Coze is ByteDance's AI team collaboration platform where people and multiple agents work on shared projects. It hosts native agents alongside Claude Code, Codex CLI, OpenClaw and Hermes, with project-level files, long-term memory and credit-based billing.
Used this tool? Rate it
Used this tool? Rate it
Coze — known as 扣子 in Chinese — is ByteDance's platform for working with AI agents as a team rather than one assistant at a time. Coze is a ByteDance product, from the same parent company as TikTok, and reporting on its Coze Space workspace noted that it is built on ByteDance's in-house Doubao large language model and supports MCP, integrating with Lark and Amap. Coze positions itself as a next-generation AI team collaboration platform for the agent era, where people and agents work together like a real team, and that framing is a deliberate departure from where the product started. Coze launched in 2024 as a no-code bot builder, and the version described here is a substantially different product: it is not just a chat tool but a platform that carries projects, accumulates assets, dispatches agents and keeps delivering results, with each project holding its own conversations, files and assets.
An important orientation point before going further. There are two Coze properties: the domestic Chinese version at coze.cn and the international version at coze.com. This page describes the domestic version, which is the one this listing points to. The two differ in pricing, model availability and where data is stored, so figures here should not be assumed to transfer to the international product. The two also carry different emphases in their own marketing — the homepage still describes an all-in-one AI office assistant covering AI writing, PPT generation, spreadsheets, design, podcasts, image and video generation, while the documentation leads with team collaboration. Both are accurate; they describe the same platform from different angles.
The practical unit of work is the project. Inside a project you hold conversations, store files, assign work to different agents, and accumulate the output over time. What makes this more than a group chat with a bot is that agents in Coze carry long-term memory, can be given skills, and can act on real resources — cloud machines, local files, calendars and mailboxes — rather than only producing text.
Coze uses credit billing across three plan families: personal, team and enterprise. Billing runs on credits as the single settlement unit, and credits cannot be converted back into cash. Personal plans range from a free tier through several paid tiers, team plans target small and growing teams with member management and per-employee usage limits, and enterprise plans add organisational controls. The free personal tier grants activity credits on daily login, while the enterprise flagship tier adds SSO and VPC private network connectivity.
Two settlement mechanics deserve attention because they behave very differently. Personal and team plans do not allow overage: when credits hit zero the services stop until you buy more or the next cycle resets. Enterprise billing settles through Volcano Engine, deducting cash automatically at a 1000:1 credit-to-cash ratio so that business operations continue uninterrupted — and a Volcano Engine balance cannot be applied to personal plans. In other words, individuals face a hard stop while enterprises face a continuing charge; neither is better in the abstract, but they demand different monitoring.
Also check what your workload actually consumes. Credits cover most platform resources — Coze tasks, cloud device time, coding tasks, model calls, plugins, knowledge base storage and so on — but Volcano Ark models, Volcano knowledge bases, paid skills, paid templates and third-party paid plugins cannot be paid for with credits and settle in cash instead. A workflow built around those resources will cost more than a credit-only estimate suggests. Specific prices are not reproduced here because tiers and credit allowances change; check the official pricing page for current figures.
Local agents are substantially less capable than their positioning suggests. By the official capability matrix, local agents lack calendar, mailbox, channels, skills, cloud phone, cloud computer, background tasks and model switching, keeping only conversation sharing, history search and collaboration. If you connect a local agent expecting the full workbench, most of the platform's distinguishing features will be unavailable to it. Choose the type based on this table rather than on the marketing.
Hard stops on personal and team plans. The no-overage rule is a real operational risk for long-running work. A task that exhausts credits mid-execution stops, and recovery means purchasing more credits or waiting for the cycle to reset. Teams running scheduled or background work should keep a buffer rather than budgeting to the last credit.
Some costs sit outside the credit system entirely. Because several categories of resource bill in cash regardless of credit balance, the headline plan price can understate what a real workflow costs. This is a documented behaviour, not a hidden fee, but it is easy to overlook when comparing plans.
Self-hosting the open-source version is not turnkey. The repository carries 489 open issues, 89 of which mention docker. The most-discussed threads concern deployment and integration rather than feature requests: failures wiring up local embedding models over the OpenAI protocol, NAS deployment errors, coze-server failing to reach MySQL after a docker compose deploy, sandbox errors in code nodes, and the frontend not running properly on Windows. Budget real infrastructure time if you go this route.
Feature parity within the open-source version is uneven. Users repeatedly report knowledge bases that answer correctly in one place but fail inside workflows, along with authentication errors on knowledge base write nodes. Validate your specific path end to end before committing.
The product has changed identity more than once, and documentation lags. Coze has moved from bot builder to office assistant to multi-agent collaboration platform across successive major versions. Third-party tutorials, and even some directory listings, frequently describe an earlier product. When you read a guide, check which version it targets — the open-source repository still describes itself as an all-in-one visual AI agent development platform, which is accurate for that codebase but not a description of the current hosted product.
Model availability is regionally shaped. The supported model list is centred on Chinese models, which is a strength for domestic users and a constraint if you specifically need frontier Western models. Notably, ByteDance first explored external models such as DeepSeek-R1 but found their tool invocation lacking, and after testing six Chinese LLMs settled on an in-house suite led by Doubao 1.5 Pro — a useful reminder that tool-calling reliability, not benchmark scores, is what determines whether an agent platform works.
Data residency and account requirements. The domestic service operates within China and enterprise settlement runs through Volcano Engine. International users should evaluate the coze.com product separately rather than assuming equivalence, and organisations with data residency obligations should confirm specifics before deploying.
Currently a collaboration platform. The official documentation describes an AI team where people and multiple agents work together on shared projects, and explicitly states it is no longer just a conversational tool. Older guides describing a no-code bot builder reflect the earlier product generation.
There is a free personal tier that grants activity credits on daily login, enough to evaluate the product. Sustained use requires a paid personal, team or enterprise plan, all billed in credits.
They are the domestic Chinese and international versions of the same brand. Pricing, model availability and data storage location differ between them. This page describes coze.cn; do not assume the figures apply to the international service.
Yes. Both are supported as agent frameworks, either as cloud agents running on Coze cloud machines or as local agents running on your own computer. Local agents can access your local files and code but lose access to most workbench features such as calendar, mailbox, skills and model switching.
On personal and team plans, services stop — there is no overage allowance. You buy more credits or wait for the next cycle. Enterprise accounts instead deduct cash automatically from a Volcano Engine balance at a 1000:1 ratio so work continues.
Doubao-Seed, Kimi, GLM and Minimax can be switched freely on supported agent types. Local agents do not support model switching.
Partly. Coze Studio and Coze Loop are on GitHub under Apache-2.0, which permits commercial use, and measured 21,488 and 5,699 stars respectively. The hosted coze.cn service itself is not open source — the open-source projects are a self-hostable agent development and observability stack, not a copy of the collaboration platform.
Local agents can, and desktop authorisation lets an agent work on local files. You control which folders are accessible, so scope permissions to the specific project directory rather than granting broad access.
Yes, agents can be published to those channels in one step for direct conversation, which is often more practical than asking colleagues to adopt another web app.
It is possible but not turnkey. With 489 open issues concentrated on docker deployment, external model integration and Windows compatibility, expect meaningful setup and troubleshooting effort. Teams without infrastructure capacity will generally get further with the hosted service.
