AutoGLM

Short answer. AutoGLM is Zhipu AI’s open-source phone-use agent: a vision-language model (AutoGLM-Phone-9B, from the GLM-4.1V-9B family) reads the phone screen, plans a chain-of-thought action sequence, and executes it over ADB (Android), HDC (HarmonyOS NEXT) or WebDriverAgent (iOS). It was the first phone agent with true Phone Use capabilities (2024-10-25) and is research/learning-licensed.

Key facts.

• Ships as a Python framework (PhoneAgent API + CLI) built around AutoGLM-Phone-9B; the commercial AutoGLM 2.0 product runs the same approach on cloud virtual phones.
• Execution spans ADB (Android), HDC (HarmonyOS NEXT) and WebDriverAgent (iOS), with remote ADB over Wi-Fi/network for flexible control.
• Free framework under Apache-2.0 (code) and MIT (models); the hosted autoglm-phone API on BigModel was limited-time free as of 2026-09-20.
• Auto-detects sensitive screens (payment, password, banking) and requests human takeover; login and CAPTCHA scenarios support manual takeover.
• Local deployment needs roughly 24GB+ VRAM; Android use requires developer mode + USB debugging (plus ADB Keyboard for text input).

What this means. AutoGLM makes phone-level computer use an open, inspectable capability in China’s ecosystem — the same GUI-operation primitive that ByteDance’s Doubao “Work” mode and Butterfly Effect’s Manus productize, but shipped as a model-plus-framework anyone can run locally or license for research.

What is uncertain. The paid price of the AutoGLM-Phone API after its promotional period is not publicly disclosed, and no persistent-memory feature is documented in the README. Independent evaluation of its task-completion reliability across the three OS backends (ADB/HDC/WebDriverAgent) is not recorded in this database.

Sources.

evidence_id	source_name	source_url	source_type	published	verified	confidence	conflict
src-agents-autoglm-1	Open-AutoGLM GitHub repository	https://github.com/zai-org/Open-AutoGLM	Official documentation	—	2026-09-20	high	—
src-agents-autoglm-2	AutoGLM Goes Open Source blog	https://autoglm.z.ai/blog	Official	—	2026-09-20	high	—
src-agents-autoglm-3	AutoGLM-Phone model card (Hugging Face)	https://huggingface.co/zai-org/AutoGLM-Phone-9B	Model card	—	2026-09-20	high	—
src-agents-autoglm-4	BigModel AutoGLM-Phone API docs	https://docs.bigmodel.cn/cn/guide/models/vlm/autoglm-phone.md	Official documentation	—	2026-09-20	high	—

Why it matters

AutoGLM is the purest demonstration of computer use as a model capability rather than a wrapper product. Its entire function — reading a screen, planning an action sequence, tapping — is downstream of the AutoGLM-Phone-9B VLM’s grounding and planning quality; there is no separate tool registry or orchestration layer doing the work. That makes it the reference case for how GUI automation depends on vision-model grounding, a relationship the database captures by binding the agent to Zhipu AI while the GLM-4.1V-9B lineage itself sits outside the tracked flagship models.

China AI Hub analysis indicates AutoGLM matters less as a production tool (it is research/learning-licensed, with sensitive-screen handoff by design) and more as the clearest evidence that phone-level computer use is technically mature in China’s open ecosystem — a capability ByteDance’s Doubao and Butterfly Effect’s Manus keep behind closed surfaces. The open release lowers the cost for any team to study or build GUI agents on the GLM VLM stack.

How it differs from Manus

AutoGLM and Manus both “operate a screen,” but they are opposite bets on the same primitive.

• Manus is a closed, cloud-only general agent — a subscription product that does whole tasks (slides, websites, research, email/Slack) across a browser and files, with its underlying model orchestration undisclosed.
• AutoGLM is an open, model-specific phone agent — a framework and 9B VLM you can run locally or license, scoped to phone GUI operation across three mobile OS backends, research/learning-only.

China AI Hub analysis: the difference is openness-versus-product and phone-versus-browser. Manus trades transparency for a polished, general, subscription surface; AutoGLM trades polish and generality for an open, inspectable, phone-specific substrate. A team wanting a deployable general assistant buys Manus; a team wanting to study or build phone automation on an open VLM starts from AutoGLM.

Practical implications

For researchers. AutoGLM is the reference open implementation of a phone-use agent — screen perception → chain-of-thought planning → action execution over ADB/HDC/WebDriverAgent, with remote ADB for Wi-Fi/network control. The setup costs are ~24GB+ VRAM locally plus per-OS backend tooling (developer mode, USB debugging, ADB Keyboard; a separate WebDriverAgent setup for iOS).

For product teams. The commercial AutoGLM 2.0 runs the same approach on cloud virtual phones, so a team can validate phone-agent value without managing physical devices; the hosted autoglm-phone API was free during its promotional period as of 2026-09-20.

For evaluators. Sensitive-screen auto-detection and human takeover are a safety feature, not a capability gap — but they also mean AutoGLM’s autonomy is deliberately bounded around payment/password/banking surfaces, so end-to-end “finish a purchase” tasks stop at the handoff rather than completing unattended.

What the evidence shows

The evidence is strong on what AutoGLM is and thin on how well it generalizes. The GitHub repo and blog document the architecture (VLM screen perception + planning + ADB/HDC/WebDriverAgent execution), the first-Phone-Use claim (2024-10-25), the sensitive-screen handoff, and the research/learning license; the Hugging Face model card confirms AutoGLM-Phone-9B’s GLM-4.1V-9B lineage. The repo has also expanded past the core framework — a WeChat community, a Zhipu AI input-method (Autotyper) product and a Midscene.js integration — showing the phone-agent surface growing beyond the README’s original scope.

What is missing is independent reliability data. No third-party benchmark of AutoGLM’s task-completion rate across the three OS backends is recorded here, and no persistent-memory feature is documented — a boundary the README leaves explicit. China AI Hub analysis indicates the honest reading is “capability demonstrated and open,” not “reliability independently proven”: the availability evidence (repo, model card, API) precedes any reliability evidence, and buyers should treat the sensitive-screen handoff as a fixed constraint rather than a setting they can turn off.

Where this fits

Workload	Relevance
Phone GUI automation research (screen → plan → action)	High
Cloud-phone batch operations (AutoGLM 2.0)	High
Local phone-agent deployment (ADB/HDC/WebDriverAgent)	High
Payment/password/banking flows	Low (auto human-takeover by design)
General-purpose, non-phone task execution	Low (phone-scoped, research-licensed)
Independent reliability evaluation	No evidence recorded

Relevance judgments are China AI Hub analysis based on documented capabilities, not vendor claims.

Field reference

Field	Value	Evidence type
Underlying model	AutoGLM-Phone-9B (GLM-4.1V-9B family)	Official
Target users	Researchers and developers (research/learning use)	Official
Platform	Python framework (PhoneAgent API + CLI); AutoGLM 2.0 cloud product	Official
OS	Android (ADB), HarmonyOS NEXT (HDC), iOS (WebDriverAgent)	Official
Browser / computer use	Phone GUI computer use	Vendor-reported
Coding	Not publicly documented	Not publicly documented
Autonomous task execution	Yes (chain-of-thought planning)	Vendor-reported
MCP	Not publicly documented	Not publicly documented
Tool calling	Yes	Official
Memory	No	Official
Workflow	Screen reading → chain-of-thought planning → action execution	Vendor-reported
API	Yes (BigModel autoglm-phone)	Official
Pricing	Framework free; API limited-time free as of 2026-09-20	Vendor-reported
Region	China (BigModel)	Official
Open-source	Yes — Apache-2.0 (code), MIT (models)	Official
Deployment	Cloud and self-hosted	Official
Limitations	Research/learning only; ~24GB+ VRAM for local; sensitive screens need human takeover	Official
Source	Open-AutoGLM GitHub	Official
Last verified	2026-09-20	Official

See the Zhipu AI company profile, the Manus general agent, the choosing-an-agent guide, and the site’s computer use and AI agents technology pages.

Labels used above: Official fact (from the Open-AutoGLM GitHub repo, AutoGLM blog and model card), Vendor-reported claim (capability and pricing statements by Zhipu AI), and China AI Hub analysis (our synthesis, always introduced as such). No independent evaluation of AutoGLM’s phone-task reliability is currently recorded.