Chinese AI office agents 2026: which of the 21 take your own API key
As of 2026-11-11, 8 of the 21 products in the Ifenxi 2026 China AI Office Ranking (October) have official docs that let a personal user enter their own API key and endpoint: ZCode, DeepSeek Harness, Qoder, Kimi Code, MiniMax Code and LobsterAI support both the Anthropic and OpenAI protocols, and WorkBuddy and AutoClaw support the OpenAI protocol. 4 more have a custom-model entry whose protocol is not documented. The table lists all 21 in ranking order.
Updated 2026-11-11
Point this tool at the QCode endpoint — one key for Claude, GPT and Chinese models.
- Billed per token — live rates on /models
- Pay by card (Visa / Mastercard), Apple Pay, Google Pay or crypto
- Self-service: top up and activate instantly
21 office agents: October MAU and whether you can bring your own API key
| Rank · Product (vendor) | MAU (Ifenxi, October) | Your own API key (personal user) | Protocol and where to enter it |
|---|---|---|---|
| 1Doubao Work (Volcengine) | 55.00M | Not in the official docs | Personal plan: built-in Doubao 2.1 models; enterprises can add their own |
| 2WorkBuddy (Tencent) | 41.49M | Yes (OpenAI protocol only) | OpenAI Chat; Settings → Models → Custom (CodeBuddy uses the same protocol) · Setup guide |
| 3QwenWork (Alibaba) | 47.48M | Not in the official docs | Personal plans: built-in model tiers only; enterprise flagship has BYOK |
| 4ZCode (Zhipu) | 1.79M | Yes | Anthropic / OpenAI; Model Settings → Add Provider · Setup guide |
| 5DeepSeek Harness (DeepSeek) | 2.10M | Yes | 3 protocols; Settings → Models → Custom model API · Setup guide |
| 6DuMate (Baidu) | 7.32M | Not in the official docs | Personal docs show no model setting; model use is billed in points |
| 7Nami Work (360) | 1.61M | Entry exists; protocol not stated | The model menu shown on the official site ends with a custom-model entry |
| 8Kimi Work (Kimi) | 0.50M | Not in the official docs | Help center: ships Kimi K3 since version 3.1.0; no custom endpoint |
| 9Kuku AI (Baidu) | 1.07M | Not in the official docs | Official intro covers experts, skills and MCP; no model setting |
| 10Qoder (Alibaba) | 0.82M | Yes | Desktop app and CLI: OpenAI / Anthropic Compatible; Settings → Models · Setup guide |
| 11Kimi Code (Kimi) | 0.41M | Yes | 3 protocols; CLI file ~/.kimi-code/config.toml · Setup guide |
| 12Skywork (Kunlun Tech) | 1.64M | Not in the official docs | Built-in Claude, GPT, Gemini, DeepSeek and more to choose from |
| 13Xiaohuanxiong (SenseTime Raccoon) | 1.16M | Not in the official docs | Plans are split into standard and upgraded models |
| 14MiniMax Code (MiniMax) | 0.13M | Yes | 3 formats; Settings → Usage & model · Setup guide |
| 15Marvis (Tencent) | 0.32M | Not in the official docs | Device-cloud and on-device models; custom API listed as a future upgrade |
| 16AutoClaw (Zhipu) | 0.21M | Yes (OpenAI protocol only) | OpenAI; Settings → Models & API → add a custom model · Setup guide |
| 17TeleAgent (China Telecom) | 0.22M | Not in the official docs | Product page says you can switch models; no endpoint field |
| 18LobsterAI (NetEase Youdao) | 0.14M | Yes | Anthropic / OpenAI; Settings → Custom Model → Add Custom · Setup guide |
| 19WPS Lingxi Pro (Kingsoft Office) | 0.50M | Entry exists; protocol not stated | Other-model entry takes keys from Tencent Cloud, Volcengine, Zhipu and others |
| 20EasyClaw (Cheetah Mobile) | 0.07M | Entry exists; protocol not stated | Add Custom Model, version 1.3.24 or later |
| 21AiPy (Knownsec) | 0.20M | Entry exists; protocol not stated | Custom model takes an API address; needs the Zhuji rank (50 points) |
Rank and MAU are from the Ifenxi 2026 China AI Office Ranking (October). The source gives MAU in wan (10,000 users); this table converts it to millions, so 179 wan = 1.79M. The API key column does not come from the ranking: it follows each product's public official docs, checked on 2026-11-11. Not in the official docs means those docs show no place where a personal user can enter a custom model address or API key.
Four points first
Documented for personal users
As of 2026-11-11, official docs for 8 of the 21 products let a personal user enter a custom model address and API key: ZCode, DeepSeek Harness, Qoder, Kimi Code, MiniMax Code and LobsterAI support both Anthropic and OpenAI, while WorkBuddy and AutoClaw use OpenAI Chat Completions.
The entry is there, the settings are not
Nami Work, WPS Lingxi Pro, EasyClaw and AiPy show a custom-model or own-API-key entry in official material but do not say whether it speaks Anthropic or OpenAI. WPS Lingxi Pro lists keys from providers such as Tencent Cloud, Volcengine and Zhipu; AiPy requires the Zhuji account rank (50 points) for custom models.
How QCode fits
One QCode API key gives you both Anthropic- and OpenAI-compatible endpoints: Claude models on https://api.qcode.cc/api, GPT models on https://api.qcode.cc/openai/v1, and the GLM, Kimi, DeepSeek and Qwen families on either. All 8 products marked Yes can take it.
How to read the ranking figures
MAU is cited from the Ifenxi 2026 China AI Office Ranking (October), for October 2026, in units of 10,000 users, and only shows how many people use each product. Whether you can bring your own key has nothing to do with the ranking; it comes from each product's official docs, checked on 2026-11-11.
How to read the table
Rows follow the Ifenxi ranking order, with four columns: product and vendor, October MAU, whether a personal user can bring an API key, and the protocol plus where to set it. The third column uses four labels only. Yes means the official docs let a personal user enter a custom address and support both the Anthropic and OpenAI protocols; then Yes (OpenAI protocol only); Entry exists; protocol not stated; and Not in the official docs, meaning that as of the check date the public official material shows no place for a personal user to enter a custom model address or API key. The 8 products with a QCode setup article have a Setup guide link in the fourth column. WorkBuddy, Kimi Work and DeepSeek Harness also have product guides at /workbuddy-guide, /kimi-work-guide and /deepseek-harness-guide, and the OpenClaw ecosystem around AutoClaw and LobsterAI is covered at /claw-ecosystem-guide.
Doubao Work and QwenWork: own models on the enterprise side
Doubao Work and QwenWork, ranked first and third, offer only built-in models on their personal plans: Doubao Work's work task mode lists Auto, Doubao 2.1 Lite and Doubao 2.1 Pro, and the QwenWork desktop app picks one of Advanced, Standard or Qwen3.8-Max. Both have an enterprise route: Doubao Work's team and enterprise quota page says Doubao work partners and Feishu aily enterprise partners and workflow apps can connect models the company deploys itself or third-party model services it buys, and QwenWork's enterprise flagship BYOK (bring your own key) is limited to customers whose BYOK entitlement is turned on through sales.
Timeline
Qoder 0.1.8 release notes: Personal plan BYOK now supports custom Base URLs for any OpenAI- or Anthropic-compatible model service. This applies to the new Qoder desktop app; Qoder IDE, from the same family, still lists preset providers only on its custom models page.
LobsterAI release 2026.10.23: custom providers now support OpenAI Responses and Gemini native APIs.
Ifenxi publishes its first 2026 China AI Office Ranking for October 2026, scoring monthly active users and user mentions at 50% each; 21 products are listed.
Documented, and not spelled out yet
Documented and tested
As of 2026-11-11, the official docs that let a personal user enter a custom address cover ZCode, DeepSeek Harness, Qoder (new desktop app and CLI), Kimi Code, MiniMax Code and LobsterAI (Anthropic and OpenAI), plus WorkBuddy and AutoClaw (OpenAI Chat Completions). Tested on 2026-11-10 with ZCode CLI 0.17.1, MiniMax Code 0.7.6 and Kimi Code CLI 2.5.2: the QCode Anthropic, OpenAI Chat Completions and OpenAI Responses addresses all work, with no sign-in to each vendor's account.
Not spelled out by the vendor yet
1) Nami Work: the model menu shown in the product demo on its official site ends with a custom-model entry, and its user agreement mentions configuring a third-party model yourself, but no fields or protocol are documented. 2) Marvis: the site describes a device-cloud efficiency mode and an on-device local mode; its FAQ lists a custom API interface among future upgrades. 3) TeleAgent: the product page says you can switch models freely and shows no endpoint or key field. 4) DuMate and Kuku AI: as of 2026-11-11 their personal docs and official introductions show no model setting. These rows follow each product's later official docs.
Which products speak which protocol
Anthropic and OpenAI (6 products)
Custom models in ZCode, DeepSeek Harness, Qoder, Kimi Code, MiniMax Code and LobsterAI support both the Anthropic and OpenAI protocols: put Claude models on the QCode Anthropic address, GPT models on the OpenAI address, and Chinese models on either. Qoder here means the new desktop app and CLI, and Kimi Code is configured in the CLI file ~/.kimi-code/config.toml. Step-by-step setup: /zcode-custom-model-api-setup, /qoder-custom-model-byok-setup, /kimi-code-custom-provider-setup and /lobsterai-custom-model-api-setup.
OpenAI protocol only (2 products)
Custom models in WorkBuddy and AutoClaw use OpenAI Chat Completions. On QCode that protocol serves GPT and Chinese models (DeepSeek / GLM / Kimi / Qwen) such as gpt-6.1-sol and glm-5.3; for Claude models, use ZCode, Qoder or Claude Code, which support the Anthropic endpoint. For WorkBuddy the full path https://api.qcode.cc/openai/v1/chat/completions is the recommended address; CodeBuddy, from the same company, also uses the OpenAI protocol, see /codebuddy-custom-model-models-json.
How to enter the same QCode key
1) Copy your API key, which starts with cr_, from the QCode dashboard. 2) Pick the address for the protocol the product supports: Anthropic is https://api.qcode.cc/api (the tool appends /v1/messages); OpenAI Chat Completions is https://api.qcode.cc/openai/v1, or https://api.qcode.cc/openai/v1/chat/completions where a full endpoint path is required (WorkBuddy, or models.json in CodeBuddy); for OpenAI Responses, tools that append /responses to the base take https://api.qcode.cc/openai/v1, and the Codex-style form is https://api.qcode.cc/openai; the two are equivalent. 3) Copy model IDs exactly as listed on /models, for example claude-sonnet-5-5, gpt-6.1-sol and deepseek-v4.1-flash. If your network is unstable, swap the host for asia.qcode.cc or us.qcode.cc and keep the path.
One QCode key, two compatible endpoints
As of 2026-11-11, one QCode API key gives you both Anthropic- and OpenAI-compatible endpoints: Claude models on the Anthropic endpoint https://api.qcode.cc/api, GPT models on the OpenAI endpoint https://api.qcode.cc/openai/v1, and the GLM, Kimi, DeepSeek and Qwen families on either, so all 8 products marked Yes can share one key. Usage is billed per token with per-model rates on /models. Three steps: 1) register and top up (Alipay and other major payment methods are supported); 2) buy a Personal Plan or a Pay-as-you-go Pack and copy your API key, which starts with cr_; 3) enter the address and key in each product as shown in the table. Not ready for a monthly plan? Pay-as-you-go Packs (/pay-as-you-go) start at $4 for a one-time quota you use within 7 days; if you already have a Personal Plan, add a booster pack when you need more quota.
Frequently asked questions
Can Doubao Work use my own API key?
As of 2026-11-11, the personal docs show no place to enter your own API key: work task mode offers Auto, Doubao 2.1 Lite and Doubao 2.1 Pro, and the product introduction says it runs on ByteDance's own models with built-in third-party models such as Kimi, GLM and DeepSeek. The enterprise side is different: the team and enterprise quota page says companies can connect models they deploy themselves or third-party model services they buy. To use your own key in an office agent, see the 8 products marked Yes.
Can WorkBuddy connect to Claude?
WorkBuddy's custom model supports OpenAI Chat Completions, so you can enter the QCode OpenAI address (the full path https://api.qcode.cc/openai/v1/chat/completions is recommended) and call GPT and Chinese models such as gpt-6.1-sol and glm-5.3. For Claude models, use ZCode, Qoder or Claude Code, which support the Anthropic endpoint. On desktop, open Models in Settings, choose Custom and enter the URL, API key and model name; the international edition keeps its config in ~/.workbuddy-ai/models.json.
Does QwenWork support a custom model?
As of 2026-11-11, the personal plans offer built-in models only: the desktop app and the web version both pick one of Advanced, Standard or Qwen3.8-Max. The enterprise flagship has BYOK (bring your own key), and the official text says the BYOK custom-model capability is limited to customers whose BYOK entitlement is turned on, and to contact sales to enable it. The personal column therefore says Not in the official docs, with a note that the enterprise flagship has BYOK.
Which office agents support both the Anthropic and OpenAI protocols?
As of 2026-11-11, six: ZCode, DeepSeek Harness, Qoder (new desktop app and CLI), Kimi Code, MiniMax Code and LobsterAI. Where to set them: ZCode under Model Settings → Add Provider; DeepSeek Harness under Settings → Models → Add model provider → Custom model API; Qoder under Settings → Models → Add model → Custom; Kimi Code in the CLI file ~/.kimi-code/config.toml; MiniMax Code under Settings → Usage & model; LobsterAI under Settings → Custom Model → Add Custom.
Where do the ranking figures come from?
MAU is cited from the Ifenxi 2026 China AI Office Ranking (October). It is the first edition, covering October 2026; MAU counts desktop clients, web versions and related aggregated entry points, in units of 10,000 users (wan), and monthly active users and user mentions each make up 50% of the score. The English table converts wan to millions, for example 179 wan = 1.79M. The API key column does not come from the ranking: it follows each product's official docs, checked on 2026-11-11.
How many of these can one QCode key run at once?
All 8 products marked Yes can use the same key, at the same time: the Anthropic and OpenAI addresses share one cr_ key, and all usage is recorded on that key. Claude models are available only on the Anthropic address and GPT models only on the OpenAI address. Usage is billed per token with per-model rates on /models.
Sources
Ranking: Ifenxi 2026 China AI Office Ranking (October); the methodology notes come from the EE Times China repost of 2026-11-10. Official product material, checked on 2026-11-10 and 2026-11-11: the Doubao Work help center (work task mode, product introduction, team and enterprise quota rules), the QwenWork pages in the Alibaba Cloud help center (model selection, enterprise flagship BYOK), the WorkBuddy and CodeBuddy docs, the ZCode configuration docs, the model guide in the DeepSeek Harness repository, the Qoder custom models docs and release notes, the Kimi Code and MiniMax Code docs, the AutoClaw model settings page and official blog, the LobsterAI README and changelog, the Lingxi Pro help docs, the EasyClaw FAQ, the AiPy manual and FAQ, the Nami Work site and user agreement, the Kimi Work help center, the Skywork site, the Xiaohuanxiong pricing page, the Marvis site and docs, the TeleAgent product page, the DuMate help docs and the Kuku app store listing. Hands-on tests: 2026-11-10 in an isolated environment. QCode endpoints and Pay-as-you-go wording come from docs.qcode.cc and qcode.cc/pay-as-you-go.
Connect all 8 products marked Yes with one QCode key
After you register, copy your API key (it starts with cr_) and enter it with the protocol named in the table: Anthropic for Claude, OpenAI for GPT, either for Chinese models; billed per token, with per-model rates on /models.
Related reading
Pay-as-you-go Packs
One-time quota packs valid for 7 days, no subscription; extra packs add to the same key
ZCode custom provider: how to use your own model API (2026-11)
Using your own API key in ZCode: desktop Add Provider and CLI provider_config.json, whether each base URL needs /v1, and the CLI 0.17.1 test results.
Qoder BYOK: how to set a custom model Base URL and API key (2026-11)
Qoder can use your own API key: on the desktop app pick OpenAI Compatible or Anthropic Compatible under Settings > Models > Add model, or use the Custom wizard in /model in the CLI; with the QCode values.
Crypto payment guide
Nine coins, 56 networks, zero platform fee for AI coding API credits.
Rank and MAU in the table are from the Ifenxi 2026 China AI Office Ranking (October), as published; the custom-model column follows each product's public official material, checked on 2026-11-11, and later official docs take precedence. Menu names follow the version you install, and QCode model IDs follow /models. QCode.cc is not affiliated with any vendor in the table.