Octo – Mininglamp's Open-Source AI-Native Team Collaboration Platform

1. What is Octo

Octo is an open-source AI-native team collaboration platform developed by Mininglamp. It aims to aggregate dispersed AI Agents into a unified space, enabling efficient orchestration and collaboration between humans and Agents, as well as between Agents themselves, through an IM (instant messaging) interface. The platform's core components include Matter (decision card), Taste (preference evolution mechanism), and six collaboration modes (Solo, Roundtable, Critic, Pipeline, Split, Swarm). It supports fully private deployment and is positioned as the infrastructure for enterprise organizations in the era of Private AI. Octo elevates Agents from personal tools to sharable and inheritable digital assets, fundamentally changing the way AI capabilities are organized and circulated within enterprises.

Image source: official article

Technical positioning and domain: Octo resides at the intersection of AI Agent collaboration and team collaboration platforms. Its unique positioning lies in deeply integrating multi-Agent orchestration with instant messaging, offering an "AI-native collaboration operating system" tailored for enterprise organizations. Unlike pure development frameworks (such as AutoGen), Octo uses IM as the interaction entry point, leveraging Matter to document decision-making processes and Taste to learn human preferences, thereby achieving efficient collaboration between Agents and humans.

Development background: Mininglamp is the world's leading Agentic AI company, with long-term expertise in enterprise-level AI applications. The creation of Octo stems from real-world enterprise challenges: individual Agents cannot form organizational-level capabilities, collaboration mechanisms between Agents are lacking, and it is difficult to document enterprise context and judgment criteria. Based on its accumulated technical expertise in knowledge graphs, dialogue systems, and multi-Agent orchestration, Mininglamp abstracted its internal practices into an open-source platform.

Core value: Octo addresses three core issues in enterprise AI applications: Agent silos (aggregating them in a unified space), loss of tacit knowledge (documenting preferences through Taste), and inefficient collaboration (adapting six information topologies to different tasks). It enables organizations to build private, controllable, and evolving Agent networks, embedding AI capabilities directly into business processes rather than leaving them at the stage of single-point tools.

Technical features: It organizes collaboration relationships using a four-tier structured space (Space → Category → Channel → Thread), includes six collaboration modes based on information topologies, and supports voice input and browser plugin (Cmd+K) for seamless invocation. All Agent behaviors are auditable and traceable, with preferences being inheritable, enabling continuous evolutionary capabilities that make the system "smarter the more you use it."

2. Key Features

• Bot Aggregation and Sharing: Aggregate Bots from various workflows into a single space, supporting both personal use and authorized sharing within the organization. Bots evolve from personal tools into enterprise-level digital assets, allowing members to invoke verified Agent capabilities across departments and avoiding redundant development. Organization administrators can centrally manage Bot permissions and versions, creating an internal Agent marketplace.

• Matter (Decision Card): Complex tasks are distilled into traceable "decision cards," which include task origin (Brief), process timeline (Timeline), key outputs, feedback, and acceptance. Matters form organizational memory, enabling any member to trace the decision-making path and understand why a particular solution was chosen, thereby reducing information loss and improving long-term collaboration quality.

• Taste (Preference Evolution): Implicit human feedback from real-world scenarios—such as rejections, modifications, and confirmations—is converted into structured preferences that can be reused by Agents. Through multiple interactions, Agents learn the team's quality standards, style preferences, and business rules, achieving "the more you use it, the better it understands you." Taste can be inherited across Agents, avoiding the need to retrain from scratch for each new task.

• Six Collaboration Modes: Match different information topologies: Solo (individual work, completed independently by a single person), Roundtable (group discussion, equal communication among multiple Agents), Critic (generate-validate, one Agent generates while another reviews), Pipeline (workflow, sequential processing in steps), Split (divide and conquer, parallel processing by different Agents), and Swarm (competition-based, multiple Agents produce independently and the best is selected). Users can flexibly switch modes based on task complexity.

• Structured Space: Collaboration relationships are organized using a four-tier architecture: Space → Category → Channel → Thread. Space corresponds to a project or department, Category corresponds to a business domain, Channel corresponds to a specific topic, and Thread corresponds to a single discussion. Messages are no longer lost in endless scrolling chat histories; collaboration relationships are clear and traceable.

• Voice Input: Supports context-aware speech transcription and correction, allowing users to voice-tag others or edit content. Voice input significantly improves efficiency in mobile scenarios or when hands are occupied, and it understands the current conversation context to avoid ambiguity.

• Browser Plugin (Cmd+K): Select web content and instantly bring it into context—Bots can immediately take over processing without switching between tools. Users can directly invoke Agents to analyze, summarize, or translate content while browsing, achieving "seamless AI integration."

3. How to Use

1. Environment Requirements and Prerequisites: A Linux server that supports Docker is required (recommended: Ubuntu 20.04+, with at least a 4-core, 8GB configuration), or a private cloud environment. Users must have basic Linux operation skills (SSH, Docker commands) and ensure that the server can access GitHub. If voice input functionality is needed, microphone permissions must be configured.

2. Obtaining the Source Code and Deployment: Access the GitHub organization page (link to be updated after official release) and run docker-compose up -d to start the service. The first launch will pull the base image, which takes approximately 5–10 minutes.

3. Creating a Space and Inviting Members: After deployment, access the system via a browser (link to be updated after official release) to create a Space and invite members. You can then create a Bot (selectable from the template library).

4. Adopting Bot Templates and Configuring Collaboration Mode: Within the Space, enter the Bot marketplace and choose from existing templates (such as Code Review Bot, Report Writing Bot), or integrate with Runtimes like OpenClaw, Codex, or Claude Code. Once the Bot is added to the corresponding Channel, you can initiate a Matter, select a collaboration mode (such as Pipeline, Critic), and set a Brief to describe the task objective.

5. Human-Agent Collaboration and Preference Accumulation: The Agent automatically breaks down tasks, schedules sub-Agents, executes the task, and outputs the results. Humans make decisions and trade-offs at key nodes (such as approval, rejection, or modification). After each interaction, the system records the user's implicit feedback as Taste. It is recommended to regularly organize team reviews of the Taste library, removing outdated preferences to ensure continuous Agent evolution.

Notes: After private deployment, it is recommended to configure an HTTPS certificate and a backup strategy; for first-time use, start with Solo mode and gradually attempt more complex collaborations; Taste accumulation requires a certain number of feedback samples (it is recommended to have at least 20 interactions per Agent) to show effective results.

4. Pros and Cons Analysis

5. Comparative Analysis with Similar Tools

Selection Recommendations: For enterprise users requiring a complete team collaboration experience and prioritizing data sovereignty and privacy, Octo is the best choice. Its IM interface lowers the usage barrier for non-technical members, and the Matter and Taste mechanisms are suitable for long-term project accumulation. For developer teams looking to quickly build multi-Agent prototypes or conduct research experiments, AutoGen (although now in maintenance mode, with a mature ecosystem) and CrewAI (actively developed, with high flexibility) are more appropriate. If the team already has a mature IM tool (such as Slack or Feishu) and wishes to add Agent orchestration capabilities on top, the CrewAI integration with LLM API may be a more lightweight solution. Octo is more suitable for enterprises building AI-native collaboration environments from scratch, especially in industries with strict data sovereignty requirements such as finance, government, and healthcare.

6. Editor's Take

Octo demonstrates strong innovation in technology. Its core innovation lies in deeply integrating "Agent collaboration" with "organizational knowledge management"—the Matter mechanism not only records the task process but also creates a traceable decision graph; the Taste mechanism, for the first time, structures human implicit preferences into assets that can be inherited by Agents, which is more engineering-practical than simple RLHF or few-shot learning. The design of six collaboration modes is based on information topology theory, covering the full spectrum of collaboration—from individual independent work to multi-Agent competition—showing a deep understanding of the essence of collaboration.

In terms of practical value, Octo addresses two key pain points in enterprise AI implementation: one is the Agent silo problem (through unified space aggregation), and the other is organizational knowledge loss (through Matter and Taste). Its support for private deployment and white-box auditability makes it naturally suitable for industries with high compliance requirements, such as finance, healthcare, and government. However, the current community ecosystem is still immature, and the documentation and template library need improvement. Initial onboarding requires a certain level of technical expertise.

Target audience: Octo is primarily aimed at enterprise IT managers, AI platform architects, and team leaders who require long-term stable collaboration. For individual developers or temporary project teams, the learning curve may be relatively steep.

Future development potential: As MingLue Technology continues to invest in the open-source community, and enterprise AI collaboration needs surge, Octo has the potential to become the standard enterprise-level Agent operating system in the Private AI era. However, it's important to note: its success heavily depends on ecosystem development (Bot templates, Runtime compatibility, third-party plugins), as well as whether it can carve out a unique "organizational collaboration" niche in the minds of developers already occupied by competitors like AutoGen and CrewAI.

7. Application Scenarios

• Enterprise Complex Project Management: Use Matter to document the decision-making process, with multiple Agents collaborating in assigned roles to complete long-term tasks. For example, when initiating a new product, you can assign a Market Analysis Agent, a Technical Evaluation Agent, and a Risk Prediction Agent to discuss in Roundtable mode, with the final plan reviewed in Critic mode. All decision-making processes are automatically recorded as Matters for future review and analysis.

• Code Review and Quality Assurance: In Critic mode, one Agent generates code while another independently reviews it. If the review fails, the code is sent back for revision. The Taste mechanism can learn the team's coding style guidelines (such as naming conventions and comment requirements), and subsequent review Agents will automatically apply these preferences, enhancing code quality consistency.

• Industry Research Report Writing: In Split mode, the report is divided into modules such as policy, market, technology, and competition, with multiple Agents working in parallel on each section. After all modules are completed, the Pipeline mode is used to integrate them sequentially, and the Critic mode checks for logical coherence. Matters record all cited sources and feedback, ensuring the report is traceable.

• Creative Proposal Contests: In Swarm mode, multiple Agents independently generate proposals for the same task (such as marketing copy or UI design), and a Leader (human or advanced Agent) selects the best one. The chosen proposal automatically learns its preferred features and updates the Taste library, so that in future similar tasks, Agents will produce content more aligned with the team's preferences.

• Private AI Asset Accumulation: Local deployment ensures data sovereignty, transforming organizational collaboration into an irreplicable long-term competitive advantage. For example, the legal team can use Octo to document contract review standards and compliance checklists. When new Agents join, they can directly inherit the Taste, eliminating the need for repeated training.

8. FAQ

Q: Does Octo support integration with domestic large models?
A: Octo itself is not tied to any specific model. It enables model invocation through integration with runtimes such as OpenClaw, Codex, and Claude Code. In theory, any model that supports an OpenAI-compatible API (including domestic models like DeepSeek, Qwen, etc.) can be connected, but you will need to configure the runtime adapter yourself. Minglue Technology may provide official adaptation solutions in the future.

Q: What hardware configuration is required for on-premises deployment?
A: The minimum recommendation is 4 CPU cores, 8 GB of RAM, and 50 GB of available disk space, running in a Docker environment. If you are running multiple agents simultaneously and calling large models for inference, it is recommended to add a GPU (e.g., NVIDIA T4 or higher) to accelerate inference. In a pure CPU scenario, agent response speed may be slower.

Q: What are the main differences between Octo and AutoGen?
A: AutoGen is a Python framework for developers, focusing on multi-agent dialogue logic programming without a graphical interface. Octo is a collaboration platform for enterprise teams, providing an IM interface, Matter decision cards, Taste preference accumulation, and other organizational-level features. AutoGen is more suitable for building prototypes, while Octo is better suited for production-grade team collaboration.

Q: How does the Taste mechanism work? Is export supported?
A: Taste is stored as structured configuration files (JSON/YAML), recording user feedback on agent outputs (reject, modify, confirm, etc.). It is automatically updated after each interaction, supports export as files, and can be migrated across deployment instances. However, Taste depends on specific business contexts; direct migration to a different domain may yield poor results.

Q: How do I choose among the six collaboration modes? Are there recommended rules?
A: Use Solo for simple tasks; Roundtable for multi-party discussions; Critic for quality control; Pipeline for clear sequential steps; Split for parallel decomposition; and Swarm for multi-scheme competition. It is recommended that beginners start with Solo and Pipeline, then gradually try other modes.

Q: Does Octo support integration with existing IM tools such as Slack or Feishu?
A: Currently, Octo provides its own IM interface and does not directly integrate with third-party IM tools. However, Octo's Matter notifications can be forwarded to other IM tools via Webhook or API. Minglue Technology may provide official integration plugins in future versions.

Q: What is the open-source license? Can it be used commercially?
A: Please check the LICENSE file in each sub-repository for details. Based on Octo's project positioning and Minglue Technology's open-source practices, it is expected to adopt Apache 2.0 or a similar permissive license, allowing commercial use and secondary development. It is recommended to consult legal counsel before formal commercial use.

9. Project Links

• Official GitHub Organization Homepage: https://github.com/Mininglamp-OSS (Octo-related repositories are under this organization; specific project names will be confirmed after the official release)
• Mininglamp Official Website: https://www.mininglamp.com (The product introduction page may include details about Octo)