Clickable evidence bundles for failed AI agent runs

Your agent failed. AgentDesk shows the node, prompt, tool call, artifact, and replay.

Debug the run before you ship the workflow: click from trace to graph, replay the failed step, inspect stdout/stderr, JSON, markdown, screenshot artifacts, and cost, then export a redacted replay session for an issue, PR, or handoff.

BYOK cloud calls are browser-direct and can be blocked by provider CORS or organization policy. Prompts and responses become trace evidence; API keys are session-only and not exported.

Open AgentDesk Inspect Export
AgentDesk failure replay demo loop
What sets it apart

A debugger, not another workflow builder.

Failure-first replay Click a failed trace event, highlight the graph node, replay that step, and keep the original evidence intact.
Inspectable evidence Prompts, model/tool calls, stdout, stderr, JSON, markdown, simulated screenshots, costs, and validation stay visible.
Local and honest Ollama runs locally, cloud BYOK is session-only, and local/MCP execution requires the loopback CLI Runtime mode.
Screenshots

Launch-ready product tour.

AgentDesk failed-step debugger
Failure Replay Lab with node-linked prompt, tool call, result, and stderr.
AgentDesk artifacts viewer
Artifact viewer for replay context, screenshots, JSON, markdown, stdout, and stderr.
AgentDesk BYOK LLM configuration panel
Session-only BYOK LLM config with provider/model presets and CORS/security caveats.
AgentDesk workflow run screenshot
Graph canvas, trace events, costs, validation, Doctor, LLMs, and MCP import in one local app.
Try the loop

Bring a failed run. Leave with replayable evidence.

Run the demo, inspect the failed step, export a redacted replay session, then wire in local Runtime mode when you are ready.

Read Launch Details