Beta v0.1.13 MacOS App Release

Highlights

• Improved agent recovery when providers reject oversized prompts with context-window errors. The agent now compacts once and retries instead of surfacing a raw provider failure.
• Added a desktop response format setting with HTML5 artifact and Markdown options. Desktop runs now pass the configured format through RPC into prompt construction.
• Fixed visible transcript truncation for retry flows, including one-past-end no-op behavior and standalone tool-call bubbles.

Validation

• Passed bun run verify before release.
• GitHub release workflows completed successfully for both the CLI binary and desktop app.

Full Changelog: v0.1.12...v0.1.13

Beta v0.1.12 MacOS App Release

Full Changelog: v0.1.11...v0.1.12

Full Changelog: v0.1.11...v0.1.12

Beta v0.1.11 MacOS App Release

Full Changelog: v0.1.10...v0.1.11

Full Changelog: v0.1.10...v0.1.11

Full Changelog: v0.1.10...v0.1.11

Beta v0.1.10 MacOS App Release

Full Changelog: v0.1.9...v0.1.10

Full Changelog: v0.1.9...v0.1.10

Beta v0.1.9 MacOS App Release

What's Changed

• Desktop app chat - Show how long in duration a chat response has lasted so that user knows how long each response has taken

Full Changelog: v0.1.8...v0.1.9

What's Changed

• Codex/local dev unrelated changes by @kinwo in #153

Full Changelog: v0.1.8...v0.1.9

Beta v0.1.8 MacOS App Release

Ouroboros Public Beta Release Notes

Release: first public beta

Ouroboros is now available as a public beta for technical users who want a
local, inspectable AI agent for real development work. This release packages the
CLI agent, Electron desktop app, shared protocol layer, memory system, Agent
Skills runtime, artifact viewer, subagent workflows, MCP integration, and
permissioned self-improvement loop into a macOS beta build.

Download the beta from the GitHub release:
Ouroboros-0.1.8-mac-universal.dmg

Highlights

• Desktop beta for macOS: install Ouroboros.app from the universal .dmg
  and use the desktop UI as the primary workspace surface.
• CLI agent runtime: run Ouroboros directly from the terminal, including
  single-shot prompts, interactive sessions, streaming output, JSON-RPC mode,
  provider selection, and configurable autonomous step limits.
• Local workspace control: choose Simple mode or Workspace mode, attach
  files, inspect command output, approve tool calls, and keep the agent grounded
  in the repository you select.
• Typed desktop bridge: the Electron app talks to the CLI over a typed
  JSON-RPC protocol, keeping the desktop as a presentation layer over the same
  agent core used by the CLI.
• Provider setup: use Anthropic, OpenAI, OpenAI-compatible endpoints, or the
  ChatGPT subscription auth flow, depending on how you want to connect models.
• Agent Skills: discover and invoke SKILL.md capabilities from bundled,
  user-global, and workspace skill roots, including slash-invoked skills.
• Memory and reflection: preserve durable memory, checkpoints, transcripts,
  observations, dream cycles, crystallized lessons, and evolution history.
• HTML artifacts: create sandboxed self-contained HTML artifacts and inspect
  them in the desktop artifact panel, including version switching, fullscreen
  viewing, download, and open-in-browser flows.
• Subagents and teams: use subagent lifecycle surfaces, team graph views,
  task assignment, debate/review workflows, permission leases, and worker diff
  approval flows.
• MCP support: configure local or remote MCP servers, inspect connection
  status, restart servers, and route MCP tool calls through the approval model.

What's Included

Desktop App

• Chat workspace with streaming model output, tool-call visibility, approvals,
  status updates, and cancellation.
• Session list, session loading, session renaming, deletion, and new-session
  flows.
• Simple mode and Workspace mode for choosing how much local context the agent
  can use.
• Settings for model providers, API keys, ChatGPT subscription auth, reasoning
  controls, workspace configuration, skills, MCP, and update policy.
• Artifact panel for HTML apps and documents created during agent sessions.
• RSI, memory, checkpoint, and evolution-log views for inspecting how the agent
  reflects on prior work.
• Subagent and team surfaces for tracking delegated work.
• Renderer contract and main-process E2E coverage for the desktop runtime.

CLI

• ReAct agent loop with planning, tool calls, observations, streaming, parallel
  tool execution, steering, and turn completion notifications.
• Built-in tools for shell commands, file reads/writes/edits, todo tracking,
  web search/fetch, artifacts, memory, reflection, dream cycles, skills,
  subagents, teams, MCP, and self-tests.
• JSON-RPC server mode for long-running desktop and automation integrations.
• Runtime validation with Zod and Result<T, Error> tool execution contracts.
• Structured transcript, memory, checkpoint, artifact, and evolution-log
  storage under local runtime paths.

Safety And Control

• Human approval remains central. File edits, shell commands, MCP tool calls,
  self-modification paths, and worker diffs are surfaced for review.
• Workspace access is explicit. Ouroboros works in the project you select and
  relies on local filesystem permissions plus the configured permission tier.
• Self-improvement is inspectable. Memory updates, reflection outputs,
  crystallized changes, and evolution history are visible instead of hidden.
• The app is local-first but not offline-only. Model calls, MCP tools, web
  tools, updater checks, package managers, and approved shell commands may use
  the network.

Installation

1. Download Ouroboros-0.1.8-mac-universal.dmg from the GitHub release.
2. Open the .dmg.
3. Move Ouroboros.app to Applications if macOS asks you to.
4. Launch the app and configure at least one model provider or supported auth
   flow.
5. Start with a repository you are comfortable testing with, preferably one with
   a clean git state and backups for important work.

Known Limitations

• This is beta software. Expect workflow, packaging, update, memory,
  subagent/team, MCP, and data-retention behavior to change.
• macOS is the primary published desktop path for this release. Windows build
  scripts exist, but the public beta is focused on macOS.
• Agent output still requires human review. Ouroboros can make mistakes,
  misunderstand context, produce unsafe commands, or generate incorrect code.
• Long sessions can consume substantial model context, provider quota, disk
  space, and time.
• Provider quality, latency, pricing, and tool-calling behavior depend on the
  model and provider you configure.
• Live LLM tests are manual and are not part of the default automated
  verification gate.
• The beta does not provide enterprise administration controls, offline-only
  model execution, or fully unattended autonomous code modification.

Recommended Beta Practices

• Use a clean git worktree before approving agent edits.
• Review permission prompts, command text, generated diffs, network access, and
  credential usage before approving.
• Avoid opening workspaces that contain secrets or private data you do not want
  sent to configured model providers.
• Keep important work backed up outside the beta app.
• Configure only MCP servers and provider credentials you trust.
• Report issues with reproduction steps, relevant logs, platform details, and
  whether the problem occurred in CLI mode, desktop mode, or both.

Developer Verification

The release is backed by the repository verification workflow:

bun run verify

That gate runs linting, type checks, CLI tests, and desktop E2E coverage. Live
LLM behavior remains a manual validation path:

bun run test:cli:live

Full Changelog: v0.1.7...v0.1.8

Beta v0.1.7 MacOS App Release

Full Changelog: v0.1.6...v0.1.7

Beta v0.1.6 MacOS App Release

Full Changelog: v0.1.5...v0.1.6

Beta v0.1.5 MacOS App Release

Full Changelog: v0.1.4...v0.1.5

Beta v0.1.4 MacOS App Release

Full Changelog: v0.1.3...v0.1.4