Open-source UI for the screens after chat.
Plans, tool calls, permission gates, memory, traces, cost, failures, and human review. Free and open source, built on React and Tailwind.
Get notified when it ships. Free and open source.
Active run
Build the approval surface
Most AI UI kits stop at chat.
Message bubble. Input field. Typing indicator. Real agent products are not chatbots. They plan. They call tools. They ask for permission. They fail in specific ways. They cost money per run.
A component for every part of a run.
Eight categories. Built around what agents actually do, not what chatbots look like.
schema.prisma · 1.2s
Tool calls
Receipts for every action. Tool name, inputs, outputs, status, duration.
Thought for 2s
Reasoning and traces
Step-by-step execution log. Chronological, timestamped. For debugging and trust.
Plans and tasks
What the agent is about to do. Status pips. Edit before run.
12 records · cannot be undone
Permissions and trust
Permission requests, scope maps, human approval, audit rows.
Memory and context
Stored facts, source, confidence, forget-all. Users can see what the agent remembers.
$0.04
actual · 3 tool calls
$0.04 / $0.12 budget
Monitoring and cost
Estimate before, ticker during, summary after. Budget alerts. Token counts.
- id String @id
+ id String @id @default(uuid())
email String @unique
+ createdAt DateTime
Sandbox and code
File trees, editors, diff panels, terminal output.
writing rows…
Loaders
Progress states for long-running actions. Not generic spinners.
Every run leaves a receipt.
Checks, costs, and outcomes are logged beside the work. When a run finishes the evidence is already there: lint, types, tests, policy. Not buried in a log.
See the verification receipt
Verification passed
run/agent-ui-proof · 3m 18s
verification.finished → Verification receipt
Gate what the agent can touch.
Consequential actions pause for a human. Approve, deny, or edit the scope. Nothing lands quietly. The permission gate is a real component, not a pattern you build from scratch.
See the permission gate
Apply remote diff
The agent wants to write two files. Review before it touches the workspace.
permission.requested → Permission gate
Review every change before it lands.
Remote edits arrive as proposals. Diffs render inline with line-level markers before anything is written. Review the patch, request changes, or accept. The agent waits.
See the diff review
The code is free.
Free and open source. MIT licensed. No dependency, no subscription. Yours to read, copy, and ship.
The Figma template is paid.
The polished design system for designers: variables, light and dark, six workflows, component naming, file structure, and documentation. One purchase, yours to keep.
Get notified when it ships.
The Figma template, the registry, and more components. One email.
No spam. One email when it ships.
Build agents people can trust.
Good agents ask before they overstep, show their work, and keep the human in charge. Kanshō is the starting point. The code is free.