic workflow instead of replacing it. It's built for Claude Code and Codex today, for one developer running one card at a time — not real-time multi-user collaboration.

Cards
Search cards…
Active1
extract the incident-timeline summary widget
main-138reviewui
gate session refresh on cookie TTL — drop the per-navigation refresh
main-142planauthperformance
Needs Review2
gate session refresh on cookie TTL — drop the per-navigation refresh
main-142unmergedauthperformance
rate-limit the Stripe webhook intake
main-141unmergedwebhooks
surface monitor-timestamp diff in the incident digest
main-136blockedingest
Todo1
gate session refresh on cookie TTL — drop the per-navigation refresh
main-142planauthperformance
Done1
gate session refresh on cookie TTL — drop the per-navigation refresh
main-142mergedauthperformance
migrate uptime rollup to ISO weeks
main-133mergedrollup
Claude Code
main-142
session.ts
Claude Code
Conductor — review
zsh
Claude Code·beacon · main worktree
create a card for the session refresh — gate on cookie TTL, drop the per-navigation refresh
Loaded cards:cards · read 3 references, ran 2 shell commands
Classifying as a bug-report card; loading the references.
Vibing…·12m 17s·↓ 22.9k tokens
»bypass permissions on(shift+tab to cycle)·esc to interrupt
01 / 05

Extend your decision bandwidth.

Once you're running parallel agents, the bottleneck becomes your decision bandwidth. Reading thousand-line transcripts to supervise them is a tax. Cards replaces transcript forensics with a structured timeline. The attribution tree shows the net effect of the agent's commits, and automated gates verify the changes before they reach your main branch.

02 / 05

The session ends; the card stays.

When a session ends, the context usually goes with it—the constraints you found, the edge cases, the reasons you chose one approach over another. Cards saves each agent job on disk as its own Git repository: the description, the plan, the raw transcripts, and the commits. Close the terminal and the card stays. Your next session picks up where the last one left off, instead of re-paying tokens to rebuild what you already knew.

03 / 05

Shape the work before the agent builds.

If the first thing you see is a finished diff, your review is reduced to checking whether the code compiles. Cards brings spec-driven development to your agentic workflow: you write the description and plan before the agent starts—an Interview helps you draw it out. That plan anchors the review, so you verify the output against what you actually intended, not just the code the agent happened to produce.

04 / 05

Keep the thread.

When an agent writes code, traditional git blame is broken—attributing changes to model IDs or default system emails rather than the engineer who directed them. Cards reconstructs the attribution tree. It maps every line of code back to the specific plan, prompt, and model version that produced it. The agent runs in its execution harness; you keep the thread.

05 / 05

Own the context; the agent does the work.

Cards doesn't replace your model or your agent—you stay in control of what the agent sees. You curate the card—the description, the plan, the past transcripts—and Cards hands that context to whatever agent you run, Claude Code or Codex. You own the context; the agent just does the work. The same card works with whatever agent you use.

↓ install
Cards

Cards is local-first. It creates no cloud databases and uploads none of your code — install the extension, link your Claude Code or Codex CLI, and you're running.