Kimi Code CLI: The MIT Terminal Agent Without a Meter

Kimi Code CLI is Moonshot AI's terminal coding agent - the successor to kimi-cli, rewritten in TypeScript and shipped as a single MIT-licensed binary (a one-line install script for macOS/Linux/Windows, or npm install -g @moonshot-ai/kimi-code; the script path needs no Node.js). It reads and edits code, runs shell commands, searches and fetches web pages, and plans multi-step tasks from feedback. Built-in coder, explore, and plan subagents run in isolated context windows, MCP servers are configured conversationally via /mcp-config, lifecycle hooks gate risky tool calls, and kimi acp speaks the Agent Client Protocol so Zed and JetBrains can drive a session. It defaults to Kimi K2.7 Code but is model-agnostic: OpenAI-compatible, Anthropic-compatible, and Google-compatible endpoints all work through config.

The CLI is free and open source (MIT license, v0.42.0 as of September 2026 with 33,000+ weekly npm downloads) - but like every 2026 agent CLI, the tokens are not. Three routes: (1) a Kimi membership subscription, from \$19/mo on the English pricing page, which bundles weekly Kimi Code quotas; (2) a Kimi Platform API key billed per token (\$0.95 input cache-miss / \$4.00 output per 1M on K2.7 Code); (3) bring your own provider - the CLI accepts Anthropic-, OpenAI-, and Google-compatible endpoints, so you can point it at any model you already pay for or a local server. Installing Kimi Code also auto-migrates configuration and sessions from the old kimi-cli.

Per Moonshot's pricing documentation (September 2026): kimi-k2.7-code bills \$0.95 per 1M input tokens on cache miss (\$0.19 on cache hit) and \$4.00 per 1M output tokens, with a 262,144-token context window. The HighSpeed variant - the same model at roughly 180 tokens/s output, up to 260 in short contexts - doubles both rates: \$1.90 cache-miss input / \$8.00 output. The Batch API runs at 60% of standard pricing (\$0.57 / \$2.40 per 1M), the cheapest published route for bulk work. Third-party gateways track K2.7 Code slightly under list (about \$0.68-\$0.74 input / \$3.40-\$3.50 output). At \$4.00 per 1M output tokens, K2.7 Code runs roughly one-fifth of frontier Opus-class output pricing - the reason long autonomous agent loops are its natural workload.

The English pricing page (August 2026) lists four tiers at \$19, \$39, \$99, and \$199 per month with relative Kimi Code allowances of 1x, 5x, 15x, and 30x; the help center surfaces the same ladder in yuan (\u00a549 / \u00a599 / \u00a5199 / \u00a5699) under different tier names (Andante/Moderato/Allegretto/Allegro vs Moderato/Allegretto/Allegro/Vivace) - check the entitlements inside your own account rather than assuming a name maps across surfaces. Kimi Code is available to all paid members (model id kimi-for-coding); the HighSpeed tier requires an upper-tier plan. A free Adagio tier exists for general Kimi use, but Kimi Code quota is a paid-membership feature, and third-party reviewers reported membership signups hitting a waitlist in August 2026.

Three overlapping controls govern Kimi Code, and none is published as a simple prompt count: (1) a shared monthly membership credit pool that all Kimi features draw from, refreshed monthly with no rollover; (2) a Kimi Code weekly allowance refreshed every seven days from your subscription date, also with no rollover; (3) a separate Kimi Code usage window the help center describes as 5 hours per week and community tooling that parses Kimi's own BillingService endpoint reads as a 200-requests-per-5-hours rolling cap on every tier. The same tooling reports weekly request quotas of about 1,024 (\u00a549 tier), 2,048 (\u00a599), and 7,168 (\u00a5199). Burn any layer and Kimi Code pauses until that layer resets; unused quota never carries over.

Moonshot wound it down in favor of Kimi Code CLI. The legacy Python/Apache-2.0 project (10,000+ GitHub stars) froze at v1.49.0, and its own README states that Kimi CLI is evolving into Kimi Code CLI and that installing the new CLI auto-migrates your configuration and sessions. The successor is TypeScript, MIT-licensed, ships as a single binary with a millisecond-start TUI, and releases several times a week. Homebrew has marked the kimi-cli formula deprecated with a disable date of January 17, 2027. If you are evaluating today, install kimi-code - the old repo remains documented but is not where development happens.

Moonshot reports K2.7 Code cuts thinking-token usage by roughly 30% versus K2.6 while improving agentic performance about 10%, and third-party MCPMark Verified reads put it at 81.1% against a reported 76.4% for Claude Opus 4.8 on the same benchmark. The honest caveat: K2.7 Code had not been submitted to independent suites like SWE-bench Verified or Terminal-Bench as of mid-2026, and most headline numbers are vendor-run or cross-harness comparisons - treat them as directional. Claude keeps the verified-quality and ecosystem edge; Kimi's edge is price per token (\$4.00 per 1M output is a fraction of Opus-class pricing) plus an MIT-licensed, model-agnostic agent wrapped around it. Note K2.7 Code always thinks - thinking mode cannot be disabled and sampling parameters are fixed.

Kimi Code is a natural fit for the VibeFuse canvas: a free Windows AI harness where every agent runs as a widget over your real repositories. VibeFuse charges \$0 - no harness meter, no weekly allowance, no 5-hour window - your Kimi Platform key (or any other provider's) does the token billing, per widget and per task. That pairing fixes Kimi Code's sharpest limitation: quota burn stays isolated to one widget while your other agents keep working, and when you hit the weekly wall you can route the task to a different vendor's CLI on the same canvas instead of waiting for a reset. VibeFuse also runs an open-source marketplace where creators sell widgets, skills, and styling packs for revenue share - pair it with VocalFuse for local voice dictation while you build.