OpenMake 1.19.0 / Desktop 1.7.1

Your AI. Your server. Your rules.

A self-hosted AI workspace. One local model answers by default, and every external provider stays opt-in — your key, your policy.

OpenMake chat workspace with a sidebar containing Chat, Agent Tasks, Custom Agents, Artifacts, and Admin, and a composer showing the active tools and model
Five sidebar destinations, and a composer that names the tools, the response style, and the model behind every answer.
Choose a path

Evaluate it now, or run it yourself

Both routes use the same product, but the hosted demo is operated for you while self-hosting makes you responsible for the application, model endpoint, and deployment boundary.

Try it now

Open the hosted demo, or install the Desktop app on an Apple Silicon Mac. The desktop app is currently available for macOS only, with no infrastructure to stand up first.

Setup
None
Desktop app
macOS only · Apple Silicon · 1.7.1
Guest access
Default local model only

Run your own

Install it on your own machine so the model, the database, and the network boundary all stay yours.

Runtime
Node 24, PostgreSQL, Docker for sandbox services
Model endpoint
Any OpenAI-compatible
Setup time
Depends on the model endpoint and deployment
Verified numbers

What is actually in the box

Everything below is read from the running 1.19.0 source rather than from a roadmap. Where a feature is optional or off by default, this site says so.

262K
Effective context window
22
Built-in MCP tools
18 / 100
Industry agents / specialists
4
Interface languages
  • Overflow is handled, not hidden: input is truncated, the output cap is reduced, and an unrecoverable request returns HTTP 413 with an audit record and an alert
  • When MCP sandboxing is enabled, stdio servers run in Docker with an isolated execution boundary
  • Interface languages are Korean, English, Japanese, and Simplified Chinese
Composer mode menu listing Discussion, Thinking, Deep Research, Web, Agent, Image, Artifact, Structured, and Select notebook
One composer, eight modes. Thinking is active here; Select notebook pins a NotebookLM source to the current conversation.
In practice

Five ways to work, one context

Chat is the entry point, not the whole product. The same conversation can hand a goal to an autonomous task, ground an answer in sources you supply, or return a reviewable artifact instead of a wall of text.

Chat and routing

One local model answers each request. Response style and mode stay visible controls, not hidden presets.

OpenMake chat workspace with the sidebar and a composer showing active tools, style, and model
The composer names the model before you send.

Agent tasks

When the opt-in task sandbox is enabled, multi-turn runs use a Docker workspace with shell, Python, browser, and file tools. Desktop local work uses a separate, confirmation-gated bridge.

Agent task list showing completed runs with turn counts, progress bars, elapsed time, and token totals
Turns, duration, and token cost per run. Captured with the Korean interface.

Deep Research

Decompose the question, gather sources, cross-check them, then write a report that keeps its citations.

Deep Research history with a four-stage pipeline and a cited report
The pipeline is visible while it runs.

MCP tools and artifacts

Governed tool servers, and results you can open, re-run, and share instead of scrolling back through a thread.

Settings, Connectors tab listing installed MCP servers with transport type, tool count, and status
Every connected server, its transport, and its tool count.

Discord gateway

Chat with OpenMake from Discord through an optional gateway bot. Messages relay to the OpenAI-compatible API, with per-user sessions and role- or mention-based access control.

Real Discord channel showing a Korean enterprise AI platform market-analysis report posted by an app
A real Discord channel view with a structured Korean report.
Chat runtime

A normal chat route, with a deliberate MoA-style switch

OpenMake does not fan every message out to a panel of specialist agents. The default route resolves one model and runs the regular chat/tool loop. Turn on Discussion when the question benefits from independent specialist views and an explicit synthesis pass.

This is MoA-style orchestration in a bounded mode: the system selects a panel, gives each expert the same evidence package and context priorities, then compares the work before composing the answer. The progress events are visible in the chat instead of hiding the pipeline behind a spinner.

01 / Select

The topic is matched to two or more related agents. The panel can use the conversation, documents, web context, and image descriptions without giving every agent every credential.

02 / Fan out

Within each round, independent expert opinions run through a bounded parallel batch. A failed opinion is retried selectively; a partial result is not presented as a full panel.

03 / Review

A cross-review pass identifies strengths, shared claims, and conflicts. When enabled, the same evidence package is collected before the opinions so every expert reasons from the same sources.

04 / Synthesize

The composer receives the opinions, review, evidence, and consistency signals, then returns one answer with participants, sources, and degraded-state truth when the panel shrinks.

  • Live phases: selecting → discussing → reviewing → synthesizing → complete
  • The shared context builder prioritizes conversation history, documents, web results, and images under token budgets
  • Optional auto-dispatch is off by default. With ORCHESTRATION_AUTO_DISPATCH=true, an intent prefilter exposes discussion or background-task delegation tools only on matching turns; shadow instrumentation records decisions without changing execution.
  • Implementation anchors: apps/api/src/agents/discussion-engine.ts, services/chat-strategies/discussion-strategy.ts, and workflow/graph-engine.ts
Models and connections

Keep the local route. Add outside capacity deliberately.

With a local default endpoint configured, the standard answer path stays inside your gateway. When a job genuinely needs a different model, register a key and assign it to a single functional role — agent, judge, research, spawn, review, or summary — instead of switching the whole workspace over.

Resolution is fail-open. A per-user mapping wins first, then the admin-set global default, then the environment default, then the local qwen3.6-35b-a3b. If an external call fails, the local model still answers.

  • OpenRouter API key
  • ChatGPT Plus or Pro subscription OAuth — an unofficial, policy-dependent device flow
  • Ollama Local and Ollama Cloud
  • NVIDIA NIM
Model Roles Admin screen showing server-shared external keys and a global role-to-model mapping list
Role mappings live in the admin console. An empty field falls back to the local qwen3.6-35b-a3b default.
Trust and control

Built for an operator, not a black box

The reference deployment runs the application on a single host. PostgreSQL is required, Redis is optional for shared multi-instance storage, and Docker is used when you enable the task, MCP, or artifact isolation services.

Access is governed by JWT authentication and role-based permissions, privileged actions are written to the audit log, and the audit log is wired to alerts.

  • A Docker-sandboxed stdio MCP server runs with --cap-drop ALL, a non-root user, memory limits, and its configured network policy
  • Pending database migrations are applied at server startup by default; set DB_AUTO_MIGRATE=false when you run them by hand
  • The Desktop updater accepts only HTTPS or loopback origins and verifies SHA-256 before replacing the app
Since 2013

From a Raspberry Pi community to a self-hosted platform

OpenMake is not a company. It is a team of people who enjoy open source and open hardware, sharing what they build and keeping the habit of running their own machines alive. That same spirit now guides the openmake_llm project.

  1. RaspberryPi Village

    rasplay.org opens as a maker community, alongside a printed Raspberry Pi practical guide.

  2. OpenMake team

    People who enjoy open-source software and open hardware gather at openmake.cc as a team, not a company.

  3. Raspberry Pi JAM of Korea

    Registered with the Raspberry Pi Foundation as Korea's official Raspberry Pi JAM.

  4. openmake_llm

    The team's open-source, local-first AI workspace, currently at OpenMake 1.19.0.

Start

Pick the next useful step

The demo is the fastest way to form an opinion. The installation guide is the shortest route to an environment you own. The release notes are where you check what changed before you upgrade.