Input→Tool→Review→Output
Your idea edits the goal. The selected template determines the executable workflow; change nodes to change behavior.
A WORKBENCH FOR AGENTIC THINKINGBuild your own
Build your own
AI agent.
No coding required.
Turn your ideas into intelligent AI agents.
Design workflows. Connect local tools. Simulate decisions.
Build a plan you can inspect, test and reuse.
Core mode: deterministic simulation. No model calls, browsing or external actions.
Ten practical starting points.
Your projects stay in this browser. No signup required.
Define a useful job.
Permitted local operations
External writes are unavailable and always blocked. Browser approval is not authorization for a future server connector.
Local tool catalog
100%
Drag nodes to move. Drag empty canvas to pan. Use arrow keys on a focused node to move it. Click an output handle, then a target node to connect; select the port below.
Connect a path
Make the agent’s contract explicit.
Generate from the actual goal, tools and constraints. Refinements are predefined rules, not AI reasoning.
Before
This document is reusable guidance. Engine permissions and transitions are enforced separately. Editing it does not silently rewrite the graph.
Run it. Pause it. Understand it.
Failure injection (local experiment)
Ready for a simulation.
Human approval required
Review the mapped input and current state. No external action will occur.
Final output
No output yet.
Execution timeline
Current variables
{}The current view shows your supplied data. Exported logs redact common secret field names and token patterns, not every possible secret. Never paste credentials.
Compare two runs
Latest two manually retained runs, kept in this tab only. Compare routing and outputs; simulated differences are not evidence of real AI quality.
Assertions, not confidence scores.
Test expected status, an output substring and an optional visited node. Scenario approval choices apply only to these local test runs.
Take your agent with you.
Portable configuration
Complete local definition, sample, tests, instructions and snapshots. Import into another copy of this studio.
Review documents
Keep your tests and redacted run evidence alongside the workflow.
Runnable JavaScript package
The same deterministic engine, tools, actual graph, CLI approval hook and README. No model or external connector is installed.
Live integrations are a separate step.
No provider credentials are collected. To add real AI or external tools, implement a server-side adapter with secret management, permission checks, user consent, input/output validation and sanitized logs. Approval in this browser cannot authorize a server action. Python generation and managed deployment are not included in this release.
Official agent-development documentation ↗
Saved configuration snapshots
Ten concepts. One inspectable workflow.
Your local agents.
Projects are saved in IndexedDB in this browser and can be cleared by browser cleanup. Export backups; there is no cloud sync. Imports create a copy.
Saved agents
Understand the workflow before you trust the output.
Agents versus chatbots
A chatbot primarily exchanges messages. An agent design adds goals, tools, state and rules for deciding what happens next. Those rules should be visible, testable and proportionate to the consequences.
Start with one useful path
Define inputs, a transformation and a measurable output. Connect them, test missing data, then introduce branches. More steps or agents do not automatically produce better results.
Keep a human in the loop
Drafting is different from sending. Approval nodes make review explicit. Real external actions still need server-side authorization, trustworthy tools and clear limits.
Is this running real AI?
No. Local tools perform real deterministic transformations. AI Task nodes render clearly labeled example text. No model, web search or external account is connected.
Can input text change permissions?
No. Input remains data. It cannot rewrite graph permissions or cause code execution. External connectors are unavailable and blocked even after approval.
Does export deploy my agent?
No. JSON and SVG are portable design artifacts. The JavaScript package runs this deterministic simulation locally; live AI, multi-agent isolation and hosting require additional implementation.