01 · Describe
Say what you need
The setup assistant names your Agent, picks a model, chooses tools, attaches knowledge bases and writes the prompt. Prefer Claude Code? Let it do the setup over MCP.
MCPSet up Agents from your coding agent
From a single sentenceto a liveAgent
The setup assistant builds your Agent. Connect it to knowledge bases and tools, publish it to the web, an API and LINE, then keep improving it with evaluations and traces.
From description to production
One platform for the whole life of an Agent
02 · Connect
Plug in knowledge and tools
Upload documents or crawl a whole website into a knowledge base. Connect MCP servers, your own APIs, or ready-made skills from the marketplace.
03 · Publish
Go live on every channel
Share a public link, embed it on your site, call the Agent API, or connect LINE, Slack, WhatsApp, Telegram and Discord.
04 · Improve
See it, then fix it
Evaluate answers against a golden set, open the trace of any turn to see what it searched and how long it took, and let knowledge gaps tell you which documents are missing.
01 / 05
One knowledge base, every document
Upload PDFs, Word, Excel and slides, or crawl a website. This HR knowledge base holds 18 documents, and your Agent searches all of them.
02 / 05
Split into passages, turned into vectors
Each document is split into small passages, and an embedding model turns each one into a vector, a list of numbers. Passages that mean similar things end up close together, whichever file they came from.
03 / 05
The question becomes a vector too
The same model turns the question into a vector, and it lands next to the passages closest in meaning.
04 / 05
Searched by meaning and by keyword
The search reaches out from the question for passages close in meaning, and also matches the question’s keywords directly, so even a distant passage turns up.
05 / 05
The five most relevant passages
Both sets of results are merged, and the five most relevant passages go to your Agent. Its answer shows which document each one came from.
Try another question
Knowledge
Answers grounded in your own documents
Upload files or crawl a whole website, and Mizar keeps the knowledge base in sync. Hybrid search and reranking find the right passage, and every answer can show where it came from.
- Upload documentsPDF, Word, PowerPoint, Excel, CSV, Markdown or images, or paste text directly.
- Crawl a websiteFollow a sitemap or a path, then re-fetch daily or weekly.
- Tune retrievalVector, full-text or hybrid search, with reranking and a score threshold.
Memory
It remembers who it’s talking to
Turn on memory and your Agent keeps what signed-in users tell it: who they are, what they’re working on and how they like their answers. Every Agent with memory on shares it, so nobody has to repeat themselves.
- A person profile sums up each user, with every version kept.
- See which Agent noted each memory and how often it has been used.
- Admins can review or delete any memory, and users can ask it to forget.
Channels
One Agent, everywhere your people are
Build it once. The same Agent, knowledge and settings serve a public link, your website, an API and the messaging apps your people already use, and every conversation lands in the same logs.
Evaluation & monitoring
See every answer. Know what to fix.
Test every change before it goes live, see exactly what each turn did once it has, and get the questions it couldn’t answer sorted into topics.
Evaluation
Grade answers against reference answers, key facts and expected sources, three times per question.
Traces
Every turn, step by step: what it looked up, which tools it called and how long each part took.
Knowledge gaps
Questions your Agent couldn’t answer from its knowledge, grouped so you know what to write next.
Tasks & teams
Hand off work that takes more than one answer
Put longer work on a board. Agents pick it up, split it into steps, report after every run, and a manager Agent can delegate to its team. Anything that reaches outside waits for your approval.
- Approve the plan before it starts, or an action once and always allow it.
- A reviewer sends weak work back, and asks you after three tries.
- Automations start tasks on a schedule or when another system calls.
Enterprise
Built for the way your organization runs
Sign in with your company accounts, decide who can change each Agent, and keep a record of everything that happens. Set the rules and the budget, choose the models, and run it on your own servers when you need to.
Sign-in and roles
Single sign-on with your company account, admins and members, and per-Agent access for who can use or manage it.
Activity log
Every settings change and every run is recorded. Entries can’t be edited or deleted.
Content moderation
Check questions and answers against keywords or topics, then reply with set text or hand off to a person.
Spending limit
A monthly cap per Agent, with alerts at 80% and 100%.
Any model
OpenAI, Anthropic, Google and xAI, or open models you serve yourself.
Run it on your own servers
One Docker image for the whole platform. Bring it up with Docker Compose next to a database, cache and file storage, choose your vector database, and point it at the models you host.
Developers
Bring your coding agent, or call the API
Give Claude Code or another coding agent a developer key and it builds Agents, knowledge bases and skills for you over MCP. Your own systems get answers from the Agent API.
- Same permissions and audit trail as the admin site. Issue a read-only key when it only needs to look.
- Pass your own user IDs to keep each end user’s conversation separate.
- Stream the answer as server-sent events, or get one JSON response.
$ claude mcp add --transport http mizar \ https://<your-deployment>/api/v1/mcp \ --header "Authorization: Bearer mzr_…" > Build an HR help desk Agent from the files in ./handbook ✓ create_knowledge_base Employee Handbook 2026 ✓ create_knowledge_upload 18 files ✓ create_agent employee-help-desk ✓ set_agent_assets knowledge, tools ✓ test_agent "Can I take half a day of leave?"
FAQ
Questions teams ask before they start
Do I need to write code?
No. Describe the Agent you want in the admin site and the setup assistant picks a model, connects tools and knowledge bases and writes the prompt. You can change any of it yourself afterwards. Developers can also work through MCP or the Agent API.
What can I put in a knowledge base?
PDF, Word, Excel, PowerPoint, CSV, Markdown, plain text and images, or text you paste in. Scanned PDFs are read with text recognition. You can also crawl a website and re-sync it daily or weekly.
Does it show where an answer came from?
Yes. Answers cite the passages they used, and each citation opens the original file. Questions it couldn’t answer are grouped under knowledge gaps, so you know which document to add.
Which models can I use?
OpenAI, Anthropic, Google and xAI, or open models you serve yourself behind an OpenAI-compatible endpoint such as vLLM. Each Agent can use a different model.
Which languages does it answer in?
By default it replies in the language of the user’s latest message. You can also fix an Agent to Traditional Chinese, English, Japanese, Korean, Indonesian, Vietnamese or Thai. The admin site is available in Traditional Chinese, English and Japanese.
Where can people talk to it?
A public link, a chat button or iframe on your website, the Agent API, and LINE, Slack, WhatsApp, Telegram and Discord. Every conversation ends up in the same logs.
Can our data stay in our own environment?
Yes. The whole platform ships as one Docker image you can run on your own servers, and it can use models you host yourself.
Who can use and manage Agents?
People sign in with Google or your company’s single sign-on, such as Microsoft Entra. Admins invite members and decide, Agent by Agent, who can use it and who can manage it. Every settings change and run is kept in the activity log.
Can conversations and memories be deleted?
Users can delete their own conversations, and admins can delete any conversation or memory, including everything one person’s Agents remember about them. Prompt and tool details in traces are kept for 30 days.