Choosing a Chinese AI Agent: Underlying Models, Tool Calling and MCP

Short answer. The ten agents tracked in the database are mostly distribution channels for their vendor’s own models: seven are bound to a single lab’s models, and only two are genuinely cross-model — Qoder (closed, smart-routes across Qwen, DeepSeek, GLM, Kimi and MiniMax) and Qwen Code (open, multi-protocol across OpenAI, Anthropic, Gemini, Qwen and others). So the first question is not “which agent” but “am I buying a tool or a model contract.” If you need model freedom, pick a cross-model agent; if you have already standardized on one lab’s model, its native agent is usually the tightest integration.

Decision criteria

Criteria	Relevance / Notes
Model binding	7/10 bound to vendor models; only Qoder and Qwen Code are cross-model; DeepSeek Harness is neutral by architecture
Tool calling	Present on all ten agents as a documented capability
MCP support	`mcp: true` on 7/10 — every open developer agent; absent on Doubao, MiniMax Agent, AutoGLM
Open source	6/10 open — but 5 of those 6 are still model-bound; only Qwen Code is open and cross-model
Deployment	Self-hosted (DeepSeek Harness, Qwen-Agent, Qwen Code) vs cloud-only (Doubao, GLM Coding Plan, MiniMax Agent) vs both
Computer use / browser	Desktop-capable agents (Kimi Code, MiniMax Code, Qoder, AutoGLM) vs CLI/API-only
Underlying model ceiling	A bound agent’s ceiling is its vendor’s model — Kimi Code’s 1M output is K3’s; GLM Coding Plan’s 1M context is GLM-5.3’s

The ten agents at a glance

Agent	Vendor	Underlying model(s)	Binding	Open	MCP
AutoGLM	Zhipu	AutoGLM-Phone-9B (GLM-4.1V-9B)	Bound	Yes	No
DeepSeek Harness	DeepSeek	V4.1-Flash, V4-Pro	Neutral (configurable)	Yes	Yes
Doubao	ByteDance	Not publicly documented	Unspecified	No	No
GLM Coding Plan	Zhipu	GLM-5.3, GLM-5.3-Flash	Bound	No	Yes
Kimi Code	Moonshot	K3, K2.7-Code, K2.7-Highspeed	Bound	Yes	Yes
MiniMax Agent	MiniMax	M3, M2.7	Bound	No	No
MiniMax Code	MiniMax	M3, M2.7, M2.7-Highspeed	Bound	Yes	Yes
Qoder	Alibaba	Qwen, DeepSeek, GLM, Kimi, MiniMax	Neutral (routing)	No	Yes
Qwen-Agent	Alibaba	Qwen>=3.0	Bound	Yes	Yes
Qwen Code	Alibaba	Multi-protocol + local	Neutral	Yes	Yes

Entity routing

Route by what you actually need the agent to be.

Scenario	Best-documented fit	Why
Cross-vendor routing, closed product	Qoder	Smart-routes tasks by credit tier across five vendors
Cross-protocol, open framework	Qwen Code	OpenAI/Anthropic/Gemini/Qwen + DeepSeek/MiniMax/Z.AI/Kimi/local
Configurable-provider open runtime	DeepSeek Harness	“Everything is a Plugin,” MCP, self-hosted
Whole-repo coding bound to Moonshot	Kimi Code	Open, MCP, browser/computer-use, 1M context via K3
Subscription coding bound to Zhipu	GLM Coding Plan	GLM-5.3/5.3-Flash, quota-based, powers 20+ tools
Open desktop coding bound to MiniMax	MiniMax Code	M3/M2.7, MCP, memory, macOS/Windows desktop
Phone-use autonomous agent	AutoGLM	AutoGLM-Phone-9B, computer use, ~24GB+ VRAM local
Consumer/closed surface	Doubao	Unnamed chat model, cloud-only, no consumer API
Cloud assistant ecosystem	MiniMax Agent	M3/M2.7, 24/7 assistant, no self-host
Python framework bound to Qwen	Qwen-Agent	Qwen>=3.0, RAG, MCP, self-hosted

What the evidence shows

The dependency map is lopsided in a way that matters. Seven of ten agents route usage to their lab’s own models, and openness does not imply model freedom — five of the six open-source agents are still model-bound. The only open-source and cross-model agent is Qwen Code; the only closed cross-model agent is Qoder, and both are Alibaba products. China AI Hub analysis indicates that concentration is structural, not accidental: a genuinely neutral agent lets the buyer swap the model under the tool, which threatens every model vendor — so the one lab with a broad enough platform (Model Studio hosts competitors’ models) is the only one positioned to sell neutrality.

MCP support is near-universal on the open agents (7/10) but absent on the three closed/consumer surfaces — the tool-interoperability standard is an open-agent phenomenon, which makes sense because MCP is how a neutral agent reaches arbitrary tools, and a bound agent has less need for it. Deployment splits along the same self-hosted-open / cloud-closed line as the model layer: the do-it-yourself surface is open, the managed surface is closed, and where a vendor sits in the stack is highly predictive of how its agent is priced and locked — which in turn predicts how dependent the buyer becomes.

A second structural finding: model-bound agents track their vendor’s capability ceiling. Kimi Code is bound to K3/K2.7-Code, so its long-horizon whole-repo strength is exactly K3’s 1M-token output. GLM Coding Plan is bound to GLM-5.3/5.3-Flash, so its 1M-token context claim is GLM-5.3’s context. Within a bound agent there is no fallback to a stronger competitor model — the agent’s ceiling is the vendor’s model ceiling.

Selection procedure

Work through these steps in order.

Decide model freedom versus integration depth. If you need to swap models under the tool, the shortlist is two entries — Qoder (closed, cross-vendor routing) and Qwen Code (open, multi-protocol). If you are already standardized on one lab’s models, that lab’s native agent is the tightest integration.
Check whether you can self-host. Three agents are self-hosted frameworks (DeepSeek Harness, Qwen-Agent, Qwen Code); three are cloud-only (Doubao, GLM Coding Plan, MiniMax Agent); four support both. If you cannot use a managed cloud endpoint, that cuts the field in half.
Confirm the capability surface you actually need. MCP is near-universal on the open agents but absent on the closed ones; computer use and browser control are desktop-capabilities (Kimi Code, MiniMax Code, Qoder, AutoGLM). Match the flag to your workload, not to the vendor’s pitch.
Trace the underlying model ceiling. A bound agent’s ceiling is its vendor’s model: Kimi Code’s 1M output is K3’s, GLM Coding Plan’s 1M context is GLM-5.3’s. If your job needs a capability only a different vendor’s model has, a bound agent cannot give it to you.
Read the pricing and lock-in before adopting. Subscription plans carry quota caps, peak-hour multipliers and expiry rules that are separate from pay-as-you-go API pricing, and they are not fully extractable from sign-in-gated pages.

Limitations

Agent products have no independent evaluation data in the database — every capability here is a vendor-listed capability flag, not a measured result. “Underlying model” is recorded as documented by the vendor; where a field is not publicly stated (Doubao’s chat model, several first-release dates) the database records the absence rather than inferring. Model-binding classification is a description of documented fields, not a performance ranking. MCP support being absent on a closed agent means the vendor does not list it, not that it is impossible. DeepSeek Harness is a developer preview — its own SAFETY.md notes it is not security-audited or production-ready.

Sources

Labels used above: Official fact (framework, license, MCP and deployment fields from primary sources and repositories), Vendor-reported claim (underlying-model and capability statements published by the vendor), and China AI Hub analysis (our synthesis, introduced as such). No third-party evaluation evidence is currently recorded for any of the ten agents.