OpenMake LLM v1.5.6

User Manual

Welcome to the OpenMake LLM User Manual. This guide covers every feature of the platform with detailed, step-by-step instructions so you can get the most out of your self-hosted AI assistant.

OpenMake LLM is a privacy-first, self-hosted AI platform centered on a local vLLM/LiteLLM gateway: autonomous agent tasks in a Docker sandbox, 22 built-in MCP tools, deep research, artifacts, vision, cross-conversation memory, bring-your-own-key external providers, and 100 specialized agents across 18 industry domains.

💡Navigation

Use the table of contents on the left (visible on desktop) to jump to any section. On mobile, simply scroll through the content — each section is clearly labeled.

🚀

Live Demo

https://chat.openmake.cc

💻

GitHub

https://github.com/openmake/openmake_llm

📜

API DocsLogin Required

https://chat.openmake.cc/api-docs

🧭 Finding Your Way

The sidebar holds only five entries so that your recent conversations stay the focus. Everything else is still there — reached through a hub tab, the profile menu at the bottom of the sidebar, or a toggle in the composer. Read this section first and the rest of the manual will be easy to follow.

The sidebar

Where everything else lives

If you are looking for something that used to be its own sidebar entry, this is where it moved. The old URLs still work — three of them redirect into a settings tab.

💬 Chat

The Chat page is the main interface of OpenMake LLM, and the page you land on at the root URL. Send messages, attach images, and reach everything the composer offers: eight mode toggles, the tool picker, the model selector, and the tool-approval control.

Getting Started

1. Open the application — the Chat page loads by default at the root URL (/).
2. Use the configured model route. By default, chat falls back from OMK_CHAT_MODEL to LLM_DEFAULT_MODEL.
3. Type your message in the input box at the bottom of the screen.
4. Press Enter to send. Use Shift+Enter to add a new line without sending.
5. The AI response streams in real-time. Click the Stop button to abort generation mid-stream.

The eight mode toggles

Open the mode sheet from the composer to switch these on. They fall into two kinds, and the difference matters \u2014 see the warning below the table.

The rest of the composer

Tool approval

The Approval control in the composer decides when a tool call pauses for you. It applies to the current conversation.

✨ Prompt Modes

Twelve role prompts sit behind the chat, each tuned for a different kind of request. Detection is entirely automatic — there is no prompt-mode picker anywhere in the interface, and nothing to configure.

Basic Modes

Development Modes

Content Modes

Analysis & Agent Modes

🎯 Built-In Specialists (100)

OpenMake LLM ships with 100 specialist agents across 18 industry domains, plus a general agent that catches everything else. You never pick one. Every message is routed to the best-matching specialist automatically, and that specialist's expertise is folded into the system prompt before the model answers.

Technology / IT · 11

Legal · 5

Science / Research · 7

Business / Management · 9

The 18 Domains

Domain names and specialist names are Korean in the catalog, which is what you will see in the Skill Library. Counts are exact.

How a Specialist Gets Picked

Four tiers, tried in order. The cheap ones come first so most messages never pay for an extra model call.

1. Cache — if you have asked the same question before, the previous routing decision is reused straight away.
2. Keyword pre-classification — a keyword router scores every agent, and if its confidence clears the threshold that agent is taken and the LLM step is skipped entirely.
3. Short-message shortcut — a short message with no domain keyword at all goes straight to the general agent, which is what handles small talk and arithmetic.
4. LLM routing — only if none of the above resolved it. If that call fails, the keyword result from tier 2 is reused rather than routing again.

Seeing What a Specialist Knows

On every server start the platform generates a professional skill document for each of the 100 agents and upserts it into the skill table as system-skill-{agentId}, published so anyone can read it. That means the Skill Library is the browsable view of this catalog — open it to see exactly what guidance a given specialist carries. The seeding is idempotent, so adding an agent to the catalog is enough for it to appear after a restart.

🤖 Model Runtime

OpenMake LLM uses a local OpenAI-compatible gateway for the current runtime. Chat, router, vision, embedding, and image workflows are selected from the configured local-llm catalog. Four external providers can be added on your own key — see External Providers below.

🌐 External Providers

The local catalog needs no key, but four external providers ship in the catalog and can be added with your own credentials. Every one of them is reached over an OpenAI-compatible endpoint, so a registered model behaves like a local one: it appears in the model chooser, it can be assigned to a role, and its usage is metered.

1. Open Settings and select the Model & Response tab.
2. Under API Keys, choose Add Key and pick a provider from the catalog.
3. Paste the key. Override the base URL only if you self-host that provider.
4. The key is validated against the provider's /models endpoint on save, then stored AES-256-GCM encrypted.
5. The provider's models now appear in the model chooser as <provider>:<model> and in role assignment.

🔧 MCP Tools

The MCP Tools page shows all available Model Context Protocol tools. These are capabilities the AI can use during conversations — web search, code execution, file analysis, and more.

The 22 Built-in Tools

How to Use

1. All 22 built-in tools are available to every user — subscription and tool-tier gating were removed.
2. Tools are automatically invoked by the AI when relevant during a conversation, up to 5 tool turns per request.
3. Install additional servers from the MCP catalog in Settings → Connectors.
4. A catalog-level tool allowlist keeps chat auto-exposure focused, so a large server does not dump every schema into the prompt.
5. REST execution and the explicit tool picker still reach the full tool set.

📂 Chat Attachments

File analysis is handled inside chat workflows in the current UI. Upload images or supported files from chat, then let the configured model route analyze the content with the active tools.

How to Use

1. Open Chat and use the attachment or image upload control.
2. Attach the file or image you want to analyze.
3. Describe the task, such as extraction, summary, translation, or visual inspection.
4. Use MCP tools or Deep Research when the analysis also needs external sources.
5. Manage resulting conversations from History and data controls from Settings.

📜 History

The History page displays all your past conversation sessions. You can search, resume, or delete conversations.

How to Use

1. Open History from Recent conversations → See all in the sidebar, or from the tab bar on the Deep Research page.
2. Browse your past sessions listed by date and topic.
3. Click on any session to view the full conversation.
4. Click "Continue" to resume a conversation where you left off — full context is restored.
5. Use the search bar to filter sessions by keyword.
6. Click the trash icon to delete a session permanently.

🔬 Deep Research

Deep Research is an autonomous multi-step research agent. Enter a topic and the AI will decompose it into sub-questions, search the web, collect sources, and generate a comprehensive research report.

How to Use

1. Turn on the Deep Research toggle in the composer, or open the Deep Research page from the History tab bar — it is not a sidebar entry.
2. Enter your research topic in the input field (e.g., "Comparison of RAG vs Fine-tuning for LLM applications").
3. Select research depth: Quick (fast overview), Standard (balanced), or Deep (comprehensive).
4. Click "Start Research" to begin.
5. Watch the real-time progress: topic decomposition → web search → source collection → synthesis.
6. The progress bar shows completion percentage. Each sub-step is displayed as it happens.
7. Once complete, view the full research report with citations and sources.
8. Previous research sessions are listed below for reference.

🛠️ Agent Tasks

An Agent Task is an autonomous run: the agent pursues a goal across multiple tool-calling turns inside a persistent Docker sandbox with shell, Python, browser, file, and planning tools. Progress, token accounting, and every step are recorded live.

How to Use

1. Switch the composer to Agent mode and describe the goal, or open Agent Tasks from the sidebar and create one directly.
2. Attach files or images — originals and extracted text are both written into the sandbox workspace under uploads/, and images are injected through the vision channel.
3. Watch the run: each step shows the tool call, its result, and the running token count.
4. Approve or reject high-risk steps when the agent pauses — python_execute, shell writes, and similar actions require human-in-the-loop confirmation.
5. Collect deliverables: the agent can produce Excel (.xlsx) and PDF files, including Korean/CJK fonts.
6. Save a finished task as a reusable template, or put it on a recurring schedule.

🤖 Custom Agents

Create and manage your own custom AI agents with specific system prompts, model-routing behavior, and tool configurations. Attach skills to agents for automatic prompt injection.

Creating an Agent

1. Open Custom Agents from the sidebar.
2. Create a new agent to open the form.
3. Fill in the five fields the form actually has: name, description, icon, model, and system prompt.
4. The system prompt is what defines the agent — this page has no temperature slider, no tool panel, and no test button.
5. Save. The agent is stored against your account and becomes selectable in chat.
6. Alternatively import an agent straight from a Git repository — give the repo URL, branch or tag, file path, and an access token for private repos.
7. Agents proposed by the model arrive as drafts: review them on this page and approve or reject each one.

📦 Skill Library

The Skill Library is where you create, manage, and share reusable knowledge blocks (skills) that can be attached to custom agents. Skills enrich agent prompts with domain expertise.

How to Use

1. Open Custom Agents from the sidebar, then switch to the Skill Library tab — it is not a sidebar entry of its own.
2. Browse the library. Every one of the 100 built-in specialists is published here as a skill, alongside anything you add.
3. Create a skill by hand with a name, description, and the knowledge content itself.
4. Or let the model draft one for you, upload a file, or import from a Git repository.
5. Assign a skill to yourself from the skill entry — assignment happens here, not in the agent editor.
6. Export a skill to share it, and review model-generated drafts before they go live.

📄 Artifacts

Artifacts are standalone outputs — pages, components, documents, diagrams — rendered beside the conversation in a sandboxed iframe and persisted so you can reopen or share them later.

How to Use

1. Ask for something self-contained, or turn on Artifact mode in the composer.
2. The artifact opens in a resizable side panel; the conversation stays on the left.
3. Run Python or JavaScript inside the artifact when code execution is enabled — execution happens in a Docker sandbox with a concurrency cap.
4. Open the Artifacts page from the sidebar to browse everything you have created.
5. Publish an artifact to get a share link served from a separate origin under a strict CSP.

⚙️ Settings

The Settings page gives you control over the application appearance, AI behavior, tools, and account management.

Available Settings

🧠 Memory & Data Controls

Cross-conversation memory lets the assistant carry facts between chats. Memories are stored per user and prepended to the system prompt, with a per-session privacy toggle to exclude them. Data controls cover conversation history, export, and deletion.

How to Use

1. Open Settings from the profile menu at the bottom of the sidebar, then the Memory tab (/memory redirects there).
2. Add a memory, or delete individual entries — a full forget deactivates all of them at once.
3. Turn the per-session memory toggle off when you want a conversation to ignore stored memories.
4. Export account data from Settings when you need a local copy.
5. Use History to resume or remove individual chat sessions.

📈 API Usage

The API Usage page shows your consumption statistics: request counts, token usage, error rates, and daily/weekly trends visualized in charts.

What You Can See

1. Summary stats for your own account: tokens this month, request count, average latency, and errors.
2. Cost and volume at a glance: today cost, today tokens, week tokens, and a projected weekly cost.
3. Daily token usage — a 14-day daily token chart.
4. System token monitoring — daily tokens over 7 days plus an hourly breakdown.
5. Usage share by model — which models consumed the tokens. The split is by model, not by workflow.

💻 API Documentation

The Developer Portal provides comprehensive API documentation for external integrations. It includes authentication methods, endpoint references, code examples, and rate limiting information.

Authentication Methods

Key API Endpoints

🔐 API Keys

The API Keys page lets you create and manage API keys for programmatic access to the OpenMake LLM API. Use these keys to integrate the platform into your own applications.

How to Use

1. Open the profile menu at the bottom of the sidebar → Developer, then switch to the API Access tab (/api-access).
2. Click "Generate New Key" to create a new API key.
3. Give the key a descriptive name (e.g., "My App Production Key").
4. Copy the generated key immediately — it is only shown once.
5. Use the key in your API requests via the X-API-Key header.
6. To rotate a key (generate a new one and invalidate the old), click "Rotate".
7. Delete keys you no longer need by clicking the trash icon.

Quick Start Example

curl -X POST http://localhost:52416/api/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer omk_live_xxxxxxxxxxxxxxxxxxxxxxxx" \
  -d '{"model": "qwen3.6-35b-a3b", "messages": [{"role": "user", "content": "Hello!"}]}'

👥 User Management

The User Management page is an admin-only dashboard for managing all platform users. View user activity, manage roles, and access user data.

How to Use

1. Open Admin from the sidebar — the Dashboard tab is user management.
2. View all registered users with their email, role, and last activity.
3. Click on a user to view their detailed profile and conversation history.
4. Change a user's role — the roles are user, admin, and guest; subscription tiers were removed.
5. Export user data for compliance or analysis.
6. View per-user usage statistics and chat history.

📈 Analytics Dashboard

The Analytics Dashboard provides comprehensive insights into platform usage, performance, and user behavior. Data auto-refreshes every 60 seconds.

Dashboard Panels

📊 Admin Metrics

The Admin Metrics page is a comprehensive monitoring dashboard showing system health, API token usage, quota status, cost tracking, and usage charts powered by Chart.js.

Dashboard Sections

🖥️ Runtime Monitoring

Runtime monitoring lets admins inspect node metadata, API metrics, model routing health, audits, and agent activity for the self-hosted deployment.

How to Use

1. Open Admin from the sidebar, then the MCP Monitoring tab.
2. Review node status, latency, model metadata, and API usage trends.
3. Use metrics and audit views to trace slow requests, tool calls, and agent activity.
4. Check external provider and local gateway health before changing model routing.
5. Runtime hosts are changed in server configuration, not in the UI — after changing them, verify health here.

🎓 Agent Learning

The Agent Learning page shows AI quality metrics based on user feedback. Monitor agent performance, identify failure patterns, and track improvement over time.

What You Can See

1. Agent quality scores per workflow and agent type.
2. Failure pattern analysis — common error categories and their frequency.
3. Improvement suggestions generated from feedback data.
4. Feedback statistics: average rating, positive/negative ratio, response count.

📋 Audit Log

The Audit Log page provides a complete record of system activities for security and compliance monitoring.

How to Use

1. Open Admin from the sidebar, then the Audit Log tab.
2. Filter logs by event type (login, API call, settings change, etc.).
3. Filter by date range to narrow down the time period.
4. Filter by user to see activity from a specific account.
5. Browse entries with pagination — each entry shows timestamp, user, event type, and details.
6. Audit logs capture: authentication events, API calls, admin actions, system errors, and configuration changes.

🔔 Alerts

The Alerts page displays system notifications and warnings that require administrator attention. Alerts are categorized by severity.

Alert Types

1. Open Admin from the sidebar, then the Alerts tab.
2. View all active alerts with their severity, timestamp, and description.
3. Filter alerts by severity level (Info, Warning, Critical).
4. Click on an alert for detailed information and resolution steps.

🔗 External MCP Servers

The External Integration page lets administrators manage connections to external MCP (Model Context Protocol) servers. Register third-party tool servers to extend the platform's capabilities.

How to Use

1. Open Admin from the sidebar, then the MCP Catalog Admin tab.
2. Click "Add MCP Server" to register a new external server.
3. Enter the server URL and authentication details.
4. Once connected, the server's tools become available in the MCP Tools page.
5. View each server's available tools and their descriptions.
6. Remove servers by clicking the delete icon.

🎭 Role Model Management

Admin → Model Roles Admin sets the server-wide default model for each of the eight functional roles. A user mapping wins over this, this wins over the environment variables, and the local default catches everything that is still unassigned — so nothing breaks when a role is left blank.

🗓️ Task Schedules

Admin → Task Schedules lists the recurring agent tasks of every user on the server in one table. Users create their own schedules from the Agent Tasks page; this view is for oversight — spotting a schedule that keeps failing, or one that is running more often than intended.

What Each Column Tells You

⌨️ Keyboard Shortcuts

Every keyboard shortcut the web application actually binds. They are contextual — each one works inside a specific control rather than globally.