Should you try Kun?
Try Kun in a disposable workspace if its shared desktop and terminal workflow answers a specific noncommercial need. Our run completed installation and build, but the test command remained unfinished when the harness stopped it. Before adopting it, check the license, reproduce the source build on your machine and complete the tests relevant to your intended work.
Kun's README says its GUI and TUI share a local runtime, including threads, plans and approvals. That continuity is the reason to evaluate it: code, documents and agent conversations use the same local application rather than separate histories. This is a documented capability, not a workflow we verified through interactive use.
What our sandbox actually recorded
Our run took place on October 13, 2026, at commit 8e16760, whose full SHA is 8e167608c63d2ec67ea01f90962c906b878c3611. The matching sandbox record specifies 3 CPUs, 8 GB RAM and the lab-node:22 image. Installation completed in 232 seconds and the build completed in 282 seconds, both with exit 0. The stored dependency measurement is 2,162 packages occupying 2,566 MB.
Our harness records each command's outcome separately through its success flag, elapsed time, exit code and timeout flag. It also stores parsed test counts and log tails. The test command timed out after 900 seconds with exit 124. Its stored parser summary reports 10 passed, 0 failed, 0 skipped and 10 total. Those figures do not establish exhaustive unique coverage. The raw tail shows passing results across several test files, but it does not provide a reconciled final suite total. We have not summed those file results or treated them as replacements for the stored summary.
The raw tail includes a language-server exit during an initialization-failure recovery test; that test file is marked as passing. It also contains SQLite warnings and a graph-supervision reconciliation message. These messages do not establish failed tests or explain why the overall command exceeded its time limit.
The record does not demonstrate an interactive desktop session, a live provider connection, completed coding or document tasks, mobile remote access, or a deployed CLI workflow. A successful build does not establish those outcomes. No new execution was performed for this correction.
The MrKeyoor testing methodology explains command outcomes and the limits of sandbox testing. That page describes the publisher's methodology; it does not establish that its current version governed this older run.
Installation succeeded with native dependency fallbacks
The English README requires Node.js 22.19 or newer for source development and a model connection for agent work. It also lists release packages for macOS, Windows and Linux. Our source-install footprint does not measure the size or installation experience of those release packages.
The recorded installation tail reports conflicting React peer dependencies. It also says the better-sqlite3 Electron prebuild failed and Kun would use a JSONL fallback; node-pty fell back to bundled binaries. Installation still returned zero. If your work depends on storage or terminal behavior, verify those paths on your target platform rather than assuming the fallback is equivalent for your workload.
Local storage and dependency advisories answer different questions
The README says conversations, memory and logs stay on your machine, while each cloud-model request sends its context to your selected provider. Local storage therefore does not establish that all task data stays on the computer. The README also describes approval requirements for tools, directories and workbench permissions. Our run did not test whether those controls prevent every unwanted action.
The attached npm audit result records 137 advisories, including 110 high and 27 moderate. These are recorded dependency advisories, not findings from a security review. The record lacks scannedPackages, so scan coverage is unknown. Neither the dependency footprint nor a package count printed in an installation log establishes that coverage. Review the affected dependencies and their relevance before granting access to valuable files or credentials.
Check commercial permission before integration
Kun's README requires separate written permission for commercial use, hosted services or integration into commercial products. It identifies the license as PolyForm Noncommercial 1.0.0. The repository detector's NOASSERTION value does not override that explicit README statement.
Keep release history separate from the measured checkout. The supplied v0.3.6 release identifies commit 993f315; our run used 8e16760. The stored installation and build logs identify their packages as 0.4.0. Results from our run should not be presented as measurements of v0.3.6.
The supplied sources also contradict two exclusions in the existing page: the current README lists mobile remote access, and the v0.3.6 release describes XLSX editing and conflict handling. We did not test either capability, so neither belongs in a claim of verified functionality or a blanket claim of absence.
Alternatives for a narrower coding workflow
Codex CLI is a local coding agent from OpenAI. Consider it when your immediate requirement is a coding agent you can launch from the terminal. Its README also points to editor, desktop and cloud experiences, so describing the whole product as terminal-only would be misleading. This is an alternative for the coding task, not evidence of a replacement for Kun's document workspace. See the Codex repository.
Continue's repository is no longer actively maintained and is read-only. Its README describes a coding agent available through a CLI, VS Code extension and JetBrains plugin, with a final 2.0.0 release. Consider that codebase only if you specifically need to inspect or maintain those integrations and accept its maintenance status. It should not be recommended as an actively maintained default. See the Continue repository.
How to read the scores
Assessed scores use this editorial rubric: 1 means a major obstacle to the stated use; 2 means substantial friction or unresolved verification; 3 means adequate evidence with material limits; 4 means strong evidence across the dimension; 5 means comprehensive evidence for the stated scope. These are judgments, not benchmarks or security ratings. An unassessed dimension has no numerical score.
Setup ease is 2/5 because our source installation had a substantial footprint and native dependency fallbacks. Documentation is 5/5 for the supplied entry-point coverage: the English README covers installation, model connection, platforms, source requirements, commands, TUI operation, contribution rules and licensing. That score does not certify the accuracy of every linked guide. Maturity is 2/5 because installation and build completed but the test command timed out and interactive behavior remains unverified. Community is unassessed because the supplied evidence does not establish contributor participation, support responsiveness or current adoption.