# WTAgent — Open-source CLI that drives your existing ChatGPT, Claude or DeepSeek web session as a local coding agent.

> Source: The Agents Index — https://theagentsindex.com/wtagent (structured, researched, re-verified)
> Facts last verified: 2026-09-21

WTAgent is an open-source Node.js command-line coding agent that drives a signed-in web chat session in a dedicated Chrome profile instead of an API key. It reads, writes and runs code locally under a policy and approval layer.

| Fact | Value |
| --- | --- |
| Website | https://xixian42.github.io/wtagent |
| Price | Open source · MIT |
| Free tier | Yes |

The CLI is free and MIT-licensed with no paid tier; the practical limit is the quota of the web AI account you sign in with, since WTAgent uses your own account, models and quota.

## What it does

**Writes code** · Yes — Takes a plain-language task in the terminal and creates source files in the working directory, then runs them to check the result.

> WTAgent creates the files in the current directory, runs the script locally, and checks its output.
> — https://github.com/XiXian42/wtagent#quick-start · README, Quick start section, after the hello.js example

**Debugs and fixes** · Yes — The documented V1 loop runs builds, tests and linters, reads errors from the output and keeps editing until verification passes.

> 4. 安装依赖并执行构建、测试和静态检查。 5. 启动开发服务并读取日志。 6. 根据错误继续修改。
> — https://github.com/XiXian42/wtagent/blob/main/docs/technical-design.md · Technical design, section 2.1 首版目标, numbered loop

**Ships it** · Partly — Pushing, releasing and deploying are permitted but classified as external-risk actions that always stop for an explicit human approval prompt.

> | `external` | `git push`、发布、部署 | 审批 |
> — https://github.com/XiXian42/wtagent/blob/main/docs/technical-design.md · Technical design, section 8.1 风险分级, risk table final row

_Not established: Writes tests, Reviews code, Whole-codebase work_

## How much it does unattended

**Runs autonomously** · Partly — Reads, writes and ordinary project commands run automatically, but a human must sign in to the web AI first and approve anything sensitive or destructive.

> 项目目录内的常规开发操作可自动执行。 - 访问项目外、读取凭证、提权、批量删除、推送或部署必须确认。
> — https://github.com/XiXian42/wtagent/blob/main/docs/technical-design.md · Technical design, section 2.2 固定边界, bullet list

**Agent permissions** · Partly — A local policy engine grades every tool call by risk: reads, writes and builds inside the project run unattended, sensitive or destructive calls need approval.

> | `read` | 项目内读取、搜索、列目录 | 自动 | | `write` | 项目内创建、编辑文件 | 自动 | | `sensitive` | 访问项目外、读取凭证、外部网络写操作 | 审批 |
> — https://github.com/XiXian42/wtagent/blob/main/docs/technical-design.md · Technical design, section 8.1 风险分级, risk table

_Not established: Multi-agent_

## Which models it runs on

**Claude** · Yes — Claude is a first-class provider selected with a flag; it uses whichever model your claude.ai browser profile already has chosen.

> Claude keeps the model selected by claude.ai in its browser profile; WTAgent does not override it.
> — https://github.com/XiXian42/wtagent#quick-start · README, Quick start, paragraph after the --model examples

**GPT** · Yes — ChatGPT is the default provider and the product's headline use case, including Pro accounts, driven through the signed-in web session.

> Turn ChatGPT Web (including Pro) into Codex.
> — https://xixian42.github.io/wtagent/ · Hero, product mapping line

**Open models** · Partly — Providers include DeepSeek, Kimi and GLM, whose weights are published openly, but WTAgent drives their hosted web chat rather than running any weights locally.

> Supported values are `chatgpt`, `claude`, `deepseek`, `gemini`, `kimi`, and `glm`
> — https://github.com/XiXian42/wtagent#quick-start · README, Quick start, --model provider sentence

**Model choice** · Yes — You pick the provider on the command line from a published list, and two of them inherit whatever model your browser profile has selected.

> Choose a provider with `--model` (ChatGPT is the default).
> — https://github.com/XiXian42/wtagent#quick-start · README, Quick start, provider selection paragraph

## Where you use it

**In your editor** · No — There is no editor extension or plugin: the vendor scopes the first version to a terminal CLI and says outright that no plugin is involved.

> V1 是 CLI，不做桌面 GUI 和远程控制台。
> — https://github.com/XiXian42/wtagent/blob/main/docs/technical-design.md · Technical design, section 2.2 固定边界, last bullet

**On the command line** · Yes — The terminal is the only interface: you install globally with npm, run wtagent in a project directory and type tasks at a prompt.

> npm install -g wtagent
> — https://xixian42.github.io/wtagent/ · How to use, step 1 Install WTAgent command block

**In your pipeline** · No — It needs a visible, non-headless Chrome with a manually signed-in web account, which rules out unattended pull-request or CI execution.

> 以非 headless 方式启动 Chrome
> — https://github.com/XiXian42/wtagent/blob/main/docs/technical-design.md · Technical design, section 4.2 BrowserManager, responsibilities list

**In a browser** · No — No hosted dashboard or remote console exists; the vendor explicitly excludes a desktop GUI and a remote console from this version.

> V1 是 CLI，不做桌面 GUI 和远程控制台。
> — https://github.com/XiXian42/wtagent/blob/main/docs/technical-design.md · Technical design, section 2.2 固定边界, last bullet

## Whose machine it runs on

**Self-hosted** · Yes — Everything but the chat itself runs on your own machine: files are created and commands executed locally by the installed CLI.

> Files and commands run locally, but code and output returned to the chat are sent to your web AI provider.
> — https://xixian42.github.io/wtagent/ · Usage notes, second paragraph

**Open source** · Yes — The full source sits in a public GitHub repository under the MIT licence, with the licence linked from the site footer.

> Open source · MIT
> — https://xixian42.github.io/wtagent/ · Hero facts row beneath the install command

## What it costs to run

**How it meters** · Yes — The tool itself charges nothing; consumption is metered by whichever web AI account you sign in with, under that provider's own quota.

> WTAgent itself does not require an API key. It uses your own web account, available models, and quota.
> — https://github.com/XiXian42/wtagent#quick-start · README, Quick start, paragraph on accounts and quota

**Free tier** · Yes — A single global npm install with no sales contact, no account with the vendor and no API key required to start.

> No API key. No MCP. No tunnel. No plugin.
> — https://xixian42.github.io/wtagent/ · Hero lede line above the action buttons

**API access** · No — It ships only an interactive terminal binary; the vendor documents no HTTP API or library surface you could call from your own stack.

> V1 是 CLI，不做桌面 GUI 和远程控制台。
> — https://github.com/XiXian42/wtagent/blob/main/docs/technical-design.md · Technical design, section 2.2 固定边界, last bullet

**MCP server** · No — The vendor advertises the absence of MCP as a selling point, so no other agent can query WTAgent over that protocol.

> No API key. No MCP. No tunnel. No plugin.
> — https://xixian42.github.io/wtagent/ · Hero lede line above the action buttons

**Bring your own key** · Partly — You bring your own inference, but as a signed-in web chat account in a dedicated Chrome profile rather than an API key or your own cloud.

> A dedicated Chrome profile opens. Sign in with your own web AI account when prompted. ChatGPT is the default provider.
> — https://xixian42.github.io/wtagent/ · How to use, step 2 Open a test project and start

## Buying it for a team

**Admin controls** · Partly — Constraints are local, not organisational: a policy engine bounds paths to the project root and forces approval prompts, but no central admin console exists.

> 所有文件路径必须： 1. 相对项目根目录解析； 2. `realpath` 后仍位于项目根目录；
> — https://github.com/XiXian42/wtagent/blob/main/docs/technical-design.md · Technical design, section 8.2 路径边界, numbered rules

**Audit log** · Partly — Each session writes a local event log and raw tool output to disk, which is a developer record rather than an administrative audit trail.

> sessions/<session-id>/ session.json events.jsonl rollout-<timestamp>-<session-id>.jsonl tool-output.jsonl
> — https://github.com/XiXian42/wtagent/blob/main/docs/technical-design.md · Technical design, section 9.1 本地 Session 记录, directory listing

## What happens to your code

**Opt out of training** · No — WTAgent offers no opt-out control of its own; whatever code it sends into the chat falls under your provider's terms and settings.

> Use your own account and follow your provider’s terms and limits.
> — https://xixian42.github.io/wtagent/ · Usage notes, first paragraph

**Data residency** · Partly — Your files and commands never leave your machine, but any code pasted into the chat lands in your web AI provider's jurisdiction, not WTAgent's.

> Files and commands run locally, but code and output returned to the chat are sent to your web AI provider. Do not share secrets or unauthorized data.
> — https://xixian42.github.io/wtagent/ · Usage notes, second paragraph

**Getting out** · Yes — There is nothing to export from a vendor: session state, event logs and tool output are plain JSON and JSONL files on your own disk.

> 状态持久化 | V1 使用原子 JSON + JSONL 事件日志；数据量增大后再迁移 SQLite
> — https://github.com/XiXian42/wtagent/blob/main/docs/technical-design.md · Technical design, section 1 结论, technology table row 状态持久化

_Not established: Certifications_

## Provenance

2 of 3 sources behind the facts above are independent of the vendor.

## Coverage elsewhere

None recorded. Every fact above comes from the vendor and is not independently corroborated yet.

## What changed

- 2026-09-21 — Capabilities · Commercial terms · Free tier · Price

## Correcting this record

Anything marked "Not established" is a question we have not answered — it is not a No. If you represent WTAgent, or anything here is wrong or out of date, you can supply it with a source: the operations this site accepts from an agent are listed at https://theagentsindex.com/.well-known/agents.json — that catalogue needs no key to read, and each entry states its own auth.

Supplying evidence is free. Nothing that can be bought here changes what a field says, whether a gap is closed, or how this record is scored.