mrkeyoor.com_
Wed 16 Sept 20:07 UTC
AI Toolsevaluationupdated 27 Aug 2026

lobehub review

LobeHub is an agent workspace for creating AI assistants, organizing them into groups and projects, scheduling work, and connecting them to models, messaging channels, files, and MCP-compatible tools. Its primary README is in English, and a Simplified Chinese README is also available.

+173stars / 7d
Verdict

Our LobeHub install spent 296 seconds before npm rejected workspace:*, so the first decision is whether your team will follow its pnpm workflow and accept a 198.4 MB monorepo. Trial it if agent groups, scheduled work, messaging, memory, and MCP tools genuinely belong in one workspace. Walk away if you need a small chat client, an ordinary Apache-2.0 license, or a self-hosted system without local agent execution and frequent platform change.

We ran it

Lab card: what happened when we ran lobehubScreenshot of lobehub (lobehub.com)
Install✗ · 296s
Build
Repo15260 files~2,097,297 lines of source · 198.4 MB · 36 CI workflows · Dockerfile · tests dir

Answers from our run

Does lobehub build from source?

The dependency install failed, and the project has no separate build step. We cloned commit 19f9107 into a clean Debian container with 3 CPUs and no project-specific setup.

Who should not use lobehub?

Anyone using npm install without checking the repository's package-manager instructions: our run failed on the workspace:* protocol, while the README tells contributors to use pnpm.

What are the alternatives to lobehub?

LibreChat, Open WebUI, Dify. Our LobeHub install spent 296 seconds before npm rejected workspace:*, so the first decision is whether your team will follow its pnpm workflow and accept a 198.

Setup1/5npm failed after 296 seconds on workspace protocol
Docs3/5Broad feature and hosting guide, but setup advice needs scrutiny
Community5/582,024 stars and active August 2026 development
Maturity3/5v2.2.14 is active; major agent features remain fast-moving

Discussed on

  1. hnLobe Chat: open-source, modern design AI chat framework3 points

Who it’s for

Individuals who want several model providers, agents, memory, files, and scheduled tasks in one polished interface.
Teams experimenting with agent groups, shared workspaces, messaging bots, and human approval around automated runs.
Self-hosters prepared to manage a database, model credentials, sandbox execution, storage, upgrades, and network exposure.
Developers who want MCP-compatible plugins and an API rather than a chat window alone.

Who it’s NOT for

Anyone using npm install without checking the repository's package-manager instructions: our run failed on the workspace:* protocol, while the README tells contributors to use pnpm.
Companies planning a derivative commercial distribution without a license review: the LobeHub Community License requires a commercial license for that use.
Security-sensitive teams unwilling to operate agent code execution: v2.2.14 adds a local sandbox with a real working directory and also supports an E2B-style cloud path.
Minimal chat users: the checkout held 15,260 files and about 2,097,297 source lines before dependencies.
Operators who will run an unaudited remote setup script: the Docker quick start pipes https://lobe.li/setup.sh into Bash.

Setup reality

Our sandbox install failed after 296 seconds with exit code 1. Npm reported EUNSUPPORTEDPROTOCOL and Unsupported URL Type "workspace:": workspace:*, including after a forced retry. Because installation failed, no build or tests ran.

The README's contributor path uses pnpm, so our npm result is a package-manager mismatch finding, not proof that the documented pnpm path fails. A useful instance needs model-provider credentials; the self-hosted stack also brings database, storage, authentication, messaging, plugin, and sandbox configuration.

The Docker path creates a directory, runs a remote setup script through Bash, then starts Compose. Review and pin that script before execution. Release 2.2.14 adds local code execution, scoped API keys, agents across CLI runtimes, messaging gateways, projects, goals, and memory, all of which widen the data and permission model.

LobeHub has grown from model chat into agent operations

The product now treats an agent as the main unit of work. Users create agents, assign models and tools, group them, organize work into projects, schedule tasks, and receive reports. Personal memory and shared workspaces keep context beyond one chat. Messaging gateways bring agents into existing channels. This makes LobeHub closer to an agent operating console than a themed front end for one model API.

Release v2.2.14, published August 16, 2026, adds projects and goals in alpha, 7 more CLI coding-agent runtimes, a local sandbox, a generated OpenAPI specification, scoped API keys, and workspace skills. It also introduces home customization and desktop split views. The volume matters: each feature adds permissions, persisted state, or another external runtime that a self-hoster must understand.

The 198.4 MB repository expects pnpm, while our npm install failed

Our checkout at commit 19f9107 contained 15,260 files, about 2,097,297 lines of source, and 198.4 MB. The lab identified a Node npm path and attempted installation in an unprivileged container with 3 CPUs and 8 GB of RAM. That choice did not match the README's local development commands, which explicitly say pnpm install.

After 296 seconds, npm exited with code 1. The log repeatedly reported EUNSUPPORTEDPROTOCOL and Unsupported URL Type "workspace:": workspace:*. A forced retry produced the same protocol error. The log does not show a missing compiler, network failure, or package conflict, so we should not name one. It shows that npm could not process this workspace dependency form in the measured run.

What happened when we ran it

Our 296-second install failed before dependencies were ready. No build ran, and no tests ran. That boundary is important: there is no lab evidence here about compiled output, application startup, or test health. The checkout had 36 CI workflow files, a Dockerfile, a tests directory, and monorepo workspaces, but those repository signals do not turn an aborted install into a passing check.

The practical next attempt is the documented pnpm path, not another npm flag. Contributors run pnpm dev for the full-stack application or bun run dev:spa for the SPA on port 9876. The README's debug proxy sends that local SPA through a production backend URL. Teams should decide whether sending local development requests toward a production-hosted service is acceptable before using that shortcut with private data.

Self-hosting starts with a remote script and expands into several services

The Docker quick start creates a storage folder, pipes https://lobe.li/setup.sh into Bash, then runs Docker Compose. Piping a changing network response into a shell gives the remote endpoint immediate code-execution authority. Download the script, inspect it, pin a known revision or digest where possible, and keep the resulting Compose and environment files under internal review.

Model use requires at least one provider key or compatible endpoint. A production workspace also needs persistent database and file storage, authentication, backups, TLS, outbound policy, and secrets for messaging or plugins. Version 2.2.14 adds API-key scopes and hardening around provider base URLs, restricted knowledge bases, connector deletion, and auth cookies. Those security changes are welcome evidence that the team works on boundaries, while also showing how many boundaries exist.

Local sandbox execution changes the threat model

The release notes describe a local sandbox with a real working directory and integrations for several coding-agent CLIs. Published agents can run tasks while the user is away, and goal launch still requires confirmation even in automatic mode. That confirmation is a sensible safeguard. It does not replace operating-system isolation, workspace restrictions, command policy, network controls, or audit logs around an agent that can execute code.

Treat every plugin, MCP server, model provider, CLI agent, and messaging connector as a separate trust decision. The README advertises more than 10,000 skills and MCP-compatible plugins. Catalog size is not a security review. Start with a small allowlist, use narrowly scoped credentials, keep execution away from production secrets, and test whether scheduled or remote-triggered agents can reach files and services outside their intended project.

The community license permits use but restricts derivative distribution

GitHub could not assign a standard SPDX license identifier. The repository license says it is based on Apache 2.0, permits commercial use of LobeChat as an unmodified frontend and backend service, and requires a commercial license to develop and distribute a derivative work. Contributors also agree that the producer can change the agreement and use contributed code commercially. Legal review belongs before a company builds a product around a fork.

LibreChat's MIT license may be simpler when multi-provider chat, agents, and MCP cover the requirement. Open WebUI is a more direct match for local-model conversation. LobeHub earns a trial when agent scheduling, groups, projects, memory, and messaging all reduce separate tooling. Its 296-second npm failure is recoverable through the documented pnpm path, but the larger commitment is permission design across a fast-moving agent platform.

August activity is intense enough to demand staged upgrades

GitHub showed 82,024 stars, 818 combined issues and pull requests, and a last push on August 27, 2026. The default branch is canary, and v2.2.14 says 363 pull requests from 21 contributors landed since the previous release. This is active maintenance, and it is also a lot of change to absorb between self-hosted upgrades.

Pin releases instead of deploying a moving branch. Back up the database and files, compare environment changes, test model calls and messaging, and verify sandbox restrictions on a staging instance. The product's breadth is its reason to exist, but it also means a regression can land in chat, schedules, memory, plugins, storage, or execution. Teams that cannot rehearse those paths should use the hosted service or choose a narrower client.

Alternatives

ProjectWhat it isPick it when
LibreChat gh↗A self-hosted multi-provider chat application with agents, MCP, tools, and multiuser authentication.pick this instead when a familiar chat interface and MIT licensing matter more than LobeHub's operator and scheduling model.
Open WebUI gh↗A self-hosted model interface commonly paired with Ollama and OpenAI-compatible endpoints.pick this instead when local-model chat is the main job and agent-team orchestration is secondary.
Dify gh↗A visual platform for LLM workflows, retrieval applications, agents, and production APIs.pick this instead when application workflows and RAG pipelines matter more than a personal agent workspace.

Sources

  1. LobeHub README
  2. LobeHub repository facts
  3. LobeHub v2.2.14 release
  4. LobeHub Community License
  5. LobeHub open issues and pull requests

More ai tools reviews

eve · MemOS · LongCat-Video · Concat · DLSS5-Feeder · Concat · the whole board →