mrkeyoor.com_
Sat 26 Sept 23:39 UTC
LLM Toolsevaluationupdated 26 Aug 2026

Kun review

Kun is a Chinese-first, local AI-agent workbench for coding, design, writing, document analysis, and automation through a shared desktop and terminal runtime. A full English README and English guides exist, so English-speaking users can evaluate and operate it without translating the main Chinese page.

+1stars / 7d
Verdict

Our Kun install consumed 2,374 MB, and its wider suite finished with 85 failed tests plus 114 known vulnerabilities, so v0.3.6 is not a low-risk default for sensitive work. Noncommercial users may still find the shared GUI and TUI runtime unusually coherent if they can isolate tools, review every permission, and accept a fast-moving release. Commercial teams need separate licensing before adoption.

We ran it

Lab card: what happened when we ran KunScreenshot of Kun (www.kun-agent.com)
Install✓ · 237s1939 packages · 2374 MB
Build✓ · 101s
Tests✗ · 626s2 passed · 0 failed of 2 (node:test)
Known vulns1140 critical · 100 high · 14 moderate · 0 low (npm audit)
Repo5992 files~1,187,327 lines of source · 116.4 MB · 5 CI workflows

Answers from our run

Does Kun build from source?

Dependencies installed in 237 seconds (1939 packages), and the build succeeded in 101 seconds. We cloned commit 993f315 into a clean Debian container with 3 CPUs and no project-specific setup.

Do Kun's tests pass?

Yes: 2 of 2 passed when we ran the project's own test command (node:test). Some failures need services or credentials a bare container does not have.

Does Kun have known vulnerabilities in its dependencies?

npm audit flagged 114 known advisories in the dependency tree at the time of our run.

Who should not use Kun?

Commercial teams expecting an open-source business license: PolyForm Noncommercial permits learning, research, and noncommercial use, while business use needs written authorization.

What are the alternatives to Kun?

Continue, OpenHands, Codex. Our Kun install consumed 2,374 MB, and its wider suite finished with 85 failed tests plus 114 known vulnerabilities, so v0.

Setup2/5237-second install and 2,374 MB dependency footprint
Docs5/5Full Chinese and English paths cover runtime, TUI, and workflows
Community4/56,254 stars with same-day pushes and responsive issue activity
Maturity2/5v0.3.6 is broad, but 85 tests failed and audit risk is high

Who it’s for

Individuals who want one local task history across a desktop GUI and terminal TUI.
Developers who need visible plans, approvals, diffs, tool calls, and test evidence around agent work.
Writers and researchers working with Markdown, PDFs, spreadsheets, and presentations.
Users who already have a supported model subscription, API key, or self-hosted endpoint.

Who it’s NOT for

Commercial teams expecting an open-source business license: PolyForm Noncommercial permits learning, research, and noncommercial use, while business use needs written authorization.
Small machines or bandwidth-limited setups: our source install pulled 1,939 packages and occupied 2,374 MB.
Teams requiring a green upstream test suite and clean dependency audit: our run ended with 85 failed tests and 114 known vulnerabilities, including 100 high severity.
Users who need editable Office documents: the README says Office files remain read-only.
People seeking phone control or remote access to the Kun runtime: issue 1029 requests those capabilities rather than documenting them as present.

Setup reality

Our sandbox install took 237 seconds, added 1,939 npm packages, and used 2,374 MB. The build succeeded in 101 seconds. Tests exited 1 after 626 seconds: the Node step passed 2 of 2, while the log tail reported 4,616 passed, 85 failed, and 3 skipped in the wider suite.

Kun also needs Node.js 22.19 or newer and at least one usable model connection. Release installers avoid compiling from source, but cloud providers still receive prompts, attachments, and task context under their own data policies.

Npm audit found 114 known vulnerabilities: 100 high and 14 moderate. The monorepo has no Dockerfile, and the app spans Electron, a local runtime, TUI, browser and terminal tools, extensions, MCP, scheduled work, and provider integrations.

Kun keeps desktop and terminal work in one local history

Kun's strongest idea is continuity. The Electron interface and keyboard-focused TUI connect to the same kun serve runtime, sharing threads, plans, goals, approvals, and background tasks. Code mode handles repositories, branches, worktrees, terminals, diffs, tests, review, and a Design canvas. Work mode covers Markdown, document questions, spreadsheets, presentations, and whiteboards. That range makes Kun closer to a local workbench than a coding extension attached to one editor.

Release v0.3.6 supports macOS on Intel and Apple Silicon, Windows x64, and Linux packages, plus standalone TUI archives. The app stores sessions, preferences, logs, and runtime data locally by default. Local storage does not keep a cloud model private: prompts, attachments, and task context go to the selected provider. Presets cover several subscription, API, compatible-service, and self-hosted routes, with actual models and quotas depending on provider rules.

The English path is complete, while Chinese remains primary

The main README is Chinese, but README.en.md mirrors its product, setup, source-build, license, and contribution guidance. English versions also exist for the TUI, workflow loops, extensions, local development, and contribution process. Some linked material remains shared or Chinese-first, but an English-speaking developer can get from download to a source build without machine translation. That is better than a translated summary attached to an otherwise Chinese-only project.

From source, Kun requires Node.js 22.19 or newer, npm, and at least one model connection. npm run dev builds the runtime and starts Electron; separate commands cover the TUI, type checks, lint, tests, production builds, and platform packages. Our checkout contained 5,992 files and about 1,187,327 lines of source. The size matches the feature list, but it also means newcomers are entering a large monorepo rather than a thin desktop shell.

What happened when we ran it

Our sandbox installed commit 993f315 in 237 seconds. Npm added 1,939 packages and used 2,374 MB on disk before any user projects or model caches. The production build succeeded in 101 seconds. The run used a fresh unprivileged Debian container with 3 CPUs, 8 GB of RAM, and no secrets, so it checked the repository without connecting a provider or granting real workspace access.

Tests ran for 626 seconds and exited with status 1. A Node test step reported 2 passed and 0 failed out of 2, while the supplied Vitest tail reported 565 passing files, 25 failing files, and 1 skipped file. At the test level it counted 4,616 passed, 85 failed, and 3 skipped out of 4,704. One visible assertion expected a rejected start hook to return an error object. The log does not establish one shared cause for all 85 failures.

Npm audit found 114 known vulnerabilities: 100 high, 14 moderate, 0 critical, and 0 low. The repository scan found 5 CI workflows, no Dockerfile, no tests directory, and npm workspaces. These audit counts do not prove that every advisory is remotely reachable through Kun. They do make dependency triage mandatory before putting valuable credentials, repositories, or SSH access into the app.

Broad tool access raises the cost of a bad release

Kun can edit files, run terminal commands, use Git worktrees, browse, connect through SSH, install extensions, call MCP tools, and schedule work. Visible permissions and approvals are the right design, yet breadth increases the consequences of a mistaken grant. A 2,374 MB dependency tree with 100 high-severity advisories deserves isolation from production credentials until each relevant path is understood. Start with a disposable repository and a restricted provider key.

Version 0.3.6 also changed plan execution to use agent-managed worktrees by default, added scheduled builds, SSH terminals, agent evaluations, and experimental session visualizations. Those are meaningful capabilities arriving in one v0 release. Open issue 1233 reports that returning to a stopped Design session can unexpectedly reactivate creation. Issue 1235 reports that one subscription provider still fails model synchronization in 0.3.6. Neither issue proves core file editing is unsafe, but both show active edge behavior.

PolyForm blocks ordinary commercial adoption

Kun uses PolyForm Noncommercial 1.0.0 rather than an OSI-approved commercial license. The README limits use to learning, research, and noncommercial purposes. Commercial use, distribution, hosted service, resale, or inclusion in a commercial product requires separate written authorization. A company evaluating Kun for paid engineering work should resolve that point before spending time on provider integration or internal deployment. Source access alone does not settle permission.

GitHub showed 6,254 stars, 9 open issues and pull requests combined, and a last push on August 26, 2026. Release v0.3.6 arrived August 21, so development and user feedback are current. That activity helps a fast-moving desktop product, but it does not cancel our 85 failed tests or 114 audit findings. Kun is compelling as a noncommercial trial for users who want one visible workspace across code and documents. It is not ready to receive blind trust.

Alternatives

ProjectWhat it isPick it when
Continue gh↗An open-source coding assistant for IDE and terminal workflows with configurable models.pick this instead when coding inside an existing editor matters more than a separate workbench for documents and design.
OpenHands gh↗A software-development agent platform centered on delegated coding tasks and sandboxed execution.pick this instead when autonomous software work and remote execution matter more than a local desktop productivity suite.
Codex gh↗A terminal-first coding agent with approvals, tools, and repository workflows.pick this instead when the terminal is the product and you do not need Kun's design canvas or document workspace.

What people are saying

  1. [github-trending] KunAgent/Kun

Sources

  1. Kun Chinese README
  2. Kun English README
  3. Kun v0.3.6 release
  4. Design session reactivation report
  5. Provider synchronization report
  6. Remote access request

More llm tools reviews

llm-wiki-compiler · claude-skills · Humanizer-zh · agent-beacon · MiMo-Code · pi-claude-bridge · the whole board →