Kimi Code

Short answer. Kimi Code is Moonshot AI’s terminal AI coding agent — the successor to the deprecated kimi-cli — shipped as a TypeScript CLI (MIT), VS Code extension and Desktop app. It is the coding front-end for Kimi K3’s 1M-token output ceiling, turning long-context generation into long-horizon whole-repo work.

Key facts.

• A first-party client bound to Moonshot’s own models (Kimi K3, K2.7-Code, K2.7-Code-HighSpeed); the CLI works out of the box with Kimi models and can be configured for other compatible providers.
• Ships as CLI (TypeScript, MIT), VS Code extension and Desktop app; subagents, MCP, Kimi Computer Use, browser control and multimodal input (text, images, video) are built in.
• Billed under Kimi membership (Plus $15 / Pro $31 / Max $79 / Ultra $159 per month), all clients sharing one rolling 5-hour quota plus a monthly total.
• Open Platform API pay-as-you-go: kimi-k3 $3.00/M input / $15.00/M output; kimi-k2.7-code $0.95/$4.00; high-speed variant $1.90/$8.00.
• Cloud inference only — managed endpoints (api.kimi.com / api.kimi.ai coding/v1), not self-hostable as a model service.

What this means. Kimi Code is the agent layer that operationalizes Kimi K3’s distinguishing capacity: the database’s only 1M-token maximum output ceiling is exactly what a long-horizon coding agent needs to emit large multi-file edits and long reasoning traces in a single pass.

What is uncertain. There is no dedicated OS-level sandbox engine (approval gates only), and independent benchmark evidence of Kimi Code’s coding quality versus Qwen Code or GLM Coding Plan is not recorded in this database. The membership quota’s interaction with the API’s per-token rates is documented but not summarized in a single official comparison.

Sources.

evidence_id	source_name	source_url	source_type	published	verified	confidence	conflict
src-agents-kimi-code-1	Kimi Code GitHub repository	https://github.com/MoonshotAI/kimi-code	Official documentation	—	2026-09-20	high	—
src-agents-kimi-code-2	Kimi Code CLI documentation	https://moonshotai.github.io/kimi-code/en/	Official documentation	—	2026-09-20	high	—
src-agents-kimi-code-3	Kimi Code product docs	https://www.kimi.com/code/docs/en/	Official documentation	—	2026-09-20	high	—
src-agents-kimi-code-4	Kimi membership pricing	https://www.kimi.ai/membership/pricing	Official documentation	—	2026-09-20	high	—

Why it matters

Kimi Code is the agent manifestation of Kimi K3’s distinguishing capability. Where most coding agents are bounded by context length, Kimi Code is positioned around output length: K3’s 1M-token output ceiling is a structural fact of the model, and the agent is the surface that consumes it for whole-repo, multi-file, long-horizon work.

China AI Hub analysis indicates Kimi Code is best read as Moonshot’s bet that the coding-agent market will reward output-length and long-horizon reliability over raw price — a bet that runs directly on K3’s differentiated capacity. The membership-billing model (one quota across CLI, extension and desktop) reinforces this: Moonshot sells the workflow, not just the tokens, and prices the agent as a consumer-plus product rather than a pure per-token utility.

How it differs from Qwen Code and GLM Coding Plan

The three coding agents split on ownership, openness and monetization.

• Qwen Code is Alibaba’s open Apache-2.0 multi-protocol agent — a free CLI that routes to any provider (Qwen, DeepSeek, Kimi, MiniMax, OpenAI, Anthropic, Gemini, local). Model-agnostic by design; Alibaba monetizes access, not the tool.
• GLM Coding Plan is not a client — it is a subscription that injects GLM models into third-party tools (Claude Code, Codex, Cursor, OpenClaw). The model travels to your tool; the client is someone else’s.
• Kimi Code is a first-party open client bound to its own models — the tool is free (MIT), the models are Moonshot’s, and the two are designed together around K3’s long-output edge.

China AI Hub analysis: Kimi Code is the only one of the three that is simultaneously open-source and single-vendor, the mirror image of Qwen Code’s open multi-vendor stance. A developer who wants one open client and is indifferent to the model picks Qwen Code; a developer already committed to Kimi’s long-output models gets the tightest integration from Kimi Code. GLM Coding Plan is the option for those who won’t leave their existing client at all.

Practical implications

For long-horizon work. Kimi Code’s headline fit is whole-repo tasks where a single pass must emit large multi-file edits — the 1M-token output is the reason to choose it over a shorter-output agent. The cost is that this strength is only as good as K3’s reasoning quality, which the database records as vendor-reported, not independently benchmarked.

For deployment. The CLI runs on the local machine with approval gates but no dedicated OS-level sandbox engine, so untrusted work needs external isolation. Computer Use on macOS requires Accessibility + Screen Recording permissions, and the Windows variant may briefly take over mouse/keyboard — a real interaction cost for GUI automation.

For billing. Membership (Plus and above) shares one quota across all clients on a rolling 5-hour window plus a monthly cap; the Open Platform API is separate, pay-as-you-go. Teams weighing the two should note that membership is the consumer path and API keys (max 5, shown once) are the programmatic path.

What the evidence shows

The evidence is strong on the surface and the model lineup, weaker on independent quality. The GitHub repo and docs document the CLI/extension/desktop form factors, the subagents/MCP/Computer Use/browser/multimodal feature set, the kimi-cli deprecation (auto-migration on install), and the managed-endpoint architecture; the pricing page fixes the membership tiers and the API per-token rates.

The gaps are independence and sandboxing. No OS-level sandbox engine is documented (approval gates only), and no independent benchmark of Kimi Code’s coding quality is recorded here. China AI Hub analysis indicates the honesty boundary is “capability and pricing documented, quality vendor-reported”: the 1M output is a documented model fact, but whether that output is correct for a given repo is exactly what independent evaluation — absent here — would establish.

Where this fits

Workload	Relevance
Long-horizon whole-repo work (1M-token output)	High
Terminal / IDE coding (CLI, VS Code, desktop)	High
Multimodal input (text, images, video)	High
Scheduled and background tasks	High
OS-level sandboxed execution	Low (approval gates, no sandbox engine)
Self-hosted model service	Low (cloud inference only)
Independent coding benchmark	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	Kimi K3, Kimi K2.7 Code, Kimi K2.7 Code Highspeed	Official
Target users	Developers (terminal/IDE coding)	Official
Platform	CLI (TypeScript), VS Code extension, Desktop app	Official
OS	macOS, Windows, Linux (CLI)	Official
Browser / computer use	Browser control and Kimi Computer Use	Vendor-reported
Coding	Yes (terminal coding agent)	Official
Autonomous task execution	Yes (planning, multi-agent subagents)	Vendor-reported
MCP	Yes	Official
Tool calling	Yes	Official
Memory	Not publicly documented	Not publicly documented
Workflow	Plan-implement-iterate; long-horizon whole-repo work via 1M-token context	Official
API	Yes (Open Platform API pay-as-you-go)	Official
Pricing	Kimi membership (Plus and above); API $3.00/$15.00 per 1M tokens (kimi-k3)	Vendor-reported
Region	Global (api.kimi.com / api.kimi.ai)	Official
Open-source	Yes — MIT	Official
Deployment	Cloud and self-hosted	Official
Limitations	No OS-level sandbox engine; cloud inference only; Windows needs Git for Windows	Official
Source	Kimi Code GitHub	Official
Last verified	2026-09-20	Official

See the Moonshot AI company profile, the Kimi K3 model, the K3 vs K2.7-Code comparison, the choosing-a-coding-model guide, and the site’s AI agents and computer use technology pages.

Labels used above: Official fact (from the Kimi Code GitHub repo and docs), Vendor-reported claim (pricing and capability statements by Moonshot AI), and China AI Hub analysis (our synthesis, always introduced as such). No independent benchmark of Kimi Code’s coding quality is currently recorded.