What is Accio? Alibaba's SI agent team for SMEs, Accio Work

What is Accio? Today the answer is Accio Work: a plug-and-play SI agent team built by Alibaba for SMEs and small teams that carries business tasks through to the end. You do not just get suggestions in a chat window. The agents write files, drive your browser, search Alibaba.com for suppliers, collect quotes and drop the result into your working folder. This guide covers what Accio Work does, how it works, how plugins, connectors and skills differ, how permissions work, what it costs and where it stops, all checked against the official documentation. The last section is how we read the product at Webtures: what changes for e-commerce sellers and exporters who source or sell on Alibaba.com, and where an agent can and cannot be trusted.

Product details in this article were checked against the official documentation and release notes on accio.com on 20 October 2026. The product moves fast; confirm the current in-app screen before any critical decision.

What is Accio?

Accio is the name of Alibaba International's SI product family for SMEs. The brand first appeared in late 2024 as a B2B sourcing search engine; according to Alibaba International it reached 1 million users about five months after launch. Accio Work is now the main product on accio.com, and the official documentation states that credits are shared between the two: credits spent on Accio web and in Accio Work come out of the same pool.

The official one-line definition calls Accio Work "a plug-and-play AI Agent team that automates end-to-end business tasks for SMEs". It is not a single chatbot. It is a digital workforce made of agents, each defined by a role, a memory, a tool set and a working folder. The target audience is spelled out:

• SMEs and small teams that want to run operations that used to need a large team.
• One-person companies: the documentation explicitly anticipates a wave of them.
• Enterprises: a separate Business edition exists for company needs.

Alibaba frames the difference through three comparisons. Against general chatbots such as ChatGPT and Gemini, Accio Work executes instead of only suggesting. Against agent platforms such as Claude Cowork and OpenClaw, it works out of the box, with preloaded skills, preconfigured agents and one-click platform connections. Against specialised SaaS and automation tools, it claims one subscription covers strategy, execution and tooling. The company grounds this in Alibaba's B2B data; in its own words, 26 years of B2B intelligence and a dataset of more than one million merchants.

What does Accio Work do?

Accio Work takes a task written in plain language and uses tools to finish it. Which tools an agent gets depends on the template chosen when it is created. The documentation defines three presets: Standard (search, file operations, web retrieval, image processing, memory), Full (Standard plus running commands, browser automation and scheduled tasks) and a read-only Minimal preset. The Coder and Shopify Dropshipper templates ship with Full. Two documentation pages disagree on which preset Daily Assistant uses, so check the tool list on screen when you create the agent.

On the sourcing side the headline agent is Sourcing Expert. It works with the Accio Sourcing Toolkit plugin and, according to the documentation, automates supplier matching, side-by-side comparison, bulk outreach, negotiation and order placement. The sample prompts compare suppliers on price, MOQ, product parameters and supplier performance. Give it product, quantity, destination, deadline and target price, and it returns matches from verified suppliers; once you pick targets and authorise outreach, it negotiates on your behalf.

On the store side the official FAQ is direct: native Alibaba International data support is still being integrated. For now the recommended route is to ask the agent to open the browser, log into the store backend and work there. Taobao and Tmall get the same answer. Mentioned in the release notes: version 0.35.0 of 19 October 2026 added opening the 1688 merchant app from the home page, Shopify product downloads, and Taobao and Tmall product sourcing and rankings. Earlier releases announced e-commerce templates for Amazon, Etsy, Shopify and TikTok, and per-store TikTok Ads account authorisation.

How does Accio Work work?

Every agent in Accio Work is defined by files that live in its own folder on disk. Personality and behavioural boundaries sit in SOUL.md, role and communication style in IDENTITY.md, long-term memory in MEMORY.md and your profile in USER.md. When context is assembled, the memory, soul and user files are placed last, closest to the conversation. The agent can rewrite its own SOUL.md; the documentation encourages this and says the agent tells you when it does.

Memory runs in four layers, and that is usually where the answer to "why did it forget me?" sits:

1. Local memory index: an SQLite database on your device. It is disabled by default and has to be switched on in the agent settings. Once on, interaction fragments become semantically searchable.
2. Diary: the agent's day-by-day working notes, searched when needed.
3. MEMORY.md: distilled knowledge such as preferences, agreements and decisions, loaded at the start of every session.
4. Identity files: SOUL.md and USER.md are part of the context directly.

The /dream command tidies these layers: it promotes frequently used entries into the diary, distils the diary into long-term files and prunes stale content. A similar pass runs in the background on the first message you send an agent each calendar day, and the documentation says its credit use is very low.

Multiple agents work in two different models. An Agent Group is a persistent team chat: a Team Lead analyses the task and assigns work, and up to 9 members execute in parallel according to their roles. The documentation suggests 2-3 members for focused work and 4-5 for a full workflow, and advises against going above 7. A SubAgent is temporary delegation: the main agent starts a background helper of type explore, bash, browser or general. SubAgents cannot start other SubAgents, and nesting stops at two levels. The documented sourcing team is a Team Lead acting as Procurement Manager, plus a Supplier Researcher and a Market Analyst.

Scheduled tasks run in three modes: delayed ("in 30 minutes"), timed ("tomorrow at 9:00") and recurring ("every 2 hours"). You set them up by telling the agent in chat or from the Automations page, and results land in the related conversation with a green tag.

What is the difference between a plugin, a connector, a skill and a channel?

Setting Accio Work up properly starts with separating four concepts that are easy to mix up. The documentation's own distinction: channels are where you chat with your agent, connectors are the tools your agent uses to get work done. The agent is the doer; a plugin is the ready-made toolbox you hand it.

Skills work in three layers: agent-private skills have the highest priority, account skills are shared across all your agents, and built-in skills that come with a template have the lowest. A private skill with the same name overrides the account-level version. You can load your own skill as a folder containing SKILL.md through "Install Local Skill", and typing / in a chat forces a specific skill for that turn. The documentation's warning matters: loading many unrelated skills onto one agent lowers call accuracy and burns credits, so splitting work across role-based agents gives better results.

Read the connector list carefully. Google Drive, Slack, a Notion connector, Shopify, HubSpot, Salesforce, Figma and Canva are under "coming soon". Notion is usable today as a plugin. On channels, WeChat supports one-to-one chat only, and every new outside contact first creates a pairing request that waits for your approval.

How do security and permissions work in Accio Work?

Accio Work's security model rests on every sensitive action passing a permission check before it runs. Each operation is governed by one of four policies: allow (runs directly), ask (asks every time), ask_once (runs automatically after the first approval) and deny (never runs). By default, file reads and network access are allowed, while file writes, shell commands and browser use need confirmation. Templates can change that default: according to the documentation, file writes are auto-approved in the Daily Assistant template.

When approval is needed, a card appears with four options: Allow Once, Allow This Session, Always Allow and Deny. The system decides in a fixed order. A built-in safety rule rejects dangerous commands such as rm -rf / outright, custom rules are matched top to bottom and the first match wins, then the default policy applies, and if nothing matches the user is asked. Every decision is written to permissions/audit.jsonl in the agent's private space.

• Workspace boundary: the agent cannot leave the authorised folder; a path allowlist, symlink protection and boundary checks work together.
• Blocked directories: you can write deny rules for specific tools and path prefixes.
• Sensitive external actions such as payments: require explicit user approval, according to the documentation.
• Connector access: starts only after explicit authorisation and can be revoked at any time from Accio Work or from the third-party platform.

Browser automation is the most powerful capability and the most sensitive one. The agent uses your real Chrome, with your existing logins. There are two ways to connect: the recommended Accio Browser Relay extension (loaded unpacked in Chrome developer mode, works in all Chromium browsers) and a beta direct CDP connection for Chrome M144 and later. Communication runs only over 127.0.0.1, the connection is authenticated with a random token, controlled tabs show a notification bar and tabs the agent opened close when the task ends. The documentation sells the benefit as "no re-login, real device fingerprint, your residential IP". Keep one thing in mind when you read that line: to the platform, every click the agent makes is a click you made.

How is Accio Work priced?

Accio Work runs on credits. The Free plan gives 20 base credits every 24 hours; the timer starts at first use and unused credits do not roll over. New users also get a one-time bonus of 500 credits valid for 7 days. The Free plan has no access to higher-quality models. Paid plans are billed monthly, with 15% off on annual billing.

Four factors decide how fast credits go: model choice, task complexity and context length, the number of tool calls and parallel tasks. The documentation's model table lists credit multipliers from 0.04x to 3.5x, and shows Auto's credit cost as "Free". Only paid subscribers can buy top-up packs: Small (2,000 credits, $14.99), Standard (5,000 credits, $34.99) and Extended (15,000 credits, $89.99). Payment is by bank card or Alipay. A refund can be requested if no credits were used within 7 days of subscribing, minus third-party processing fees.

Prices, credit amounts and model multipliers were checked on the Billing and Credits pages of accio.com on 20 October 2026.

What are Accio Work's limits?

Accio Work's marketing and its documentation do not always line up. The agents page presents the product as a "24/7 digital workforce"; the scheduled tasks page says plainly that tasks do not run while your computer is off. When the app restarts, short delayed tasks that were missed run immediately and recurring tasks resume from the current time. Always-on cloud execution is not described anywhere in the documentation.

Sign-in is by email or an Alibaba International account. The official route to mobile access is connecting the agent to a messaging channel.

Who is Accio Work for?

Accio Work makes the most sense for businesses that want to scale repetitive, rule-writable operations with a small team. It is clearly built for people who source regularly on Alibaba.com, sellers running stores on several marketplaces and small teams with a heavy reporting load.

• Good fit: e-commerce sellers sourcing from Alibaba.com, teams that want competitor price and trend tracking turned into a weekly report, operations teams turning quote and invoice files into tables, one-person brands.
• Proceed with care: teams handling customer personal data, anyone thinking of handing order and payment authority to an agent, sellers whose marketplace account health is critical.
• Too early: organisations expecting server-side, always-on automation, full store analytics through native connectors, or a Turkish interface.

The documentation recommends Daily Assistant as the most versatile starting template. A concrete role description changes the output directly: "handles supplier stuff" is weak, "supplier research focused on MOQ, certifications and lead time" is strong.

How we read it: what Accio means for exporters and e-commerce sellers

We read Accio Work less as a productivity tool and more as an agent stepping into the buyer's side of B2B trade. For 16 years we have managed search and SI visibility for brands from Istanbul and London, and the work has shifted from what people search for to what agents compare. Accio Work is a concrete example of that shift inside the supply chain. Webtures has no partnership with Accio; this assessment rests on the product's public documentation and our own agentic commerce work.

For suppliers selling on Alibaba.com, including Türkiye's exporters: the documented sample prompts compare suppliers on price, MOQ, product parameters and supplier performance, and the recommended sourcing role focuses on MOQ, certifications and lead time. The agent builds its shortlist from those fields and sends bulk outreach to that list. A manufacturer whose MOQ is vague, whose certificates cannot be read or whose lead time is missing ends up as an empty cell in the agent's table. More and more incoming inquiries will be written by an agent and negotiated by an agent. Keeping product data structured, complete and consistent is no longer catalogue hygiene; it is the condition for making the shortlist. We measure this layer in our Agentic Commerce Readiness work.

For e-commerce sellers sourcing from Alibaba: time-consuming work such as supplier scanning and quote comparison can be handed to the agent; the decision should not be. The quote table the agent brings back is a starting point. Samples, quality control, customs and logistics cost and contract terms stay on a human desk. For sellers expanding abroad, we handle this flow together with marketplace, logistics and visibility strategy in our E-Export Consulting practice.

We answer "where can the agent be trusted?" with three tiers:

1. Let it run: public web research, trend and competitor price tracking, reading files into tables, drafting reports. Low cost of error, easy to undo.
2. Run with approval: bulk messages to suppliers, negotiation rounds, file writes, listing or price changes in a store backend. "Allow This Session" is enough here; do not grant "Always Allow".
3. Keep human: orders and payments, accepting contracts, exporting customer personal data, marketplace account settings.

With browser automation the biggest risk is the authority itself. The agent works with your session, your IP address and your fingerprint, and its mistakes are recorded against your account. Marketplaces have different rules on automated access, and this year saw agents blocked for not identifying themselves; we covered it in sell-side agentic commerce. Our advice is to connect the browser through a separate Chrome profile, keep only the accounts the agent needs logged in there, and review permission decisions in audit.jsonl every week.

On data sovereignty, two statements need to be read together. The documentation says data stays local by default, and agent files and the memory database do live on the device. At the same time, the model table lists many models from different providers, and task context is sent to the selected model to be processed. For companies handling personal data, that can raise cross-border transfer questions under GDPR, KVKK or similar regimes. Before running tasks with customer data, review the Business edition contract and data processing terms with your legal team.

Finally, Accio Work shows both sides of agent readiness at once. One side is how ready you are to hand your own operations to agents: role definitions, a permission ladder, an audit trail, measurement. We build that side in SI Workflow Engineering. The other side is how well other people's agents can read you. You can see whether agents can read and compare your site and product data with an SI Agent Readiness scan.

Frequently asked questions

Are Accio and Accio Work the same thing?

They are two products in the same brand family. The Accio name appeared in late 2024 with Alibaba International's B2B sourcing search engine; Accio Work is the brand's agent team product that executes tasks. According to the official documentation, Accio web and Accio Work credits come from a shared pool.

Is Accio Work free?

There is a Free plan: 20 base credits every 24 hours plus a 500-credit bonus valid for 7 days for new users. The Free plan cannot use higher-quality models. As of 20 October 2026, paid plans cost $19.9 (Pro), $99 (Elite) and $199 (Ultra) per month.

How is Accio Work different from ChatGPT?

General chatbots such as ChatGPT mostly produce suggestions and text. Accio Work executes through agents with roles, memory and tools: it writes files, acts in the browser, messages suppliers on Alibaba.com and runs scheduled tasks. Sensitive steps pass through a permission card.

Can Accio Work place orders on Alibaba.com?

According to the documentation, the Accio Sourcing Toolkit runs from supplier matching through negotiation to order placement. Sensitive actions such as payments require explicit user approval. We recommend not leaving order and payment decisions to the agent.

Do scheduled tasks run when my computer is off?

No. The official documentation says plainly that scheduled tasks do not run while the computer is off. When Accio Work restarts, missed tasks are checked and recurring tasks resume from the current time.

Does Accio Work connect directly to my store data?

Native data integration for Alibaba International, Taobao and Tmall is still in progress. Today the agent works through the browser in your store backend. The documentation recommends reading results from the real page and having key figures verified item by item.

Where is my data stored?

Agent files, diaries and the memory database are kept in the workspace on your device, and the agent cannot leave that workspace. Task context is sent to the selected model to be processed. Work involving personal data needs a separate review of contract and data processing terms.

Sources

All official sources were checked on 20 October 2026.

• Accio Work: What is Accio Work
• Accio Work: Agents and Agent Tools Guide
• Agent Core and Agent Core Design
• Agent Memory and Agent Memory Guide
• Agent Group Guide
• Plugins, Skills and Skills Guide
• Connectors Guide and IM Channels
• Permissions & Security and Permissions Guide
• Browser Use Guide
• Automation Tasks and Scheduled Tasks
• Billing & Subscription and Credits & Limits
• Install & Updates, Login & Account and Common Operations
• Automated Sourcing
• Accio Work Release Notes (0.35.0, 19 October 2026)
• Digital Commerce 360, 11 March 2025 (Accio search engine user count, company statement)