Skip to content

Home

Jump to bottom
karurikwao edited this page Jul 5, 2026 · 2 revisions

AgentDesk Wiki

AgentDesk is a local visual debugger for AI agent runs. It is built for the moment after an agent workflow fails and you need to answer concrete questions: which node ran, what prompt or tool call was used, what came back, what it cost, what artifacts were produced, and what evidence can be shared safely.

The README is the launch overview. This wiki is the long-form operating manual for users, contributors, and launch reviewers.

Fast Start

Run the packaged app locally:

npx --yes --package=@papaplus/agentdesk agentdesk --port 5173

Open http://127.0.0.1:5173 and start with the Failure Replay Lab.

The shortest demo path is:

  1. Click Run demo trace.
  2. Click the failed Browser Replay trace event.
  3. Inspect Debug, Artifacts, Costs, and Validation.
  4. Click Replay failed step.
  5. Export the replay session or trace bundle ZIP.

Documentation Map

Page Purpose
Getting Started Install paths, first run, and the 30-second failure replay demo.
Core Concepts Workflows, nodes, edges, traces, artifacts, costs, validation, and replay.
Run Modes Demo, Ollama, Cloud BYOK, and Runtime mode behavior.
MCP and Runtime MCP import, discovery, local commands, official SDK adapter, and safety boundaries.
LLM Configuration OpenAI and Anthropic BYOK setup, model presets, and browser-direct caveats.
Replay Sessions and Exports Replay-session JSON, trace summaries, portable workflows, ZIP bundles, and framework starters.
Troubleshooting Common Node, npm, browser, Ollama, BYOK, MCP, and Runtime mode issues.
Security Model Redaction rules, local execution boundaries, secret handling, and non-goals.
Contributing Guide Repo workflow, verification commands, code areas, and good first contribution paths.
Launch and Promotion Public positioning, launch assets, community threads, and promotion checklist.

What AgentDesk Is

AgentDesk is a debugger-first agent tool. It gives you:

  • a visual workflow canvas
  • click-linked trace events and graph nodes
  • prompt, tool, result, stdout, stderr, artifact, and cost inspection
  • graph validation for common workflow mistakes
  • failed-step replay without deleting the original failed event
  • redacted replay-session exports
  • trace bundle ZIP exports
  • local Ollama execution for Ollama model nodes
  • BYOK browser-session execution for configured OpenAI and Anthropic model nodes
  • loopback Runtime mode for approved local commands and MCP discovery/tool calls

What AgentDesk Is Not

AgentDesk is not a production scheduler, hosted workflow runner, secret vault, queue system, or full observability platform. It can complement those systems by producing local, shareable debugging evidence before you wire a workflow into production.

Use AgentDesk when you need to explain or reproduce an agent run locally. Use a workflow builder when you need production scheduling, hosted secrets, queues, branching operations, or live third-party tool execution at scale.

Current Safety Contract

AgentDesk is intentionally explicit about execution:

  • Demo mode does not call models or tools.
  • Ollama mode only runs Ollama model nodes against local Ollama.
  • Cloud BYOK mode only runs configured OpenAI/Anthropic model nodes and stores API keys in browser session state only.
  • Runtime mode requires the packaged loopback CLI.
  • Local command execution is isolated behind explicit local runtime requests.
  • MCP discovery and tool calls are user-triggered, not automatic remote probing.
  • Exports redact common secrets and private local path prefixes.

Do not paste API keys, customer data, private traces, tokens, cookies, or unredacted screenshots into prompts, labels, stdout/stderr, artifacts, issues, or discussions.

Public Links

AgentDesk Wiki

Project

Clone this wiki locally