Reflection browser
Reflection and evolution history make the agent's learning process inspectable instead of burying behavior changes inside hidden prompts.
Ouroboros
A local first self-improving AI agent for real projects. Ouroboros reflects on work, preserves memory, grows through Agent Skills, and keeps every improvement inspectable.
Self-improvement is permissioned and visible. Beta software for technical users; review approvals, diffs, credentials, and workspace access carefully.
See it in action.
Real desktop sessions: the agent reasons in chat, calls tools with permission, and ships sandboxed HTML5 app artifacts you can open, expand, and download from the conversation.
Artemis II Orion trajectory
NASA/JPL Horizons vector data for Orion, the Moon, and the Sun rendered as an embedded WebGL/Three.js artifact with playback, zoom, and orbit controls.
Architecture diagram
Ask for a system diagram and the agent responds with a Mermaid artifact you can expand fullscreen, zoom, and inspect at any layer.
Self-improvement is the main event.
Ouroboros is built around a visible improvement loop. It observes work, reflects on what happened, records durable memory, and turns repeated useful patterns into skills under human control.
Durable memory
Memory, transcripts, checkpoints, and working state preserve context across sessions so the agent can improve without losing accountability.
Permissioned growth
Code and configuration self-modification is a disabled-by-default approval tier. Generated skills are validated and self-tested before they become available in the catalog, with every automatic improvement recorded in the evolution log.
Meta Thinking is built in, but optional.
Meta Thinking is a built-in SecondOrder skill for deliberate self-critique. It is not enabled by default; turn it on for complex plans, tradeoff decisions, troubleshooting, and moments where uncertainty should be visible instead of hidden.
classify
Choose when extra thinking is worth it
The skill classifies the task and gates meta mode so simple requests stay direct, while high-uncertainty work gets a stronger control loop.
extract
Make goals and constraints explicit
It pulls out the goal, constraints, response strategy, and compact plan before the agent commits to an answer or implementation path.
critique
Catch brittle assumptions earlier
A critique pass checks for missing context, unsupported assumptions, weak plans, and limitations that could materially change the result.
signal
Give users better trust signals
When useful, the agent can surface concise confidence, limitations, and context gaps so you can decide whether to proceed, steer, or provide more information.
Transparent autonomy, not silent automation.
Ouroboros can work through real tasks while keeping tool calls, permissions, memory, and self-improvement visible.
Visible tool calls
Agent actions are surfaced as inspectable tool events. You can see what was read, run, or proposed.
Explicit approvals
File edits, shell commands, network access, and self-improvement follow the permission model instead of assuming unattended control.
Inspectable context
Memory and reflection history make repeated behavior easier to understand and correct over time.
Works inside real repositories.
Ouroboros supports Workspace mode for repository-aware work and Simple mode for focused chat. It can reason over code, docs, tests, skills, artifacts, and project history while keeping sensitive actions behind review.
✓Use Agent Skills packaged as
SKILL.md instructions.
✓Coordinate subagents and team workflows for larger engineering tasks.
✓Create and preview sandboxed HTML5 app artifacts from agent work.
observe
Ground in the workspace
Read files, inspect context, search code, and understand the shape of the project before acting.
act
Use tools with permission
Run local tools, propose edits, create artifacts, and request approval when an action crosses a safety boundary.
reflect
Keep useful context
Update memory and evolution logs so repeated work becomes more consistent without becoming opaque.
Built for recursive work.
The product surface is more than chat. Ouroboros combines self-improvement, memory, skills, artifacts, subagents, and team workflows into one local agent runtime.
Memory and Reflection browser
Browse remembered context, checkpoints, reflections, crystallizations, and evolution history to understand why the agent changes.
Workspace and Simple mode
Use Workspace mode when the agent should operate inside a repository, or Simple mode when you want a lighter conversation surface.
HTML5 app artifacts
Create self-contained HTML app artifacts, preview them safely in the desktop app, and keep generated work attached to the session.
Agent Skills
Load reusable capabilities through the Agent Skills SKILL.md format
so learned workflows can become explicit behavior.
Subagents
Delegate focused exploration, review, or implementation work to subagents while keeping lifecycle events visible in the main thread.
Team workflows
Coordinate multiple agents through task graphs, quality gates, debate, review, and cleanup flows for larger projects.
One agent core. CLI and desktop surfaces.
The CLI owns the agent runtime. The Electron desktop app gives the same core a focused interface for onboarding, chat, approvals, settings, artifacts, and RSI history.
CLI for power users
Run single-shot prompts, interactive sessions, JSON-RPC mode, local tools, memory, reflection, MCP, and skill workflows from the terminal.
./dist/ouroboros -m "Summarize this repo"
Desktop for inspectable work
Use chat, provider setup, Workspace and Simple mode, file attachments, approvals, HTML5 app artifacts, Memory and Reflection browser, subagents, and team surfaces in one app.
cd packages/desktop
bun run dev
What Ouroboros can and cannot do.
The launch posture is controlled public beta. The goal is trust, inspectability, and repeat use rather than promises of full autonomy.
Can do
- Run an AI coding agent from the CLI or desktop app.
- Stream model output, call tools, and ask for approvals.
- Work with files, shell commands, MCP servers, Agent Skills, HTML5 app artifacts, memory, reflection, subagents, and team workflows.
- Use provider API keys or supported auth flows to connect to models.
Cannot promise
- Correct, safe, or useful agent output without human review.
- Stable beta behavior across every packaging, update, memory, and workflow path.
- Offline-only model execution or enterprise administration controls.
- Fully unattended code modification without oversight.
Beta safety notes
Start with projects and credentials you are comfortable testing with. Ouroboros can read and write workspace files when you approve tool use.
- Use repositories with clean git state and backups for important work.
- Review permission prompts, generated diffs, shell commands, and network-related actions before approving.
- Avoid opening workspaces that contain secrets or data you do not want sent to configured model providers.
- Only configure MCP servers and provider credentials you trust.