Updated on
Kimi CLI: reviews and analysis
Moonshot's terminal agent: open source with very low API prices.
Affiliate link · no extra cost to you
Our verdict
Kimi CLI competes on price: the agent is open source and Moonshot's models cost a fraction of the leaders, with membership from $19 or pay-as-you-go via API. For volume tasks where cost per token rules, the math strongly favors it. On the most demanding long-reasoning tasks the leaders stay ahead, and the English documentation trails the product.
Best for: Cost-sensitive developers running many bounded tasks.
What changed this week
Sweep of
Control rises because since 30 July plugins can ship their own agents and inject system-prompt instructions, plus four new hook events for watching execution (official changelog, v0.31.0-v0.32.0). Integration rises because on 5 August Kimi Computer Use and Kimi WebBridge become official plugins installable from the catalogue (v0.33.0). The overall score falls because this week's mood is bad: a Hacker News thread and a GitHub issue document poor performance and an auto-approve bug, with no verifiable positive signal to offset them.
- down
- down
Two sweeps down in a row. Three would make it a trend; two do not.
We have tracked it since July 25, 2026: 3 sweeps on record, with the score moving between 3.5 and 3.9. We publish the latest change here; the full series is not published.
What people said this week
Tracked from July 30, 2026 to August 6, 2026 · 3 new signals
An issue in the successor's repository documents that the VS Code extension's yolo auto-approve mode silently drops to manual after an idle re-attach, with no mode-change event in the log Read it
In a Hacker News thread about agent benchmarks, a commenter says they would have preferred the benchmark use Kimi Code's native harness instead of opencode, and notes that the variation between Kimi providers shows you cannot trust a single trial Read it
A Hacker News user tests Kimi Code after importing their Claude Code setup and finds it falls short: slow, token-hungry, and claiming tasks are done when the app will not even start, with another user confirming a similar pattern with other Chinese models Read it
This week’s sentiment is one fifth of the score. It is only computed when there are at least two dated, linkable signals: without them, the score stays on the five sweep axes.
What the internet says
The price stays the same, confirmed today on the official page: $19 a month to start. The product gains control across three releases shipped in six days, from 30 July to 5 August: plugins can now ship their own agents and inject system-prompt instructions, four new hook events land for watching execution, and Kimi Computer Use and Kimi WebBridge become official plugins installable straight from the catalogue. The week in the community, though, was bad: a user who imported their Claude Code setup finds it slow, with tasks reported done when they were not, and a documented bug makes the VS Code extension's auto-approve mode silently drop to manual after idle time, with no trace in the log. Open issues keep climbing in both of the project's repositories, 525 in the one being retired and 572 in its successor.
What the web repeats in favour
- A paid step from $19 a month with a free plan below it, well under what the western terminal agents charge
- MCP servers configured for other agents work untouched, and it speaks the agent client protocol several editors use
- Open source under Apache-2.0 with very active development in the repo
- The successor splits work across built-in coder, explore and plan subagents in isolated contexts, and can be driven from Zed or JetBrains
- Across three releases in six days, plugins can now ship their own agents and inject system-prompt instructions, and Kimi Computer Use and Kimi WebBridge become official plugins installable straight from the marketplace
What the web repeats against
- The repo notice no longer talks about a rename: it says the project will be wound down gradually and sends you to another repository, under another licence, to keep using it
- 525 open issues in the repo being retired and 572 in the one replacing it, with install, proxy and MCP failures reported by users outside the team
- The allowance has not one limit but three stacked: it refreshes every seven days, there is also a rolling five-hour window, and on top of that a monthly cap from the membership
- In July several independent voices report burning through the five-hour window on a single task and compare the limits unfavourably with the ChatGPT or Claude plans
Sweep sources: GitHub · GitHub · Official pricing · Docs · Communities · GitHub · Communities
Pros / Cons
Pros
- Cost per token far below the leaders
- Open source with pay-as-you-go available
- Performs well on bounded coding tasks
Cons
- Behind the leaders on long reasoning
- Greener documentation and community
TLDR: Kimi CLI competes on price by a wide margin: an open source terminal agent on Moonshot’s models, which cost a fraction of the leaders, with membership from $19 a month or pay-as-you-go via API. For volume tasks where cost per token rules, the math strongly favors it. On long reasoning the leaders stay ahead, and documentation outside Chinese trails the product.
What Kimi CLI is and how it works
Moonshot, the Chinese lab behind the Kimi models, entered the agent category with the challenger’s classic strategy: same format as the leaders, unbeatable price. Kimi CLI is an open source terminal agent doing what defines the genre (takes tasks, reads the project, edits, executes) on models whose cost per token is a fraction of the Western frontier ones.
The Kimi model family has surprised on coding leaderboards, performing far above what its price suggests, and that ratio is the product’s entire argument: it does not promise the best judgment, it promises the best judgment per dollar.
The payment routes are two: the Kimi membership, with tiers from about $19 a month, or direct pay-per-use on its API platform, which for agent use is the route most people choose for transparency.
What it is like day to day
On bounded, well-defined tasks (fixes, specific features, mechanical refactors, test generation), the experience is a first-division agent at second-division prices: the task comes back done and the bill barely moves. That is its niche, and for automations and batches it is a cheap workhorse.
On long, ambiguous tasks, where the agent explores, decides and chains unguided, the distance to Claude Code appears: more detours, more supervision, more iterations. The usage pattern the community has settled on is the second mount: the leader for the hard, Kimi for the volume.
The non-technical friction is the ecosystem: greener documentation outside Chinese, a Western community still forming and that discomfort of depending on a provider whose fine print reads worse. Each team weighs that factor differently, and it deserves deciding consciously.
Pricing and plans
The agent is free and open source. The cost is the model: membership from about $19 a month or pay-per-use on the API, where Kimi’s code models cost a fraction of the frontier price. For volume work, the per-task difference is whole multiples.
The useful comparison is threefold: against the leaders, you save most of the bill in exchange for judgment on the hard parts. Against free Gemini CLI, you pay a little for better code models. Against Qwen Code, the fight is between equals and gets decided by the models and limits of the moment.
Who it is for (and who it is not for)
Kimi CLI is for cost-sensitive developers with many bounded tasks: automations, maintenance, batch generation, personal projects on finite budgets. Also as the volume second agent alongside a subscription leader.
It is not for anyone delegating the ambiguous without watching the price, nor for teams whose data policies rule out providers by jurisdiction, a decision worth making explicitly here.
Kimi CLI alternatives
Qwen Code is the direct rival in the minimum-cost corner, with an even more aggressive free tier. Gemini CLI gives the free Western route with Google’s backing. And Claude Code or Codex are the judgment jump when the task demands it. They are in Kimi CLI alternatives and the ranking in the best AI coding tools.
Frequently Asked Questions
How much does it really save?
On cost per token, Kimi’s models cost a fraction of the frontier ones: for volume work, the monthly difference is multiples, not percentages. The real saving depends on how many of your tasks are the kind Kimi solves well.
How does it perform against Claude Code?
On bounded tasks, surprisingly close. On long reasoning and autonomous exploration, clearly behind. The settled pattern is using them as a pair: volume to Kimi, difficulty to the leader.
Is it open source?
The agent yes, so you can inspect and extend it. The models are Moonshot’s paid service, via membership or API.
Should I worry about where my data goes?
As with every provider: read the data policy and decide for your case. For teams with jurisdiction requirements, that factor decides before price, and it deserves being treated that explicitly.