Our uniTerm source build succeeded in 2 seconds, but tests exited 1 after 128 seconds at commit 716a95c. The application suits developers who want SSH, file transfers and databases in one connection workspace and can evaluate it with disposable credentials. Skip it as your only credential archive or when your selection requirement is documented per-client MCP host authorization. The useful decision is how much access and recovery responsibility to consolidate in one application.
What people use uniterm for
ys-ll/uniterm combines remote terminals, file transfers, desktops, databases and containers, with an optional AI agent. The official introduction describes these categories, while the repository README lists SSH, SFTP, RDP, PostgreSQL and Kubernetes. This is the Go/Vue remote-access application. The separate Rust application at uniterm.dev is a terminal multiplexer and agent-fleet supervisor, so the owner/repository name matters when looking for documentation.
The sampled users give consolidation a practical meaning: they saved credentials, opened SSH sessions, downloaded folders and used remote desktops. Their reports also show which parts of an evaluation deserve attention. The table preserves those reported outcomes as user experience, rather than treating a protocol listing as a successful connection or a support thread as our own test.
| Person and task | Reported outcome | Source |
|---|---|---|
| rockswang primarily used uniTerm as a credential manager | Encrypted credentials became unreadable after the system keychain entry disappeared; the maintainer confirmed recovery was impossible | Credential recovery report #1383 |
| macaty connected over SSH and tried the AI agent on Windows v1.11.2 | Execution stalled, and the user supplied a session log | SSH and AI report #1307 |
| lixiplus downloaded a folder through SFTP in v1.9.2 | A JSON file failed from Kylin V10 despite reported root access and matching permissions; another terminal downloaded it | SFTP download report #818 |
| zhangsir1211 used longer AI conversations on macOS | Busy-console messages and typing lag led the user back to Netcatty | AI responsiveness report #1426 |
| itnewgay tried RDP in v1.4.1 | RDP worked, but the user wanted automatic resolution adjustment and saved credentials | Historical RDP discussion #227 |
The installation guide also describes a Docker-hosted browser service where each user has separate connections, sessions, keys and data. Alongside the documented Windows, macOS, Linux and Android clients, that creates several installation choices. For a developer moving repeatedly between a shell, remote files and a database, the shared connection workspace is a concrete reason to evaluate uniTerm.
What happened when we ran it
Our November 15, 2026 run installed commit 716a95c in 52 seconds and built it in 2 seconds; tests exited 1 after 128 seconds. The recorded sandbox had 3 CPUs, 8 GB RAM and image golang:1.24-bookworm. We measured repository setup and command outcomes. Keeping the commit beside the results makes the evidence useful without assigning those outcomes to a different release.
| Recorded step | Result | Duration or count |
|---|---|---|
| Install | Succeeded | 52 seconds; 256 dependencies recorded |
| Build | Succeeded | 2 seconds |
| Tests | Failed, exit 1 | 128 seconds; summary reports 17 passed and 2 failed of 19 |
The checkout contained 1,088 files, approximately 234,289 source lines and 15 MB of files. The scan found six CI workflow files and a Dockerfile. It found no tests directory, while the Go log contains tests beside their packages. These are repository observations: the directory scan and the test command answer different questions, so the missing directory should not be translated into an absence of tests.
The log tail names six failing functions in backend/sync: TestSnapshotSaltReusedAcrossPushes, TestConfigureRepoWebDAVFreshRemote, TestConfigureRepoWebDAVAdoptsRemote, TestConfigureRepoWebDAVWrongMasterPassword, TestSyncSkipsNeverPersistedStore and TestPushSnapshotUploadsOnlyChangedBlobs. Preserve those names separately from the summary's 17 passed and 2 failed of 19. The same tail records successful results for backend/update, backend/utils and cmd/uniterm-gateway. The concrete outcome is a successful source build followed by a failing test command.
Where people get stuck
The sampled issues name Windows crashes, AI stalls, SFTP failures and input problems, with different responses from maintainers and users. Reading the response matters as much as reading the symptom: a proposed snapshot, a personal workaround and a release-note fix describe different events. The table identifies those events directly so readers can choose which historical workflow to reproduce.
| Report | Symptom | Recorded response or outcome |
|---|---|---|
| Windows crashes #1426 | Portable builds, including v1.12.2 and v1.12.3-alpha, disappeared while idle | Alpha logging exposed callback exhaustion; the maintainer identified the cause and scheduled a stable-release fix |
| AI execution #1307 | Execution stalled after an operation or repeatedly exceeded the context budget | A contributor proposed discarded parallel tool calls and broken context grouping as a possibly related cause, with PR #1312 |
| SFTP download #818 | A JSON file failed to download | The reporter said permissions matched and another terminal worked; the maintainer continued analysis |
| Missing macOS input #146 | Rapid typing lost characters | Configuration sync helped the original reporter; another user reported the symptom in v1.7.0 |
| Duplicate input #655 | Uppercase or Chinese-input typing duplicated characters on macOS v1.8.0 | A commenter used window switching as a workaround; the maintainer offered snapshot fdfb06f |
| Windows title bar #666 | Window controls were missing on Windows 10 21H2, v1.8.0 | Restarting and the suggested snapshot failed to solve the reporter's problem |
The v1.12.3 release notes, published November 15, describe fixes for callback exhaustion, busy sessions and streaming-related typing lag. The crash error was fatal error: too many callback functions. That connects the release claim to a specific diagnostic result, rather than asking readers to infer resolution from a closed issue label. The measured source checkout remains commit 716a95c.
Credential recovery depends on retaining the decryption key
In issue #1383, the maintainer confirmed that losing the sole keychain key made the existing encrypted credentials unrecoverable. The investigation found encrypted fields still present in configuration files and the key missing from the system credential manager. The distinction is consequential: retaining encrypted data does not give its owner a way to read it after the only decryption key disappears.
The credential investigation says resetting the store and re-entering passwords retains other connection fields, including hosts, ports, usernames and groups. It also distinguishes master-password mode, which derives the key from the password, from keychain mode. The maintainer advised preserving files and completing diagnosis before clearing credentials. Choose a recovery model before saving valuable secrets, and keep recoverable originals elsewhere.
The thread also records an inaccurate setup prompt for an existing keychain-mode store, which the maintainer acknowledged and promised to improve. On the upgrade question, the maintainer says updates require a manual click. The investigation found the key missing while the user still ran v1.9.0. That sequence supports the keychain-loss explanation and separates it from the reporter's initial suspicion about a forced upgrade.
MCP approvals govern commands and transfers
The MCP guide says all tools are available by default and side effects are governed by approval policy. Its default policy prompts for every command, connection and transfer. Other policies prompt according to command risk, and the approval dialog identifies the requesting client, target connection and full command. This answers the command-approval part of Obiwayneone's Reddit question.
The current MCP guide describes loopback access by default, separate bearer tokens for clients and a JSONL audit log. The endpoint uses plain HTTP; the guide describes SSH tunnelling for encryption. Upload and download paths are unrestricted, with every transfer raising a dialog showing source and destination. These current instructions replace the older bookmark-directory restriction described on the previous MrKeyoor page.
For an initial evaluation, use an OS account or separate instance containing only test connections and disposable credentials, and keep confirmation enabled. This is our recommended setup choice. It gives the operator a concrete way to control which access the application holds before introducing an agent. Review the target and command at each prompt, just as you would review the destination and overwrite consequences of a file transfer.
AI sessions affect ordinary terminal work
A macOS user in #1426 reported that longer AI conversations made command entry sluggish and repeatedly produced a busy console. In #1307, a contributor described a possible parallel-tool-call defect: later calls were discarded, leaving unfinished calls that interfered with context grouping. The v1.12.3 notes separately describe changes to context compaction, streaming rendering and foreground execution ownership.
A useful evaluation sequence follows those symptoms: type while an answer streams, interrupt a command, and continue a conversation through several tool calls. The terminal needs to remain usable while the assistant works. In #1145, a user also reported that Ctrl+C or Escape would not stop a continuous ping, and the maintainer requested clarification. That interruption belongs in the same practical checklist.
Archive analysis has a separate explanation in #1145. The maintainer says uniTerm connects directly to the model, without a server-side archive-unpacking step, and therefore does not support that workflow. Screenshot and attachment support had separately been announced for v1.10.1-alpha. Distinguishing an attachment interface from archive processing gives a first-time user a clearer expectation than grouping every file type under one upload feature.
Windows lockouts had a specific runtime investigation
Issue #957 describes failed local Windows logins associated with WebView2 startup and account lockouts under a strict policy. The report names uniTerm v1.9.4. An initial snapshot failed to stop the events. The maintainer then attributed the behavior to WebView2 Runtime 153.0.4234.x after comparative testing, including a minimal WebView2 host, and reported a locally successful mitigation through PR #1010 and snapshot 3dbd92d.
The Windows lockout investigation gives an administrator specific evidence to collect: runtime version, Windows Security events and lockout policy. The maintainer's later local test stopped event 4625 requests. Keep those details attached to the report when diagnosing a similar machine; they identify a local Windows startup interaction and provide a more useful starting point than a general instruction to reinstall uniTerm.
English documentation and client requirements
The official introduction lists English among 11 interface languages, and the repository provides an English README and installation instructions. The sampled credential, crash and input discussions contain Chinese maintainer explanations. An English-speaking operator can begin with the documentation, then translate the relevant issue thread when investigating a matching symptom. Snapshot identifiers and affected versions help connect those discussions to the installed application.
The README's installation section lists Windows installers and portable archives, macOS disk images, Linux packages and archives, an Android APK and Docker deployment. Its source-build requirements are Go 1.26+, Node.js 20+ and wails3 CLI v3.1.0. Linux source builds add GTK 3 and WebKitGTK 4.1 development libraries; Android builds add JDK 21 and the Android SDK/NDK. Select instructions for the route you intend to use.
uniTerm vs alternatives
MobaXterm documents SSH, RDP, VNC and SFTP sessions; mRemoteNG describes an open-source, multi-protocol, tabbed connection manager. Those primary descriptions provide a concrete shortlist alongside the user comparisons. The decision here is which workflow you need to retain, such as remote sessions, output logging or a tabbed connection list, rather than which application has the longest feature catalogue.
| Alternative | Supplied evidence | When to consider it |
|---|---|---|
| MobaXterm | Its feature page documents remote sessions and a graphical SFTP browser; a user in #227 wanted its output-logging behavior | When session output is part of your configuration-recovery routine |
| mRemoteNG | Its README describes a tabbed remote connection manager; the #227 user preferred its interface and cited its RDP implementation | When a conventional tabbed connection manager is your main requirement |
| Netcatty | The macOS commenter in #1426 switched back after AI busy messages and input lag | When responsive command entry is the immediate reason for comparing another application |
| Conduit | TechYorker positions it around iPhone/iPad access and team sharing | When those requirements drive further comparison of the vendors' documentation |
The historical RDP discussion also includes an NLA workaround for servers where NLA was already disabled. Preserve that condition when reading it: the maintainer was explaining credential retention in a v1.4.1 workflow. For the present shortlist, use the official product descriptions to identify candidates and the attributed user reports to identify tasks worth comparing, rather than adopting a historical configuration change as general setup advice.
What the other reviews get wrong
Android downloads and a Docker-hosted Web service appear in uniTerm's official instructions, correcting omissions in two supplied reviews. iTechGuides says no mobile platform is provided. TechYorker marks Web and self-hosted availability as not listed. The installation guide supplies the Android APK route and describes isolated browser-user workspaces.
The supplied homepage snapshot displays v1.10.0 and advertises nine languages; the introduction lists 11, and the supplied release page identifies v1.12.3. The introduction's blanket reassurance about avoiding data loss also needs the recovery explanation from #1383. A missing decryption key can leave encrypted records present but unreadable. That is the relevant correction for someone deciding whether uniTerm should hold their only credential copy.
The X record promotes protocol breadth and the AI agent. The supplied YouTube overview describes an interface prototype and safety practices. These records explain how readers may encounter the project. The source run and named issue histories provide the operational detail used here: what command failed, what a reporter experienced and what a maintainer changed or proposed.
FAQ
The maintainer says encrypted credentials were unrecoverable in #1383 after their sole keychain key disappeared. That recovery question, MCP access and reported terminal failures are addressed briefly below. Each question comes from the supplied synthesis or search intents; the answers point to the relevant source so readers can move directly to the investigation or instructions they need.
Can encrypted credentials be recovered after the system keychain entry disappears?
For the keychain-mode incident in #1383, the maintainer's answer was no. Resetting and re-entering passwords restored usability while retaining other connection fields. The dedicated recovery section above explains the keychain and master-password distinction.
Can MCP access be restricted to individual saved connections and separate read-only commands from commands that change things?
The MCP guide documents policies for command approvals, including write-command confirmation. All tools are available by default. The approval dialog shows the target connection and full command, letting the operator accept or deny the requested action.
How can I keep an agent testing a server away from production connections?
Our recommendation is a separate evaluation account or instance containing only test connections and credentials, with command confirmations enabled. The guide describes discovery of saved SSH connections, so choose what the evaluation instance holds before connecting the agent.
Which uniTerm project is this review about, and where is its GitHub repository?
This review covers ys-ll/uniterm, the remote-access application. The supplied search intents also include indexing and commercial names. The separate uniterm.dev project describes a Rust terminal multiplexer and agent-fleet supervisor; use the repository owner to distinguish the software.
Is ys-ll/uniterm related to Uniterm indexing, Uniterm Direct or Uniterm Nigeria Limited?
The entity reviewed here is the ys-ll/uniterm software repository. Uniterm indexing is a subject-indexing system. The other names occur in the supplied search-intent record and are unnecessary for locating this application's repository or official guide.
Why does uniTerm crash even when it is sitting idle?
Users in #1426 reported idle crashes, and added logging exposed callback exhaustion. The maintainer's v1.12.3 notes list a callback-leak fix. Match the actual crash log to that error when using this investigation.
Why does the AI console stay busy or become slow after a longer conversation?
A macOS commenter reported those symptoms in #1426. The v1.12.3 release notes describe streaming-rendering throttling and a change to foreground executions that leave sessions busy. These are the maintainer's stated changes for the relevant behaviors.
Why does the AI stop working after its first command or report that the context budget is exceeded?
A contributor in #1307 proposed discarded parallel tool calls and unfinished context groups as a possibly related cause. The contributor linked PR #1312 and requested session-log evidence showing tool calls without corresponding results.
Does uniTerm automatically force upgrades, and can an upgrade destroy saved credentials?
The maintainer says upgrades require a manual click in #1383. The investigation found the missing key while the reporter still ran v1.9.0. The documented failure concerns loss of that keychain entry.
Under what conditions can WebView2 trigger failed Windows logins and account lockouts?
The #957 investigation concerns WebView2 Runtime 153.0.4234.x and a strict local Windows lockout policy. The maintainer reported a locally successful mitigation in snapshot 3dbd92d after an earlier snapshot failed.
Why does rapid typing lose characters on macOS?
The #146 discussion records missing input, a configuration-sync workaround for the original reporter and a later recurrence in v1.7.0. Those are the thread's concrete outcomes; use the affected version and input pattern when comparing a similar symptom.
Why are uppercase characters duplicated when using a Chinese input method on macOS?
The #655 discussion records duplicate input on macOS v1.8.0. A commenter said switching windows helped, and the maintainer offered snapshot fdfb06f. The thread identifies both the input-method context and the proposed build to examine.
What we have not tested yet
Our evidence covers repository commands, documentation and attributed reports; it does not include a completed desktop, Android or Web workflow. The sandbox image label does not establish actual runtime versions. The failing test tail supplies no assertion messages, so it cannot identify an environment, network or encryption cause. The summary counts and six named failing functions also use different levels of reporting.
The separate November 18 run checked out c9a98753cf73 in a fresh, unprivileged golang:1.24-bookworm container with 2 CPUs, 4 GB RAM, no secrets and no published ports. The clone step took 2 seconds. Its quickstart-cli recipe selected a generic Go package-install route rather than the project's documented packaged-client installation.
The installation step failed in 15.4 seconds with exit 1 at go install github.com/ys-ll/uniterm/...@latest. It downloaded v1.12.3, reported a requirement of Go 1.26.2 or newer and switched to go1.26.11. The failure record identifies module replace directives as the stopping condition. The teardown record reports successful cleanup. This probe supplies no desktop-reliability result.
Credential restoration across machines, current SFTP folder transfers, RDP authentication, Android installation, browser-user separation and current long-session AI behavior remain outside our completed checks. The MCP guide does not document a per-client saved-connection allowlist; action approvals alone do not establish host containment. Browser-user isolation likewise does not establish shared team vaults. The supplied material provides no current comparative reliability test of the alternatives.
The issue extracts lack reporter confirmation for the later Windows lockout mitigation and several offered snapshots, and the release-note fixes have no matching workflow retest here. The cause of the missing keychain entry remains uncertain. The recovery incident demonstrates neither sync-caused deletion nor a data leak. These are the outstanding boundaries on the credential, release and stability conclusions, rather than reasons to discard the recorded failures.
No screenshots or known-answer task results accompany these runs. Rendered layout, resolved links and site recommendation modules are outside this text review. Search volume, demand, adoption, staffed support and production uptime are unmeasured, and the supplied Bilibili results establish no relevant walkthrough. For an evaluation, start with a disposable server and recoverable credentials, then exercise transfers, streaming input and interrupted commands in the workflow you actually need.


