AgentFlow

中文文档

AgentFlow is an out-of-the-box AI agent workflow system for running multi-agent workspaces, streaming conversations, project-aware tools, task tracking, and model/MCP management from a web console or mobile client.

• Login with administrator credentials and captcha verification.
• The default administrator can be created from the configured admin_user and admin_password; first login can require password change.
• Manage personal profile information and password from the account menu.
• Use role-based permissions to control access to the dashboard, workbench, models, agents, skills, MCP tools, projects, members, roles, prompts, system tools, and system settings.

• Add and manage models using these protocols:
  • OpenAI-compatible chat completions
  • OpenAI Responses
  • Anthropic Claude
  • Google Gemini
• Configure display name, local model ID, upstream model ID, base URL, API key, token limits, and reasoning effort.
• Mark model capabilities with tools, vision, and image_gen.
• Enable models for automatic fallback/switching.
• Test model connectivity from the model list.
• Fetch upstream model lists when the upstream provider supports listing models.
• Import and export model configurations as JSON.

• Select, create, and delete workspaces from the workbench.
• Create, open, delete, and stop conversations.
• Stream assistant responses with SSE.
• Reconnect to a running conversation and replay available streamed events.
• Switch the active agent for a conversation.
• View assistant text, reasoning content, tool calls, and tool results.
• Use main conversations and sub-conversations created by agent delegation.
• Enhance a drafted prompt before sending it.
• Stop a running task without deleting the conversation history.

• Upload supported attachments including images, text files, Markdown files, and PDFs.
• Preview uploaded image attachments and remove pending attachments before sending.
• Browse uploaded files and browser screenshots for the current workspace.
• Read and save editable uploaded text content.
• Browse project files for workspaces linked to a project path.
• Read existing project files and save edits to existing files.
• Read and save workspace documents under specs/ and tasks/.
• Read and save project documents such as AGENTS.md, README.md, and Markdown files under docs/.

• Create projects with a name, description, and bound project path.
• Create workspaces under projects or directly from the workbench.
• Bind a default agent to a workspace.
• Keep workspace data isolated by workspace UUID.
• View workspace task lists and task progress.
• Delete a workspace and its workspace files when no conversation is running.

• Use built-in agents loaded from the configured context directory.
• Create and edit custom agents from the UI.
• Configure each agent with:
  • name, title, description, keywords, icon, and status
  • model selection or automatic model mode
  • role, rule, workflow, skill, and memory files
  • auto-loaded context file types
  • allowed tools
  • linked skills
• Manage reusable skills with label, description, keywords, level, content, status, and sort order.
• Agent memory is stored per agent and can be updated by agent tools or context tidy-up.

AgentFlow exposes built-in tools by capability category instead of a fixed tool count:

• Import MCP servers from standard mcpServers JSON.
• Export selected MCP server configurations as mcpServers JSON.
• Configure command, arguments, environment variables, timeout, label, category, and version.
• Start and stop MCP servers from the UI.
• View server status, discovered tool count, and tool names.
• Enabled MCP servers are started on application startup and their tools are registered automatically.

• View built-in system tools synchronized from the tool registry.
• Enable or disable built-in tools from the system tools page.
• View and edit system prompt templates.
• Reset prompt templates to their default content when available.

• Configure site settings and security-related settings from the system configuration page.
• Important configuration keys include:
  • session_max_age for browser session lifetime
  • admin_user and admin_password for the initial administrator account
  • llm.timeout, llm.max_retries, and llm.stream_idle_timeout for model calls
  • ai.vision_model, ai.image_gen_model, and ai.prompt_enhance_model for AI service routing
  • search.engine, search.searxng_url, and search.jina_api_key for web search
  • pagination defaults for list pages

• Login through the API with a bearer token.
• Store a separate token per server.
• Manage multiple AgentFlow servers and switch between them.
• Browse workspaces, conversations, and tasks.
• Use SSE streaming conversations with reconnection support.
• Display text, reasoning content, tool calls, tool results, attachments, and code blocks.