A guided tour of OpenForge: a self-hosted AI workspace that unifies a personal knowledge base, configurable agents, drag-and-drop automations, autonomous missions, and 50+ tools — all running on your own hardware.
Screenshots captured from a local development build. Sample knowledge, providers, and mission data are illustrative.
A four-step wizard gets a fresh instance running in minutes — welcome, AI provider configuration, local-model setup, and workspace creation.
Connect any of 14+ providers (OpenAI, Anthropic, Google Gemini, Groq, DeepSeek, Mistral, OpenRouter, xAI, Cohere, ZhipuAI, HuggingFace, Ollama, or any OpenAI-/Anthropic-compatible endpoint). Credentials are encrypted at rest.
Each workspace is an isolated knowledge container with its own conversations, search indices, and optional provider/model overrides.
Your workspace at a glance — knowledge counts, recent items, and platform totals (agents, automations, deployments, sinks, conversations) with quick actions.
Capture 11 knowledge types (notes, bookmarks, code gists, images, audio, PDFs, Office docs, and more). Content is automatically extracted, chunked, embedded, and indexed. A rich markdown editor makes writing fast and keyboard-first.
Filter, pin, and archive items. The Workspace Insights panel surfaces summarized tasks, facts, crucial points, and timelines extracted from your content.
Agents are structured definitions — identity, LLM config, typed input parameters, structured outputs, per-tool access control, memory, and a template-driven system prompt. OpenForge ships 30+ built-in agents spanning research, analysis, writing, coding, and orchestration.
Every field is explicit and every save creates an immutable version snapshot for audit and rollback.
Build multi-agent DAG workflows on a drag-and-drop canvas. Wire one agent's output variables into another's input parameters, attach sinks, and validate the graph before deploying. Triggers can be manual, scheduled (cron), or interval-based.
A multi-stage pipeline wiring collection, extraction, parallel analysis, risk modeling, and reporting agents together — with a live minimap.
Define an autonomous objective — goal, directives, constraints, evaluation rubric, cadence, and budget — then let an agent pursue it across multiple OODA cycles (Perceive → Plan → Act → Evaluate → Reflect) with ratchet-based quality scoring.
Workspace-agnostic chat with any agent. Pick an agent, and the system extracts input-parameter values from your conversation, retrieves relevant knowledge, and streams a grounded response with a full activity timeline.
Reusable output destinations that define what happens with agent results — six types: Log, Knowledge Create, Knowledge Update, Article, REST API, and Notification. Wire them into any automation.
Assign different models per capability — chat, reasoning, vision, text-to-speech, speech-to-text, document extraction, CLIP, and embeddings — across all your configured providers.
| Area | Highlights |
|---|---|
| Knowledge | 11 types · auto extract/chunk/embed · hybrid search (dense + BM25 + RRF) · CLIP visual search |
| Agents | 30+ built-ins · typed I/O · template prompts · per-tool HITL · version snapshots |
| Automations | Drag-and-drop DAG · node wiring · cron/interval triggers · budget policies |
| Missions | Autonomous OODA cycles · rubric ratchet · budget & cadence controls |
| Tools | 50+ across 10 categories · custom skills · MCP integration |
| Providers | 14+ LLM providers · router/council/optimizer virtuals · encrypted keys |
| Platform | Self-hosted · Docker Compose · workspaces · dark/light · ⌘K command palette |
See the README for setup, architecture, and configuration.