Ralph TUI
ralph-tui.com- Category
- AI Agents
- Rank
- No. 2784Tools index
Previous survey · No. 2722 ·
- Pricing
- Open Source
- Type
- TOOL
- Use case
- Agent Building · Coding
- Interfaces
- CLI · API
- Builder
- subsy
- GitHub
- 2.5k stars
- Latest release
- v0.12.0
- Date
About
An AI agent loop orchestrator that manages autonomous coding agents through intelligent task routing and continuous delivery. It helps developers orchestrate AI coding agents at scale with features like dependency-aware task planning, real-time monitoring, and automated execution loops.
What it does
Ralph TUI is a terminal program that feeds a list of work items to a coding assistant one at a time and keeps going until the list is empty. Each pass picks the highest-priority open item, writes a prompt from a template, runs the assistant, and waits for it to print a fixed completion marker before marking the item done. It drives Claude Code, OpenCode, Factory Droid, Cursor CLI, Gemini CLI, Codex and Kiro CLI, and reads work from a simple JSON file or from Beads, a git-backed issue tracker with dependencies.
Why it's ranked here
The case rests on how much of the unglamorous plumbing is already built. Sessions pause, resume and survive crashes. Rate-limited agents can hand off to fallback agents. Independent items can run side by side in separate git worktrees, then merge back one at a time onto a session branch, with the agent itself called in to settle merge conflicts. CI type-checks, lints and builds on Linux and Windows and runs a large test suite with coverage. It is MIT licensed. For anyone running long unattended agent loops, that is far more than a shell script wrapped around a prompt.
What's good
Several details show real operating experience. Live agent output held in memory is capped at 250,000 characters, while full logs go to disk, so long runs do not grow without bound. Before each prompt it pulls in progress notes from the last five iterations, giving the agent memory across items. An optional sandbox wraps the agent in Bubblewrap on Linux or sandbox-exec on macOS, with a switch to cut network access. A remote listener lets one terminal watch and control instances on other machines, protected by a generated token that can be rotated.
Tradeoffs
It requires the Bun runtime and refuses to start on an older Bun than the package declares, so Node-only setups need an extra install. Completion depends on the agent printing an exact marker; an agent that finishes without it looks unfinished. Parallel mode leans on git worktrees and automatic conflict merging, which adds moving parts, and the default is three workers. One CI test batch covering the agent plugins is allowed to fail without failing the build. Linux sandboxing needs Bubblewrap installed separately.
How to use it well
It suits a developer who can write a clear list of small, well-specified items with acceptance criteria and wants an agent to grind through them overnight. Start with a single JSON task file and a capped number of iterations, turn on the sandbox, and read the per-iteration logs before trusting longer runs. Move to Beads when items depend on each other or when parallel runs are worth the merge risk. It does not write the plan for you beyond an assisted requirements draft, and it does not review the resulting code.
Technical notes+
src/cli.tsx dispatches subcommands and defaults to run; it checks the Bun version against engines.bun from package.json. src/engine/index.ts holds ExecutionEngine: completion is detected by a case-insensitive <promise>COMPLETE</promise> regex, error strategies are retry, skip or abort, rate limits are detected with a RateLimitDetector and fallback agents, and the primary agent is re-probed with a five second test prompt. src/parallel/index.ts runs task-graph groups in topological order, defaults maxWorkers to 3 and aiConflictResolution to true, and re-queues a conflicted task at most once. src/config/index.ts merges global and project TOML config files with Zod validation. .github/workflows/ci.yml marks the agent-plugin test batch with || true.
Observed
- License
- MIT
- Language
- TypeScript, React terminal UI via OpenTUI
- Runtime
- Bun required (install via bun install -g ralph-tui)
- Interfaces
- Interactive TUI, headless CLI mode, plugin library export, WebSocket remote listener
- Supported agents
- Claude Code, OpenCode, Factory Droid, Cursor CLI, Gemini CLI, Codex, Kiro CLI
- Task sources
- JSON PRD file, Beads; Jira and Linear tracker code present in the build and tests
- Sandboxing
- Bubblewrap on Linux, sandbox-exec on macOS, optional network cutoff
- CI
- Type check, lint and build on Ubuntu and Windows; test suite with coverage on Ubuntu
Read from README.md, package.json, LICENSE, src/cli.tsx, src/engine/index.ts, src/plugins/agents/index.ts, src/plugins/trackers/index.ts, src/sandbox/index.ts, src/parallel/index.ts, src/config/index.ts, testing/TESTING.md, .github/workflows/ci.yml.
What it can do
Route coding tasks to appropriate AI agents
Development tasks and available AI agents → Task assignments to optimal agents
Plan task execution with dependency awareness
Multiple coding tasks with interdependencies → Ordered execution plan respecting dependencies
Monitor AI agent execution in real-time
Running AI coding agents → Live status updates and performance metrics
Execute automated continuous delivery loops
Code changes from AI agents → Deployed applications and delivery reports
Orchestrate multiple AI coding agents simultaneously
Collection of AI coding agents and development projects → Coordinated multi-agent development workflow
Manage autonomous coding agent lifecycles
AI coding agents and operational parameters → Started, stopped, and configured agent instances
Tags
Tech Stack
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.