Skip to content
←Back to Applications

APPLICATION

AI CodingDevelopmentTop

Trae: ByteDance two-form-factor coding line - TraeCode (IDE + SOLO) and TraeWork, plus a stalled MIT-open shell

trae

Trae is the AI coding line from ByteDance, and the official docs split it into two clearly different things. TraeCode is a development tool with AI deeply integrated, in two modes: IDE mode keeps the editor, terminal, debugger, extensions and source control for scenarios that need fine-grained control over code changes and execution, while SOLO mode hands the lead to AI - describe the requirement in natural language, by voice or by uploading local files, and it decomposes the task itself and runs code generation, testing, preview, change summary and deployment, with a task panel, an AI conversation panel and a tool panel (built-in editor, doc viewer, browser) from left to right. TraeWork is the AI-native workspace grown out of the SOLO mode of TraeCode, on web, desktop and mobile, in Work / Code / Design modes, aimed at product managers, data analysts, operations and designers rather than developers only. The long-task ceilings are stated concretely: Max mode extends the context window to 1M, allows up to 200 tool-call rounds per task and reads up to 750 lines per file read to cut down on chunking. bytedance/trae-agent is MIT-licensed but has not been updated for about eight months and cannot stand for the current engineering of Trae. Boundaries: no third-party benchmark readings are published (no SWE-bench or Terminal-Bench self-evidence), so every capability claim comes from feature docs and vendor descriptions; clients are closed source; 1M / 200 rounds / 750 lines are ceilings rather than typical experience and the vendor itself says they raise cost substantially; model availability is doubly limited by region (Seed, MiniMax and GLM unavailable to US users) and by plan tier, so teams across regions do not share one model set; concurrent cloud tasks are a hard tier difference and Free excludes SOLO, so evaluation cannot look at monthly price alone. Confidence C (vendor-claim).

C
CONFIDENCE
Vendor Claim
Official model card or keynote only, no independent re-test
1M · 200 轮
KEY METRIC
Max 模式上限(上下文 / 单任务工具轮次)
Vendor Claim · 2026-09
MATURITY
Product
research → demo → product → production
Our take

We grade it C (vendor-claim). What we could verify independently is popularity and engineering fact only: bytedance/trae-agent at 12,119 stars / 1,352 forks / MIT / last push 2026-02-05 / 191 open issues, pulled by us straight from the GitHub API on 2026-09-28, with the companion paper arXiv:2507.23370 checkable; and the 18-model built-in list, the 272k-tiered unit prices, and the five plans with concurrent cloud task counts (2/2/10/15/20), all published in official docs and the pricing page. But the capability claims that actually matter in this dossier - Max mode's 1M context, 200 tool rounds per task, 750 lines per read, SOLO's end-to-end delivery chain, automatic cloud failover when a device goes offline - come entirely from official product documentation, and Trae publishes no third-party benchmark readings at all (no SWE-bench or Terminal-Bench style self-evidence). We have not re-tested any of it. Per contract section 13.5, "official docs only, no independent re-test" is vendor-claim; a star count is a popularity fact, not a capability reading, and cannot lift the grade to A.

Water level: Trae's strategic position is a model-neutral multi-frontend dispatch layer, not an in-house-model play. Three things support that. First, its model surface is among the most heterogeneous in the market - OpenAI (GPT-6-Astra/Sol/Luna, three GPT-5.6 variants, GPT-5.5/5.4/5.2), Google (Gemini-3.1-Pro / 3-Flash), ByteDance's own Seed-2.1-Turbo, plus China's GLM-5.2, DeepSeek-V4-Flash, Kimi-K3/K2.7-Code/K2.5 and MiniMax-M3/M2.7, with BYOK on top. Trae sells orchestration and experience, not the base model. Second, it makes cross-device dispatch the product's main axis: the phone is the dispatch center, managing cloud TraeWork and several personal computers at once, assigning concurrent tasks to different devices, failing over to cloud execution automatically when a device is offline, with one account system and real-time task sync across three clients. That abstracts "which machine the agent runs on" into a schedulable resource pool, which is further along than a single-machine IDE. Third, its cost language is unusually candid - the Max mode docs state directly that enabling it significantly increases cost and that the default window is usually sufficient for everyday work. A vendor talking users out of its own high tier inside feature docs is rare in this market.

Three things worth copying: the agent workflow is explicitly decomposed into requirement analysis, code research, solution design, implementation of changes, delivery and acceptance, and the built-in Agent produces an actionable plan first and only develops step by step after you confirm it - the cheapest available gate against long-task runaway. DiffView collapses agent output into one reviewable change (affected file count, total lines changed, per-file diffs). And AI code review over uncommitted changes, a single commit or a branch diff emits summaries, flowcharts and diffs, turning review into a comprehension tool rather than a compliance step.

Risk signals stated plainly: trae-agent has had no commit in about eight months while 191 open issues keep accumulating - the open-source repo and the commercial line have diverged, so that MIT repository is not a sample of Trae's current engineering. Model availability is constrained twice over (Seed, MiniMax and GLM unavailable to US users; the whole GPT-6 and GPT-5.6 families plus GPT-5.5, GLM-5.2, DeepSeek-V4-Flash and Kimi-K3/K2.7-Code require the new plan), SOLO is excluded from Free, and only Ultra gets early access to new models. The usable model set therefore differs by region, and selecting against the public model list alone will give the wrong answer for cross-region teams.

AI CodingCode HarnessSOLO ModeCloud Agent

What it is: ByteDance's two-form-factor coding line - TraeCode (IDE + SOLO) and TraeWork (web / desktop / mobile)

Trae is ByteDance's AI coding product line, and the documentation splits it into two clearly different things. TraeCode is "a development tool deeply integrated with AI capabilities", offering a complete experience across coding, project understanding, debugging and running, and change management: you can stay in control of every step as in a traditional IDE, or delegate complex tasks to agents for planning and execution. TraeWork grew out of TraeCode's SOLO mode as an AI-native workspace with web, desktop and mobile clients and three modes - Work / Code / Design - aimed beyond developers at product managers, data analysts, operations and design roles.

TraeCode's dual mode is its most recognizable structure:

  • IDE mode: keeps the familiar workflows - editor, terminal, debugging, extensions, source control - for cases needing fine-grained control over code changes and execution.
  • SOLO mode: AI takes the lead. Describe requirements in natural language, by voice, or by uploading local files, and the AI decomposes the task and runs the whole path: code generation, testing, preview, change summaries, deployment. The UI is task management panel, AI chat panel, tool panel (built-in editor, documentation viewer, browser).

Max mode: 1M context, 200 tool rounds per task, 750 lines per read

These are the three concrete ceilings Trae publishes for long tasks: context window expanded to a maximum of 1M, up to 200 rounds of tool invocation in a single task (for multi-step, multi-dependency work), and up to 750 lines read at a time (reducing segmented processing). Documented fits include: rapid first drafts for large complex projects (import dependencies, data structures and config files at once and produce a runnable global prototype), analysis and implementation from long documents (read a lengthy PRD, design doc or compliance agreement straight into code), cross-module and cross-file understanding and refactoring (SDK or framework upgrades, global naming conventions, cross-module API refactoring), automation scripts for complex multi-step processes (CI/CD pipelines, cross-service orchestration, automated test scripts), and context preservation during real-time interactive development.

The more interesting part is the warning Trae writes about itself: "Enabling Max mode will significantly increase costs... For everyday development tasks, the default context window is usually sufficient." A vendor talking users out of its own high tier, inside feature documentation rather than on the pricing page, is unusual honesty.

CUE and Agent: the completion layer and the execution layer are built separately

CUE is the completion-side capability set: code completion, chained completion, multi-line edits, next-edit prediction and navigation, plus dependency imports and reference renaming in Python, TypeScript and Golang projects.

Agent is the execution side. The documented five-stage workflow reads like a harness design reference: requirement analysis, code research (search the codebase, docs and online resources, locate relevant files, analyze existing implementation), solution design (break down steps and dynamically optimize the modification plan), implementation of changes (which may include recommending new dependencies, terminal commands to execute, and guidance for manual operations outside the client), and delivery and acceptance (hand control back after validation and summarize all modifications). Two agents are built in: Chat for quick technical Q&A and troubleshooting, and Agent for automated project development - its distinction from an ordinary conversational agent is that it first generates an actionable plan from your goal and project context, then develops step by step only after you confirm the plan. Custom agents are supported with configurable prompts, MCP servers and toolsets, and ready-made custom agents can be imported in one click.

The context types you can attach are also fully enumerated: files, folders, code snippets, terminal output, repositories, document sets, and even webpages.

SOLO's delivery chain: Figma to code, Supabase, Vercel, Stripe, DiffView

SOLO wires the key third parties along the design-to-production path into tools: Figma to code parses design files and converts elements into executable code, either a whole frame for a full page or precisely selected components (buttons, forms, cards); Supabase connects a cloud PostgreSQL database so you design screens and build the schema at the same time; Vercel deploys and returns a shareable link, with re-deployment as content updates; Stripe is callable as a tool for payments; and AI services can be configured and integrated into the web app as needed.

Two interaction details deserve separate mention. DiffView opens from the chat panel's Open Diff button and shows the number of affected files, total lines changed and the list of modified files, with per-file diffs - collapsing agent output into one reviewable change. Conversation Auto-Fold (Settings > Conversation > To-Do List) automatically folds and summarizes completed tasks, expandable on demand. Task management supports running multiple tasks concurrently within a single project, breaking the traditional serial model.

TraeWork: SOLO turned into a cross-device dispatch workspace

TraeWork positions itself as a "pocket AI agent hub". The three clients share one account system and task data with real-time sync, and the division of labour is explicit: mobile is the cross-device task dispatch center, while web and desktop handle task execution, deeper interaction and result review. The desktop app runs independently of TraeCode IDE, supports both local and cloud tasks, takes text, voice, attachments and skills as input, shows live progress with automatic output summaries, and lets you preview and accept results inside the chat surface. Mobile defaults to press-and-hold to speak, centrally manages cloud TraeWork plus multiple personal computers, dispatches concurrent tasks to different devices with live progress monitoring, and automatically fails over to cloud execution when a device goes offline so tasks are not interrupted.

The cloud agent provides unified runtime and dependency management, executing all code in a stable isolated remote environment and avoiding compatibility or performance issues caused by local environment differences. The three modes are aimed differently: Work for people who do not develop (documents, data, presentations), Code for engineers used to agent-driven development (coding, debugging, repository management, Git workflows), and Design for an end-to-end AI workflow of creating, refining and delivering designs.

Security and engineering furniture: Privacy mode, sandboxed execution, AI code review

With Privacy mode on, Trae does not use your chats, code snippets or AI-generated outputs for data analysis, product optimization or model training, and codebase files always remain on your local device. Sandboxed execution runs agent-generated commands in a restricted environment, reducing misuse risk through file access controls and high-risk command blocking policies.

The engineering furniture is not sacrificed to "AI first": a complete built-in Git workflow with AI-generated standardized commit messages; AI code review over uncommitted changes, individual commits or branch diffs, presented as summaries, flowcharts and diff views; an extension store for editor, language support and debugging tools; and Remote SSH / WSL for developing in remote environments.

Model surface: 18 built-in models plus BYOK, and two regional restrictions

The built-in list (each item checkable) spans three US labs and Chinese vendors: Seed-2.1-Turbo, GPT-6-Astra / Sol / Luna, GPT-5.6-Sol / Terra / Luna, GPT-5.5 / 5.4 / 5.2, GLM-5.2, DeepSeek-V4-Flash, Kimi-K3 / K2.7-Code / K2.5, Gemini-3.1-Pro-Preview / Gemini-3-Flash-Preview, MiniMax-M3 / M2.7, with custom models addable via API key. Unit prices are tiered at a 272k context boundary: GPT-6-Astra is $10 / $50 per million tokens (input / output) at or below 272k and $20 / $75 above it; GPT-6-Sol is $2 / $10, rising to $4 / $15; Seed-2.1-Turbo is $0.50 / $2.50.

Two restrictions must be stated: first, the Seed, MiniMax and GLM series are not available to users in the United States; second, GPT-6-Astra/Sol/Luna, GPT-5.6-Sol/Terra/Luna, GPT-5.5, GLM-5.2, DeepSeek-V4-Flash, Kimi-K2.7-Code and Kimi-K3 are only available after upgrading to the new plan.

Commercial terms

Five subscription tiers: Free $0, Lite $8/mo ($80/yr, about $6.70/mo), Pro $20/mo ($200/yr), Pro+ $60/mo ($600/yr), Ultra $200/mo ($2,000/yr, about $166.70/mo), plus On-Demand Usage. Concurrent cloud tasks scale by tier: 2 / 2 / 10 / 15 / 20; SOLO mode is not included in Free; only Ultra gets early access to new models. Lite is regional pricing, offered only in Thailand, Indonesia, Vietnam, the Philippines, India, Pakistan, Turkey, Malaysia, Brazil, Colombia, Peru and Nigeria.

trae-agent: the open-source half, and its stall signal

Trae has an MIT-licensed open-source command-line agent, bytedance/trae-agent, from ByteDance's Lakeview research team with an accompanying paper, arXiv:2507.23370. It supports multiple LLM providers, records full execution trajectories for replay and training, and is configured in YAML (max_steps defaults to 200). GitHub data as of 2026-09-28: 12,119 stars, 1,352 forks, 191 open issues, not archived. To be plain about it: the last push was 2026-02-05, roughly eight months without a commit while 191 open issues keep accumulating - the open-source repository and the commercial product line have clearly diverged in iteration pace.

Boundaries

Trae publishes no third-party benchmark readings (no SWE-bench or Terminal-Bench style self-evidence); every capability claim comes from feature documentation and vendor description. The clients are closed source, and while trae-agent is MIT-licensed it has not been updated in about eight months, so it cannot represent Trae's current engineering implementation. Max mode's 1M / 200 rounds / 750 lines are ceilings, not typical experience, and the vendor itself says costs rise significantly. Model availability is constrained twice over - by region (Seed, MiniMax and GLM unavailable to US users) and by plan tier - so the usable model set differs across teams in different regions. Concurrent cloud tasks are a hard tier differentiator and SOLO is excluded from Free, so evaluation cannot look at monthly price alone.

More in AI Coding

4
AI CodingDevelopmentTopC

Qoder: Alibaba agentic platform for real work - nine product lines on one knowledge engine

Qoder is the agentic coding platform from Alibaba, positioned officially as an agentic platform for real work: not an editor but an end-to-end loop - understand the task and context, plan, call tools, verify results, iterate toward the deliverable - resting on three stated principles (context engineering, agent autonomy, goal-directed loops), with nine product lines sharing one knowledge engine (desktop Qoder and Qoder IDE coexisting rather than replacing each other, Editor and Quest forms, a JetBrains plugin, Qoder CLI, cloud agents and more); session history and memory are stored separately but can be imported from the IDE. Repo Wiki is generated locally by multiple agents, never uploads the codebase, is off by default and supports Auto Update, Auto Export and Citation back to source locations. Quest has four drives - Agent, Experts, Goal and Spec (convertible to scheduled tasks): Spec runs requirement clarification (multiple choice, with Recommend / Continue / Skip), a structured Spec covering requirements, design, task breakdown and acceptance criteria, human review, execution, then Review/Commit/Push, while Goal takes only the desired outcome and evaluates progress at the end of every round, continuing automatically until met. Two scaled cases: building Qoder with Qoder (10 people, 3 weeks, 500,000 lines of agent code merged into a 4-million-line legacy system, 99% agent-generated, still in production at v1.4.0 with zero incidents; the method is a cognitive base plus Ultra Spec plus Experts cross-review plus a verifier agent filtering hallucinated issues, with humans only deciding SLO definitions and irreversible operations, and each person driving 20-plus Experts tasks a day); and AutoSDK for AMap in-car systems across 20-plus repositories and over a million lines, where the strict first-pass rate went from 37.3% to 61.5% (problem framing cites KoCo-Bench / arXiv:2601.13240v3: general coding reaches 90% Pass@1 while domain code generation reaches only 8.9%). Boundaries: the client and knowledge engine are closed; every scaled number comes from official cases and vendor self-reporting, and self-evidence from a product about itself carries methodological self-interest, none of it independently reproduced; Experts cost per unit is clearly above a single agent (median about 75 versus about 50 Credits) and Credits reset each cycle rather than accumulating. Confidence C (vendor-claim).

500,000 行 · 99% agent 生成10 人 3 周并入 400 万行遗留系统(厂商案例)Vendor Claim · 2026
ProductAlibabaSiteRepo
Qoder: Alibaba agentic platform for real work - nine product lines on one knowledge engine
AI CodingDevelopmentTopC

Kimi Code: a model lab that builds its own harness, then opens it to every other shell over two protocols

Kimi Code is the developer coding service from Moonshot AI at kimi.com/code, and it sits in an unusual spot in this cohort: most harnesses are shell vendors plugging into models, while Kimi Code is a model lab building its own shell and then opening the API over two protocols at once - OpenAI-compatible at api.kimi.com/coding/v1 (China) and api.kimi.ai/coding/v1 (overseas), Anthropic-compatible at api.kimi.com/coding/ and api.kimi.ai/coding/ - with dedicated integration guides for Claude Code, OpenCode, Codex and Hermes Agent, which amounts to opening its own models to every other shell. Three clients run in parallel: Desktop (released 2026-09-17 for macOS Apple Silicon/Intel and Windows, moving the CLI agent core into a GUI), the CLI (the kimi command, widest feature surface, install.sh verifies checksums; Windows relies on Git Bash from Git for Windows with KIMI_SHELL_PATH for non-standard bash.exe) and a VS Code extension. The model surface is three models across four model IDs - k3, k3-256k, kimi-for-coding, kimi-for-coding-highspeed - and the docs insist on the model ID rather than a version name (writing K3 or K2.8 Preview fails outright), since a misspelled HighSpeed ID falls back silently instead of erroring. K2.7 Code HighSpeed runs about 180 tokens/s (up to 260 on short context) at 5-6x speed for 3x quota, and it only accelerates model output, so rounds dominated by tool calls feel little faster. K2.7 Code was released and open-sourced on 2026-06-12 with official deltas over K2.6 of Program-Bench +10.4%, MCP Mark Verified +11.4%, SWE Marathon +76.2% and 30% fewer reasoning tokens. Boundaries: all readings come from the vendor or vendor-relayed external benchmarks and were not independently reproduced; the K3 technical report is not public alongside the weights; quota is tightly bound to membership tiers (the 1M tier costs about twice the 256K tier, HighSpeed three times); older models retire fast (kimi-k2 on 2026-05-25, kimi-latest on 2026-01-28, kimi-k2.5 and moonshot-v1 on 2026-08-31). Confidence C (vendor-claim).

4 个 model ID · 1M 上下文内置模型面(K3 / K2.8 Preview / K2.7 Code HighSpeed)Vendor Claim · 2026-09
ProductMoonshot AISite
Kimi Code: a model lab that builds its own harness, then opens it to every other shell over two protocols
AI CodingDevelopmentTopC

ZCode: the official GLM-5.3 harness, defined as an ADE rather than an editor

ZCode is the official harness Z.ai built for GLM-5.3, and it defines itself as neither an AI editor nor a CLI but an ADE (Agentic Development Environment): it turns the 1M context window and long-horizon capability of GLM-5.3 into a stable desktop experience covering planning, coding, review and iteration, keeping goal, files, terminal output, browser context, execution mode and Git state inside one task so continuity survives from plan through implementation to verification, with model capability, tool calling and the execution chain tuned over multiple rounds against GLM-5.3. The most distinctive design is /goal mode: after a goal is set, every round ends with an automatic check of whether the goal is met, and the agent continues into the next round on its own until completion is confirmed (the documented example is a research task running twenty-plus rounds, with a right-hand panel showing what each round did). One session holds one goal at a time; /goal shows it, /goal sets it (replacing any existing goal), /goal replace swaps it explicitly, /goal pause suspends it, and the docs are honest about the fit - work that is easy to state in one sentence but takes many rounds to finish. The card metric keeps the verifiable price fact layer: GLM-5.3 API at $1.4 input / $4.4 output per million tokens. Boundaries: the client is not open source and the deep tuning only covers the GLM-5.3 family; a stable 1M context is a vendor claim and mid-context recall is an industry-wide weakness, so stuffing a whole repo is no substitute for retrieval; idle-time tasks are being rolled out to subscribers rather than always available; the free quota is a 5-day window, not a lasting benefit. Confidence C (vendor-claim).

$1.4 / $4.4GLM-5.3 API 单价(输入/输出,每百万 token)Vendor Claim · 2026-09
ProductZ.aiSite
ZCode: the official GLM-5.3 harness, defined as an ADE rather than an editor
AI CodingDevelopmentTopA

Claude Code: the agentic coding tool that lives in the terminal, and the reference harness for Anthropic models

Claude Code is the agentic coding tool from Anthropic, and its own definition is deliberately modest: it lives in your terminal, understands your codebase, and executes routine tasks, explains complex code and handles git workflows through natural language. The official docs name three surfaces - terminal, IDE, and tagging @claude on GitHub. anthropics/claude-code stood at 148,438 stars / 24,902 forks on 2026-09-28, among the highest in this class; stated plainly, the repository has no LICENSE file (the GitHub API returns null), so it is publicly readable but not openly licensed - a material difference from the Apache-2.0 core of Codex CLI that must not be blurred during evaluation. Installation moved from npm to native installers (claude.ai/install.sh, install.ps1, brew cask, winget) and npm install -g @anthropic-ai/claude-code is officially deprecated, usually a way to escape the Node version matrix and global-permission support burden. The plugins/ directory is not a shell: it ships 13 officially maintained capability packs - code-review, pr-review-toolkit, commit-commands, feature-dev, frontend-design, security-guidance, agent-sdk-dev, plugin-dev, hookify, the explanatory and learning output styles, the opus 4.5 migration and ralph-wiggum - each a bundle of custom commands plus agents, turning personal prompts into installable team assets. It is also the reference harness for Anthropic models in third-party benchmarks (the Cursor footnote routes Anthropic model scores through Claude Code). Boundaries: no open license, closed models with unavoidable data egress, a terminal-first form factor that costs non-CLI users a learning curve, and a ceiling that depends on how well any third-party model adapts to the Anthropic protocol and tool calling when swapped in. The card metric is the verifiable fact (stars). Confidence: confirmed.

148,438GitHub stars(仓库无 LICENSE)Confirmed · 2026-09
ProductAnthropicSiteRepo
Claude Code: the agentic coding tool that lives in the terminal, and the reference harness for Anthropic models