Trae
Type: full-code · Vendor: ByteDance · Language: TypeScript (Electron IDE) + Python (Trae Agent CLI) · License: proprietary (IDE) / open-source (Trae Agent CLI) · Status: active · Status in practice: emerging
ByteDance's AI IDE family: the Trae IDE with a unified Chat-Builder interface and @Agent + MCP multi-agent system, the autonomous TRAE SOLO coding agent, and the open-source Trae Agent CLI that topped SWE-bench Verified.
Description. Trae is ByteDance's AI coding stack. The IDE provides a unified Chat-Builder interface and an @Agent system with MCP for multi-agent collaboration. TRAE SOLO is positioned as a more autonomous responsive coding agent for real-world software development, with a mobile variant added in 2026. Separately, the open-source Trae Agent CLI (github.com/bytedance/trae-agent) ranked #1 on SWE-bench Verified with Claude 4 and provides a transparent, modular architecture researchers can extend; it supports MCP servers and a toolkit of file editing, bash execution and 'sequential thinking'.
Agent loop shape. Three surfaces. (1) Trae IDE Chat-Builder: user task in chat is routed to @Agent which composes tool calls and (optionally) sub-agents over MCP. (2) TRAE SOLO: more autonomous loop that executes longer, higher-trust workflows responsively. (3) Trae Agent CLI: a Python ReAct-style CLI taking a natural-language task, running tools (file editing, bash execution, sequential thinking) and optional MCP servers, until the task is done.
Primary use cases
- AI-assisted coding inside the Trae IDE
- autonomous coding sessions via TRAE SOLO
- research and benchmarking via the open-source Trae Agent CLI
- MCP-driven multi-agent code workflows
Key concepts
- Chat-Builder unified interface (docs) — The Trae IDE's single surface combining conversational chat with a builder mode that produces concrete changes.
- @Agent system → agent-as-tool-embedding (docs) — Mention-based agent invocation that lets the user (or another agent) call a specific agent and pass context.
- MCP multi-agent collaboration → mcp (docs) — MCP is the protocol Trae uses to compose multiple agents and external tool servers.
- TRAE SOLO (docs) — Autonomous coding agent variant positioned for real-world software work; mobile variant added 2026.
- Trae Agent CLI (docs) — Open-source Python CLI for general software-engineering tasks; tops SWE-bench Verified.
- Sequential thinking tool → extended-thinking (docs) — Built-in 'thinking' tool inside the Trae Agent toolkit alongside file editing and bash execution.
Patterns this full-code implements —
- ★Agent-Computer Interface
Trae IDE is an Electron-based agent-aware IDE; the CLI variant provides a transparent agent-computer interface for software-engineering tasks.
- ★★Model Context Protocol
MCP is documented both as the IDE's protocol for multi-agent collaboration and as a configuration-level integration point in the open-source CLI.
- ★★ReAct
Trae Agent CLI executes complex software-engineering workflows via an explicit tool-call loop with reasoning and structured thinking; same shape underlies the IDE's @Agent.
- ★★Tool Use
Documented tool palette includes file editing, bash execution, and sequential thinking, plus any MCP servers added to the configuration. Configuration example lists exact tool names.
- ★Subagent Isolation
Trae IDE markets @Agent + MCP as multi-agent collaboration, with a dedicated @Agent blog ('Rules + Tools to Define Your Future Agents'). Specific isolation guarantees per sub-agent are not spelled ou…
Neighbourhood
Click any neighbour to follow the lineage. Scroll to zoom, drag to pan.