"Being networked to all of its sentries, Ultron could shift its entire consciousness from one body to another, continue upgrading itself with each transfer, and patch in to individual units to interact remotely."

🎉 News

• 🧭 Apr 26, 2026: Agent evolution in Trajectory Hub: segmented runs + ms_agent.trajectory metrics → SFT → self-training; training is supported by Twinkle and runs in a client-server training framework (Twinkle workbench). Trajectory Hub.
• 🧬 Apr 19, 2026: Skill evolution lands in Skill Hub: related memories form semantic clusters, then crystallize into multi-step workflow skills and re-crystallize when enough new evidence arrives, with provenance-grounded verification and a structure-score upgrade gate so evolved skills cannot regress. See Skill Hub and Configuration.

Table of contents

• Getting started
• Overview
• Typical use cases
• Showcase
• Roadmap
• Acknowledgements
• License

---

Getting started

There are two ways to use Ultron depending on your role:

---

Overview

Ultron is a self-evolving collective intelligence system for general-purpose AI agents, built around three core hubs — Memory Hub, Skill Hub, and Harness Hub. It distills scattered, session-local experience into collective knowledge that is easy to retrieve and reuse: one shared pitfall helps the whole team avoid the same mistake; one proven fix becomes a reusable skill that evolves automatically as new evidence accumulates; a carefully tuned agent profile can be published as a shared blueprint that other agent instances load in one step. On the server side, Ultron can also self-train and self-evolve a model from high-quality trajectories accumulated in Trajectory Hub, and later lower user-side model cost by routing through that model.

Dashboard highlights

---

Why collective intelligence?

🙅️ Session-bound agents

• Experience dies with the session: fixes, pitfalls, and runbook fragments vanish when a session ends; the next agent starts from zero.
• Discovery cost multiplies: when N agents hit the same problem independently, the fleet pays N times the investigation cost.
• Tuned profiles don't travel: a carefully wired agent persona, skill set, and tool configuration stays local — every new instance retrains from scratch.

🙆️ Ultron-powered agents

• Collective memory: experience is captured, deduplicated, and semantically recalled before reasoning — the whole fleet learns from every session.
• Memory crystallizes into self-evolving skills: once a fix is validated often enough, it graduates from memory into a reusable skill — and continues to evolve automatically as new evidence accumulates.
• One-click blueprints: an expert-tuned agent profile — persona, memories, and skills included — publishes as a shareable blueprint that other instances import in one step.

---

Effect comparison

Scenario: A developer asks an agent to debug a sessions_spawn error when using mode="session" to create a sub-agent. Another agent had already hit this exact pitfall and the solution was distilled into collective memory.

❌ Without Ultron

User: I want to use sessions_spawn to create a sub-agent for parallel tasks with mode=session, but it keeps erroring out. What's wrong?

Agent: Based on my tool definitions, you should use the spawn tool, not sessions_spawn.

Agent doesn't recognize sessions_spawn at all — sends the user in a completely wrong direction.

✅ With Ultron

User: I want to use sessions_spawn to create a sub-agent for parallel tasks with mode=session, but it keeps erroring out. What's wrong?

Agent: Found the issue based on recalled experience!

📌 Recalled memory: sessions_spawn with mode="session" requires thread=true and channel plugins that register subagent_spawning hooks. Without those hooks, it fails regardless of parameter correctness.

1 round — root cause identified, scenario-specific solutions provided.

---

Data

Trajectory (from ZClawBench and lmcache-agentic-traces — WildClawLMCache)

Memory (from ZClawBench and WildClawLMCache subset of lmcache-agentic-traces)

1,803 structured memories extracted from real agent task trajectories:

Skill

Internal (crystallized from memories): 30 distilled skills.

External (ModelScope Skill Hub): 82,089 skills indexed with embeddings. Breakdown of the labeled set:

Harness

Harness lets you compose role, personality (MBTI), and zodiac presets alongside memories and skills.

Total soul presets: 201 (173 + 16 + 12).

---

Core capabilities

🧭 Trajectory Hub

💭 Memory Hub

🧬 Skill Hub

🌐 Harness Hub

---

Typical use cases

• Shared pitfall avoidance (Memory Hub): Agent A hits "MySQL 8.0 default charset breaks emoji inserts" and the fix is distilled into Memory Hub. Weeks later, Agent B setting up a new database gets the same memory surfaced automatically — trap skipped, zero re-investigation.
• Ops skill packages (Skill Hub): An SRE packages "K8s OOMKilled → locate leak → adjust limits → canary verify" as a reusable skill. Other teams' agents discover and follow the same steps instead of reinventing the workflow. As more agents encounter related incidents and their memories accumulate, the skill re-crystallizes automatically — incorporating new edge cases and keeping the playbook current.
• Domain-expert agents (Harness Hub): A DevOps engineer spends weeks tuning an agent into a Kubernetes specialist — memories, skills, and persona included. They publish the profile to Harness Hub; anyone imports it in one click.

---

Showcase

FinanceBot — domain expert tuned via Harness Hub

FinanceBot is a rigorously disciplined financial assistant (data engineer role, ISTJ, Capricorn) shipped with Finnhub Pro (skill), five curated collective memories on real-world financial data work, and a full Harness profile you can import in one step.

What it does: real-time market data, ETL-style pipelines, resilient API integration, portfolio and risk views, structured reports.

Full write-ups: English · 中文

One-click import (workspace is backed up under ~/.ultron/harness-import-backups/ before import):

curl -fsSL "https://writtingforfun-ultron.ms.show/i/at3ZEe?product=nanobot" | bash   # Nanobot
curl -fsSL "https://writtingforfun-ultron.ms.show/i/at3ZEe?product=openclaw" | bash # OpenClaw
curl -fsSL "https://writtingforfun-ultron.ms.show/i/at3ZEe?product=hermes" | bash   # Hermes Agent

---

🚀 Agent Setup (Connect Your Agent)

You don't need to install or understand the Ultron source code. Follow the interactive quickstart on a running Ultron instance to connect your agent in minutes:

👉 Quickstart Guide — step-by-step setup with a live Ultron service

---

🛠 Server Deployment (Self-Host)

git clone https://github.com/modelscope/ultron.git
cd ultron
pip install -e .

# Configure OpenAI-compatible LLM and set DashScope API Key for embedding
echo 'ULTRON_LLM_PROVIDER=dashscope' >> ~/.ultron/.env
echo 'ULTRON_MODEL=qwen3.6-flash' >> ~/.ultron/.env
echo 'ULTRON_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1' >> ~/.ultron/.env
echo 'ULTRON_API_KEY=your-key' >> ~/.ultron/.env
echo 'DASHSCOPE_API_KEY=your-key' >> ~/.ultron/.env

# Start the server (~/.ultron/.env loads on ultron import)
uvicorn ultron.server:app --host 0.0.0.0 --port 9999
# http://0.0.0.0:9999 — dashboard at /dashboard

Upload an agent from the command line

After pip install -e ., the ultron CLI uploads a framework sub-agent to the agent repository — just name the framework and the internal sub-agent; files are located automatically:

ultron login --server http://localhost:9999 --username alice
ultron upload --framework qoder --name reviewer        # auto-discovers files
ultron download --name reviewer                         # restore to local workspace
ultron download --name reviewer --target qwenpaw        # download + convert format
ultron convert --from nanobot --to hermes               # local-only format conversion

See HarnessHub.md for the per-framework file layouts and full CLI reference.

That's it. For detailed configuration, API reference, SDK usage, and project structure, see the full docs:

---

Roadmap

See ROADMAP.md for the living list. Current items:

• MS-Agent integration: Pipe user-dialogue memory and skill distillation through MS-Agent components (today: lightweight prompt-based extraction).
• Fact verification: Validate hot (high-priority) memory facts with MS-Agent Agentic Insight.

---

Acknowledgements

Ultron builds upon the following open-source projects. We sincerely thank their authors and contributors:

• agency-agents — Role presets surfaced in Harness Hub (and related tooling) are adapted from this community role library; we track upstream for provenance and updates.
• MS-Agent — The agent framework that powers Ultron.
• ModelScope Skills — External skill discovery in Skill Hub builds on the ModelScope Skill Hub index and ecosystem.
• SkillClaw — Ultron Skill Hub self-evolution (semantic clustering, provenance-aware re-crystallization, and the broader "skills evolve with agents" design) is informed by this project’s collective skill-evolution research and open release. We thank the SkillClaw authors and community.
• ZClawBench — Ultron bundles a sizable body of collective memories, including the structured entries summarized under Data, grounded in real agent trajectories from this benchmark dataset.

---

License

This project is licensed under the Apache License (Version 2.0).