A desktop workspace for the Mastra Code coding agent
Describe what you want in plain language. The agent reads your code, proposes a plan, edits files, runs commands, and helps you review and ship the result — all inside one window. No account, no telemetry, everything local.
Install
One command downloads the latest release and installs it to
/Applications. Requires macOS on Apple Silicon.
curl -fsSL https://raw.githubusercontent.com/JJJ-Mo3/yardarm/main/scripts/install.sh | sh
Prefer a manual download? Grab the .dmg or .zip from the
latest release and drag
Yardarm.app into /Applications.
xattr -dr com.apple.quarantine /Applications/Yardarm.app — or just use the
install script above, which avoids this entirely.
More in troubleshooting.
You can also build from source on any platform with Node 22+ and pnpm 10.
Why Mastra Code
Yardarm exists because Mastra Code is a genuinely different coding agent.
Observational Memory
Instead of compacting the conversation when context runs out, a background observer continuously distills what matters. Long-running work doesn't fall off a cliff at the context limit.
Goals with a judge
Hand the agent an objective and it keeps running until an independent judge model agrees the goal is met — not just until the first "done" claim.
Stored workflows
Ask the agent to save a repeatable multi-step process — a release checklist, a codebase audit, a scaffolding routine — as a workflow, then re-run it on demand with per-step results. See use cases.
Bring any model
OAuth in with Claude, OpenAI Codex, or GitHub Copilot subscriptions, use API keys from a dozen providers, or point it at any OpenAI-compatible endpoint — Ollama, LM Studio, vLLM, and more.
Deeply extensible
MCP servers, lifecycle hooks, skills and plugins, subagents, and custom slash commands — all plain files, configurable globally or per project.
Open source and local-first
Configuration is plain files on disk, shared between the CLI, ACP editors, and this app. No accounts, no telemetry.
What Yardarm adds
A full desktop workspace around the agent.
Parallel agents on one repo
Every chat gets its own git worktree on a yardarm/… branch, so several
agents can build, test, and commit on the same project at once — without touching
your working copy or each other.
One-click rollback
Every message pins a checkpoint (a real git ref). Roll back and both the transcript and the working tree return to that moment, with the agent told what happened.
Review and ship in one place
Side-by-side diffs of exactly what the agent changed, staging, commit, and push — with GitHub PR and GitLab MR flows, plus agent-powered code review of local changes or any open PR/MR.
Terminal, IDE, CLI & Preview
A real shell, a light Monaco-based IDE with language-server diagnostics, the interactive Mastra Code TUI, and an in-app browser for localhost dev servers — all scoped to the chat's worktree.
Kanban that dispatches agents
Write task cards, drag one to In progress, and Yardarm creates a chat and sets an agent working. Cards follow the agent's live status.
Modes at a glance
A color-coded Plan Build Fast selector, tool calls as expandable cards, and plan/tool approvals as buttons.
Full sandbox mode
Run the agent's shell commands inside an OS-level sandbox — writes contained to the worktree and allowed paths, with an optional network block.
Local model conveniences
Ollama auto-detect and auto-start, dropdowns filtered to models that will actually run, and no idle timeouts — a big local model can think for minutes without being cut off.
Persistent, organized history
Projects and chats with full transcripts survive restarts in a local SQLite database. Archive chats and projects without deleting anything.
Get started in minutes
Install, run the setup wizard, add a project, and type. The Getting Started guide walks through installation, first-run setup, and every part of the app — from modes, approvals, and checkpoints to the workspace tabs, settings and slash commands, and a full FAQ.