Trae Agent
https://github.com/bytedance/trae-agent- Category
- AI Agents
- Rank
- No. 325Tools index
- Pricing
- Open Source
- Type
- AGENT
- Builder
- bytedance
- GitHub
- 12.1k stars
- Date
About
ByteDance's open-source LLM-based agent for general software engineering tasks — debugging, refactoring, feature work — across multiple model providers.
What it does
Trae Agent turns a natural-language task into an iterative command-line work session. It can inspect and edit files, run shell commands, modify JSON, and signal completion after verification. Interactive mode supports follow-up work, while trajectory recording captures model exchanges, tool activity, errors, and execution results.
Why it's ranked here
Its strongest case is inspectability. Provider selection, configurable tools, step limits, and detailed trajectories make experiments easier to reproduce and compare. Docker execution and benchmark harnesses add practical depth. However, programmatic headless access, dedicated secure sandboxing, and multi-agent coordination remain roadmap items, so the current product is chiefly a configurable command-line agent.
What's good
Provider coverage includes OpenAI, Anthropic, Google Gemini, Azure, Doubao, OpenRouter, Ollama, and custom base URLs. Built-in tools handle exact file edits, persistent shell sessions, structured reasoning, JSONPath edits, and verified completion. Trajectories preserve token usage, tool calls, results, reflections, errors, and timing. Configuration precedence is explicit, and Docker mode can create, load, build, or attach to environments.
Tradeoffs
Installation requires Python 3.12 or newer and UV. Hosted models require provider credentials, while configuration adds several layers to understand. Shell commands receive a 120-second timeout, and their output may be clipped. Recorded trajectories can contain proprietary code or other sensitive material. A headless SDK and dedicated secure sandbox are planned rather than established capabilities.
How to use it well
Use it for supervised repository tasks, provider comparisons, agent research, and debugging runs where a complete action trace matters. Start with a constrained working directory, explicit step limit, selected tools, and Docker when isolation or reproducibility matters. Review generated patches and secure trajectory files. It does not yet replace an embeddable application API, a dedicated secure execution service, or a multi-agent orchestrator.
Technical notes+
pyproject.toml defines a Python >=3.12 Hatchling package, exposes trae-cli = "trae_agent.cli:main", and includes provider SDKs, Textual, MCP, tree-sitter, PyInstaller, and YAML support. trae_agent/tools/base.py normalizes tool names, builds provider-aware schemas, and supports sequential or parallel execution. trae_agent/tools/run.py launches asynchronous shell commands with a 120-second default timeout and 16,000-character output clipping. trae_agent/agent/agent.py initializes trajectory recording automatically, attaches configured tools, starts MCP clients when allowed, and cleans them up after execution. evaluation/utils.py and evaluation/run_evaluation.py provide Docker-backed harnesses for SWE-bench, SWE-bench-Live, and Multi-SWE-bench.
Observed
- License
- MIT
- Primary language
- Python, requiring version 3.12 or newer
- Packaging
- Hatchling wheel package for trae_agent, installed with UV
- Interface
- Command-line interface with run, interactive, and configuration inspection modes
- Model providers
- OpenAI, Anthropic, Google Gemini, Azure, Doubao, OpenRouter, and Ollama
- Extensibility
- Optional Model Context Protocol server configuration
- Execution environments
- Local working directories and Docker image, Dockerfile, archive, or existing-container modes
Read from README.md, Makefile, pyproject.toml, docs/tools.md, docs/roadmap.md, docs/legacy_config.md, docs/TRAJECTORY_RECORDING.md, trae_agent/cli.py, evaluation/utils.py, evaluation/__init__.py, trae_agent/__init__.py, evaluation/run_evaluation.py, trae_agent/tools/run.py, trae_agent/tools/base.py, trae_agent/agent/agent.py.
What it can do
Debug software code
Code files with errors or issues → Identified bugs and suggested fixes
Refactor existing code
Legacy or poorly structured code → Improved, restructured code
Implement new software features
Feature requirements and specifications → Working code implementation
Generate code from natural language descriptions
Natural language task descriptions → Functional source code
Analyze code quality and suggest improvements
Source code files → Code quality assessment and recommendations
Execute software engineering tasks across multiple LLM providers
Engineering tasks and model provider selection → Task completion using specified LLM backend
Tags
Tech Stack
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.