Key takeaways
- Moonshot AI's official terminal coding agent, rewritten in TypeScript as kimi-code (2.2K+ stars, created May 22, 2026, v0.14.0 on June 10, 2026), succeeding the Python-based kimi-cli (8.9K+ stars, October 2025)
- Powered by Kimi K2.6 — an open-weights ~1T-parameter MoE model (~32B active) with a 256K context window and a Moonshot-reported ~58.6% on SWE-bench Pro
- Aggressive pricing is the wedge: the tool is free (MIT), a coding membership runs ~$19/mo, and API pay-as-you-go rates undercut US labs by an order of magnitude
FAQ
What is Kimi Code CLI?
Kimi Code CLI is Moonshot AI's open-source terminal coding agent that reads and edits code, runs shell commands, searches files, and fetches web pages, powered by Kimi models or compatible providers.
How much does Kimi Code CLI cost?
The tool itself is free and MIT-licensed. Usage requires either a Kimi coding membership (~$19/mo) or pay-as-you-go Moonshot API credits (roughly $0.60–0.95/M input tokens as of mid-2026).
What model powers Kimi Code CLI?
Kimi K2.6 (released April 2026) — a mixture-of-experts model with ~1T total / ~32B active parameters, a 256K token context window, and open weights.
How is Kimi Code CLI different from Claude Code?
Kimi Code is a much cheaper, open-source, open-weights alternative — community reviews put K2.6 at 80–90% of Claude Code's quality on routine tasks at roughly 12% of the cost, but behind frontier US models on hard reasoning.
Executive Summary
Kimi Code CLI is Moonshot AI's official terminal coding agent — the Chinese lab's entry in the foundation-lab coding agent race alongside Claude Code, Codex, and Qwen Code. It reads and edits code, runs shell commands, searches files, and fetches web pages, with read-only operations auto-approved and mutations gated behind developer approval.[1] The current repo, kimi-code, is a ground-up TypeScript rewrite (2.2K+ stars, created May 22, 2026, v0.14.0 released June 10, 2026) succeeding the Python-based kimi-cli (8.9K+ stars, launched October 2025).[2][3]
The pitch is frontier-adjacent capability at a fraction of frontier cost. Kimi K2.6 — an open-weights mixture-of-experts model with ~1T total / ~32B active parameters and a 256K context window — posts a Moonshot-reported ~58.6% on SWE-bench Pro, and the coding membership runs ~$19/mo with API rates an order of magnitude below US labs.[4][5]
| Attribute | Value |
|---|---|
| Company | Moonshot AI |
| Founded | 2023 |
| Headquarters | Beijing, China |
| GitHub Stars | 2.2K+ (kimi-code) / 8.9K+ (kimi-cli predecessor)[2][3] |
| License | MIT[2] |
Product Overview
Kimi Code CLI is a terminal-first agent in the Claude Code mold: a TUI that plans, edits, and executes against your repo, distributed as a single binary or npm package (@moonshot-ai/kimi-code).[2][1] Authentication is via Kimi Code OAuth (membership) or a Moonshot AI Open Platform API key.[1]
It works natively with Kimi models — K2.6 is positioned for long-horizon execution and agent-swarm workflows — but also supports compatible third-party providers.[4][1]
Key Capabilities
| Capability | Description |
|---|---|
| Subagents | Built-in coder, explore, and plan agents run in isolated contexts for parallel work[2] |
| Video Input | Accepts screen recordings as context — demonstrate a bug instead of describing it[2] |
| MCP Support | AI-native configuration via conversational /mcp-config command[1] |
| ACP Integration | Agent Client Protocol support for Zed and JetBrains IDEs[2] |
| Lifecycle Hooks | Command automation around agent actions[2] |
| Plugins | Plugin ecosystem with marketplace access[2] |
| Plan / YOLO Modes | Plan mode for review-first workflows; --yolo flag for batch auto-approval[1] |
Product Surfaces / Editions
| Surface | Description | Availability |
|---|---|---|
| CLI (TUI) | Single-binary terminal agent | GA (v0.14.0)[2] |
| Zed / JetBrains | IDE integration via ACP | GA[2] |
| kimi-cli (legacy) | Python predecessor with migration guide to kimi-code | Maintenance[3][5] |
Technical Architecture
Installation:[2]
# Single binary (no Node.js required)
curl -fsSL https://code.kimi.com/kimi-code/install.sh | bash
# npm (requires Node.js 24.15.0+)
npm install -g @moonshot-ai/kimi-code
The rewrite moved from Python 3.13 + uv (kimi-cli) to TypeScript (98% of the codebase) with single-binary distribution and fast TUI startup.[2][5]
Key Technical Details
| Aspect | Detail |
|---|---|
| Deployment | Local CLI (single binary or npm) |
| Model(s) | Kimi K2.6 — ~1T total / ~32B active MoE, 256K context, open weights with native INT4 quantization[4][5] |
| Integrations | MCP, ACP (Zed, JetBrains), lifecycle hooks, plugin marketplace[2] |
| Open Source | Yes (MIT) — tool and model weights both open[2][4] |
Strengths
- Price disruption — ~$19/mo coding membership; community reviews report 80–90% of Claude Code's quality on routine tasks (code generation, tests, refactors, UI prototyping) at roughly 12% of the cost[5][6]
- Open weights, open tool — MIT-licensed CLI plus an open-weights model; self-hosting K2.6 is possible, unlike Claude Code or Codex[2][4]
- Modern agent feature set — subagents, MCP, ACP, hooks, and plugins match the Claude Code playbook at launch rather than years in[2]
- Video input — screen-recording context is a genuine differentiator among terminal agents[2]
- Fast iteration — 16 releases and 289 commits in the first three weeks of the rewrite[2]
- Proven predecessor demand — kimi-cli accumulated 8.9K+ stars in roughly eight months before the rewrite[3]
Cautions
- Very young rewrite — kimi-code was created May 22, 2026; at v0.14.0 and three weeks old, expect breaking changes, and kimi-cli users face a migration[2][5]
- Model trails frontier on hard tasks — Hacker News commenters assessed K2.6 as below Claude Sonnet/Opus-class capability, only slightly ahead of K2.5, and weak on domain-specific and single-turn high-stakes reasoning[6]
- Benchmark harness disputes — HN analysis flagged that Moonshot's Terminal-Bench comparisons used a harness that scores competitors lower than their own harnesses do; treat the ~58.6% SWE-bench Pro figure as vendor-reported[6]
- Pricing volatility — Kimi's pricing changed multiple times in 2026; published API rates are community-sourced ranges, not stable list prices[5]
- 256K context ceiling — smaller than the long-context options US rivals offer[5]
- China-based provider — teams with data-residency or procurement constraints around Chinese AI vendors will need the self-hosted open-weights path
Pricing & Licensing
| Tier | Price | Includes |
|---|---|---|
| Tool (CLI) | $0 | MIT-licensed, bring your own auth[2] |
| Kimi coding membership | ~$19/mo | Steady daily coding use via OAuth[5] |
| API pay-as-you-go | ~$0.60–0.95/M input, ~$2.50–4.00/M output tokens (community-sourced, mid-2026) | Moonshot Open Platform API key[5] |
| Self-hosted | Infra cost only | Open-weights K2.6 with INT4 quantization[4] |
Licensing model: Open source (MIT) tool + subscription or usage-based API; model weights are openly released[2][4]
Hidden costs: Pricing has shifted repeatedly in 2026 — verify current rates before committing; self-hosting a ~1T-parameter MoE requires serious hardware[5]
Competitive Positioning
Direct Competitors
| Competitor | Differentiation |
|---|---|
| Claude Code | Claude Code leads on model capability; Kimi Code costs ~12% as much on routine work[6] |
| Codex | Codex spans desktop/web/mobile surfaces; Kimi Code is terminal + ACP only, but fully open source |
| Qwen Code | Closest analog — both Chinese-lab, open-source CLIs with open-weights models; different model families |
| Gemini CLI | Gemini CLI's individual tiers sunset June 2026; Kimi Code is a candidate destination for free-tier refugees |
| OpenCode | OpenCode is model-agnostic by design; Kimi Code defaults to Kimi but accepts compatible providers[1] |
When to Choose Kimi Code CLI Over Alternatives
- Choose Kimi Code when: Cost dominates — high-volume routine coding where 80–90% of frontier quality at ~12% of the price is the right trade[6]
- Choose Kimi Code when: You want an open-weights escape hatch — the same model can be self-hosted if the vendor relationship ends
- Choose Claude Code or Codex when: Hard reasoning, domain-specific work, or single-turn high-stakes tasks where K2.6 lags frontier models[6]
- Choose Qwen Code when: You prefer the Qwen model family or Alibaba's ecosystem
Ideal Customer Profile
Best fit:
- Cost-sensitive individual developers and startups with high agent-token burn
- Teams that want open-source tooling and open-weights models end to end
- Zed and JetBrains users who want native ACP agent integration[2]
- Developers leaving Gemini CLI's sunsetting free tier who won't pay US-lab subscription prices
Poor fit:
- Enterprises with data-residency or procurement restrictions on China-based AI providers (unless self-hosting)
- Teams needing frontier-level reasoning on high-stakes tasks[6]
- Anyone requiring API pricing stability for budgeting[5]
Viability Assessment
| Factor | Assessment |
|---|---|
| Financial Health | Venture-backed Chinese lab; specific funding totals not covered in cited sources |
| Market Position | Challenger — credible open-source alternative on price, behind US labs on capability[6] |
| Innovation Pace | Fast — full TypeScript rewrite shipped to v0.14.0 in three weeks; K2.6 released April 2026[2][5] |
| Community/Ecosystem | Early — 2.2K+ stars on the rewrite, 8.9K+ on the predecessor; plugin marketplace nascent[2][3] |
| Long-term Outlook | Promising but unproven — depends on K2-line model cadence keeping the price/quality gap attractive |
Moonshot is executing the full foundation-lab playbook — own model, own CLI, subagents, MCP, plugins, membership pricing — compressed into months. The open-weights model is a structural hedge competitors can't match, but the rewrite's youth and the model's frontier gap make this a fast-follower bet, not a leader bet.
Bottom Line
Kimi Code CLI is the price-disruption entrant among foundation-lab coding agents: a free, MIT-licensed terminal agent over an open-weights ~1T-parameter MoE model, with a ~$19/mo membership that undercuts Claude and OpenAI subscriptions while delivering most of their quality on routine work.[2][5][6]
The honest caveats: the TypeScript rewrite is three weeks old, benchmark claims are vendor-reported and contested, pricing has been a moving target, and K2.6 trails frontier models on hard reasoning.[6][5]
Recommended for: Cost-sensitive developers running high volumes of routine agentic coding, and teams that value an open-source tool plus open-weights model.
Not recommended for: Enterprises with China-vendor constraints, teams needing frontier reasoning, or anyone allergic to early-version churn.
Outlook: Watch whether kimi-code's star growth catches its predecessor, whether pricing stabilizes, and whether the next K2-line release narrows the capability gap — that combination would make this the strongest open challenger in the category.[7]
Research by Ry Walker Research • methodology
Sources
- [1] MarkTechPost: Moonshot AI Releases Kimi Code CLI
- [2] Kimi Code CLI GitHub Repository
- [3] kimi-cli GitHub Repository (predecessor)
- [4] Kimi K2.6 Tech Blog: Advancing Open-Source Coding
- [5] Kimi Code CLI Guide: Setup, Pricing & Tips (Lushbinary)
- [6] Hacker News: Kimi K2.6 — Advancing open-source coding
- [7] Kimi Code Website