G2

Luca P.

Chief Operations Officer DEQUA Studio | Formerly CTO in MarTech

Marketing and Advertising

Mid-Market (51-1000 emp.)

6/18/2026

"One Interface to Switch Models, Agents, Plugins, and Knowledge Base—Self-Hosted Done Right"

4/5

What do you like best about LobeChat?

The model switcher is the part I touch every working day. I run LobeChat self-hosted with my own provider keys, and a dropdown at the top of the conversation lets me move between OpenAI, Claude, Gemini, and a local model running through Ollama without leaving the thread I am in. The practical effect is that I draft with a cheap, fast model and then switch to a stronger one for the pass that actually matters, all inside the same conversation. I stopped paying for three separate chat subscriptions and keeping three browser tabs open to compare them. One interface, my keys, and whichever model fits the task in front of me.

The agent marketplace earned its place faster than I expected. I can install a pre-built assistant, fork it, change its system prompt and its default model, and save it under a name that means something to me. I keep a small library of these: a code-review agent pinned to a stronger model, a quick-rewrite agent sitting on something cheap, a SQL helper with the table context already baked into its prompt. Rather than pasting the same three-paragraph system prompt every time I want a particular behavior, I pick the agent and start typing. The marketplace itself is a reasonable starting point, but the real value for me is treating those community agents as templates I bend to my own work instead of building each one from a blank page.

Plugins and the MCP marketplace are useful, with a caveat I will get to. The plugin system runs on function calling, so an agent can reach out to a tool mid-conversation to pull a document, hit an API, or generate an image. The MCP marketplace integration matters more to me than the older plugin catalog, because it means I can point LobeChat at the same MCP servers I already use elsewhere rather than waiting for someone to ship a bespoke plugin. When a tool is wired in correctly, it is the difference between a chat window and something that can actually carry out a step of real work. Not every plugin is maintained to the same standard, which I will come back to in the dislikes, but the handful I depend on are dependable.

The knowledge base quietly changed how I use the tool for real work. I upload a set of internal documents, and retrieval pulls the relevant chunks into context when I ask a question, so I am querying our own material instead of pasting it in by hand every time. On the self-hosted server version this rides on Postgres with the pgvector extension, which means the documents and their embeddings sit on infrastructure I control rather than in some third party's store. For anything that touches internal information, that distinction is the entire reason I chose to host it myself.

Artifacts and the rendering are a smaller pleasure that adds up over a day. Code comes back in a side panel with proper highlighting, and markdown, Mermaid diagrams, and LaTeX render in place rather than arriving as raw text I have to copy somewhere else to read. When I am working through something that has a diagram or a formula in it, seeing it rendered inline keeps me moving instead of bouncing out to a separate viewer. It is not the feature that sold me, but it is the kind of detail I would miss the moment I went back to a plainer client.

Multi-modal input is more useful in practice than I assumed it would be. I can drop an image into a conversation and have a vision-capable model describe or reason about it, generate an image from a prompt when I need a quick visual, and use speech in and out on the occasions when typing is not convenient. None of this is unique to LobeChat taken feature by feature, but having vision, image generation, and voice sitting in the same place as the text models, rather than spread across three more tabs, is what makes me actually use them. A screenshot of a broken layout or an error dialog goes straight into the chat instead of being routed through a separate tool first, which is usually the moment I would have given up and just described it in words.

Running local models through Ollama is an option I am glad exists even though I do not reach for it constantly. For a prompt that involves anything I would rather not send to a hosted provider, I switch the conversation to a local model and the request never leaves my own hardware. Having that escape route in the same interface, one dropdown away from the cloud models, means privacy is a per-conversation choice rather than a separate tool I have to remember to open.

The interface itself is calm and quick, and it holds up across the desktop app and the browser PWA, including on my phone when I am away from my desk. What makes me comfortable building a workflow on it is that the project keeps moving. It picked up native support for Claude's SKILL.md skills format not long after that standard appeared, which is the sort of thing that tells me the people behind it are tracking where the field is going rather than coasting on what already shipped. Review collected by and hosted on G2.com.

What do you dislike about LobeChat?

The first hour is harder than it should be for anyone non-technical. I handed the tool to a teammate who is not a developer and watched them get lost between agents, plugins, providers, and API keys, none of which they needed to think about to send a first message. The settings run deep, and the surface area you see before you have done anything is larger than the task of "just chat with a model" warrants. The workaround I have settled on is to set people up once myself, with a couple of agents already configured and a sensible default model chosen, so their daily experience is simple even though the initial setup is not something I would leave them to do alone.

The jump from the simple deployment to the full self-hosted server version is a real step up in effort. A basic Docker container with local storage is genuinely a few minutes of work. The moment you want the knowledge base, cross-device sync, and persistent accounts, though, you are standing up Postgres with pgvector, a Redis instance, and S3-compatible object storage, and then thinking about backups and secrets for all of it. It is well documented and it does work, but it is a small piece of infrastructure to own rather than a one-click affair, and I would not hand that part to anyone who does not already run services for a living.

If you go the hosted route instead of self-hosting, the credit-based pricing is harder to predict than a flat subscription. The cloud plans bill in compute credits tied to token usage, and a heavy day of long contexts and retrieval can burn through them faster than you would guess from the headline number. I evaluated the cloud version before committing and ended up self-hosting partly for this reason, because pay-per-token on my own provider keys at least gives me line-item visibility into where the spend is going.

Enterprise controls are lighter than a tool built squarely for large organizations. SSO and OAuth are present, and for a small team that is plenty. Org-level administration and audit logging, though, are thinner than what a compliance team would want to see, so for a regulated environment this is the area to check carefully before standardizing on it. It is a question of intended scale rather than a flaw, but it is worth knowing going in.

The last one is a mix of the tool and the models sitting behind it. Every so often a single response will contradict itself or drift from the format I asked for, and a long retrieval query against a large knowledge base adds latency you can feel. A fair amount of that is model behavior rather than anything LobeChat is doing wrong, but it surfaces inside LobeChat so it is part of the day-to-day experience regardless. The plugin ecosystem is also still maturing, and a couple of the plugins I tried early on were plainly unmaintained, so I have learned to stick with the ones that show active upkeep and not assume everything in the catalog still works. Review collected by and hosted on G2.com.

What problems is LobeChat solving and how is that benefiting you?

The core thing it fixed was the sprawl of separate AI tools. Before, I had a ChatGPT subscription, a Claude one, a tab open for Gemini, and a separate flow whenever I wanted a local model, and comparing them meant copying the same prompt into each in turn. Now there is one interface, I bring my own provider keys, and changing model is a dropdown inside the conversation I already have open. The benefit is not only fewer subscriptions to manage. It is that picking the right model for a given task stopped being a context switch and became a single click, which means I actually do it instead of defaulting to whatever was already on screen.

Cost control followed directly from that. Routing drafts and throwaway questions to a cheap model, and saving the expensive one for the work that genuinely needs it, is trivial when both live behind the same switcher. Because I am on pay-per-token with my own keys rather than a bundle of fixed monthly plans, I can see what each provider is actually costing me and adjust which model my default agents point at. The spend tracks real usage instead of sitting as a stack of subscriptions I am paying for whether I touch them or not.

Keeping sensitive material in-house is the next problem it handles. Between self-hosting the application and being able to run local models through Ollama, prompts and documents that should not leave our network do not have to. The knowledge base storing its embeddings in our own Postgres instance is part of the same picture. For the work where data handling is the constraint, having the whole path stay on infrastructure I control is the difference between being able to use an AI tool for it at all and having to keep that work manual.

Standardizing AI access across the team got much easier. Instead of everyone expensing an individual chat subscription and settling on whatever they happened to land on, we run a shared self-hosted instance with central key management, so the available models and the agent library are consistent across the group and I am not chasing reimbursements for a dozen separate personal plans. Onboarding someone to our AI setup now means giving them access to one place that is already configured, not walking them through signing up for three different services.

Repeated prompt setups stopped living in my head and a scratch file. The behaviors I reach for again and again are codified as named agents, each with its own system prompt and default model. The before-state was re-pasting a long system prompt and re-selecting the right model every single time I wanted a specific kind of help. The after-state is picking the agent from the list and starting, which sounds minor until you count how many times a day it happens.

Querying our own documents in context is the last habit it changed. Rather than pasting a reference document into the prompt by hand, or skimming it myself to find the one paragraph that matters, I ask the question and let retrieval pull the relevant section into the conversation. Because the whole thing is self-hosted, that material stays on our own infrastructure while still being usable in a chat. It turned our internal documents from something I had to manually feed the model into something the model can reach when a question calls for it.

Not being tied to one model vendor is a quieter benefit that pays off over time. New models ship at a constant pace, and because LobeChat sits in front of many providers, adopting a new one is a matter of adding a key rather than migrating to a different tool. I am not betting my whole workflow on a single company's roadmap, which, given how quickly this space moves, is worth more than any one feature on the list. Review collected by and hosted on G2.com.