Open-source toolchain · Built for AI agents

Thetoolchainfor
AI-native development.

Ruah gives every agent its own worktree, turns any API into agent-ready tools, and runs it all from one CLI. Built for Claude Code, OpenCode, Codex — and whatever ships next.

Start in 60s View on GitHub

ruah workflow run feature.md

live

claude-code

src/auth/**

✓ merged

opencode

src/ui/**

✓ merged

codex

src/db/**

✓ merged

3 tasks · 3 worktrees · 0 conflicts✓ merged in order

The toolchain

Six tools. One CLI.

Guard the loop, prove done, price it, replay it, compare it, feed it APIs. Each tool works standalone.

@ruah-dev/cli

CLI Router

One command. Every agent tool.

Top-level CLI router and unified cockpit for the ruah ecosystem. Install one package to get instant access to guardrails, verification, cost profiling, replays, and conversions with sub-millisecond dispatch.

Unified CLI namespace (ruah <tool>)Sub-millisecond command dispatchZero-config auto-detection

Explore

$ ruah init
$ ruah task create api \
    --executor claude-code
$ ruah status

@ruah-dev/guard

Guard

Can I trust the agent not to do damage — in any harness, with proof?

One deterministic policy file. The same verdict in Claude Code, Codex, Cursor, and CI, backed by a bypass-attempt test suite. Portable and testable — not just another rm -rf blocker.

Sub-50ms deterministic command verdictsHigh-entropy secret scanner (sk_live, AWS, PEM)Drop-in Claude Code PreToolUse hook adapter

Explore

$ ruah task create auth \
    --files "src/auth/**"
$ ruah task start auth
# ✓ Worktree created
# ✓ Claims established

@ruah-dev/verify

Verify

Did the agent actually finish?

The independent, deterministic referee for agent outputs. Turns acceptance criteria into machine-runnable checks. Anything that cannot be machine-checked is marked unverifiable — never silently passed.

6 criteria types (shell, regex, schema, no-diff, file, script)Parallel criterion execution with deterministic orderAuto-detect stack (scripts -> starter criteria)

Explore

$ ruah conv curate spec.yaml
# 29 endpoints → 10 tools
# ~226 tok (was ~540)

@ruah-dev/opt

Optimizer

Where did my tokens go?

Read Claude Code session transcripts and canonical traces to reveal exactly where your tokens and money went. Pinpoint context bloat, repeated file reads, and oversized tool outputs with actionable fixes.

Claude Code JSONL transcript adapter (~/.claude/projects/)Waste heuristics H1–H4 (oversized tools, repeated reads, dead context)Self-contained offline HTML reports (zero CDN, strict CSP)

Explore

$ ruah guard check --cmd 'rm -rf /'
# deny  deny-rm-root

@ruah-dev/watch

Watch (Lite)

Can I show someone what this session did?

Convert raw, messy agent sessions into beautiful, self-contained HTML replays. Includes interactive turn timelines, syntax-highlighted diffs, and automatic privacy redaction. Zero server required.

Single-command replay generation (ruah watch render --latest)Automated privacy redaction (--redact strips paths, tokens, emails)Syntax-highlighted diff view for Edit and Write tool calls

Explore

$ ruah verify run
# VERDICT: PASS

@ruah-dev/eval

Eval

Same task — which executor wins, and what did it cost?

A fast micro-benchmark harness for AI coding agents. Run one task specification across multiple executors in fresh, isolated sandboxes with deterministic acceptance criteria and cost scorecards.

Ephemeral sandbox isolation per run (/tmp/ruah-sandbox-*)Multi-executor orchestration (Claude Code, Codex, OpenCode, Scripts)Deterministic criteria referee (powered by Ruah Verify)

Explore

$ ruah opt analyze --json
# tokensIn / tokensOut / cost

@ruah-dev/conv

Converter

How do I make this API agent-sized?

Curate and convert massive API specifications into compact, agent-friendly tool surfaces. Compress 400-endpoint specs into 5–10 task-level MCP tools with minimal context overhead.

5 input formats (OpenAPI 3.x, Swagger 2.0, Postman, GraphQL, HAR)Smart endpoint curation (collapses bloated APIs into task-level tools)MCP Tool Definitions & TypeScript/Python server scaffolds

Explore

$ ruah watch render --latest
# wrote replay.html

@ruah-dev/orch

Orchestrator

Parallel agents. Zero collisions.

Multi-agent orchestration that coordinates simultaneous code changes. Each task runs in an isolated git worktree, file-level claims prevent merge conflicts, and DAG scheduling enforces dependency order.

Git worktree isolation per taskFile-level claim locking (owned / shared / read-only)Workflow DAG scheduling with automatic dependency order

Explore

$ ruah eval run spec.json --runs 5

How it works

Six tools. Composed.

Each tool stands alone — together they form the toolchain. Click a step to jump. Demo cycles automatically — hover to pause.

~/.project · 01 · convert

$ ruah conv generate stripe.yaml \
    --target mcp-ts-server

→ Parsed 47 operations
→ Generated 47 tools
→ Risk: 32 safe · 12 moderate · 3 destructive
→ Wrote ./out/mcp-server (TypeScript)

translating the spec into tools…

The roadmap

An expanding toolchain. Eleven more on the way.

Open-core by default. The CLI and runtime stay OSS; team dashboards, policy management, and the hosted registry ship as Ruah Cloud. Each new tool earns its slot — shipping when the signal says so, not when a calendar does.

In development

Prototypes building now.

Ruah Architect
Architect
PRDs into DAGs.
Brief, issue, or meeting notes in. A typed task graph with claims, acceptance criteria, and rollback plans out.
Ruah Optimizer
Optimizer
Cost & context, measured.
Token counts, wasted-token detection, model routing, and budget policies per workflow. Profiler (`@ruah-dev/opt`) is shipped; routing is later.
Ruah Code
Code
One IDE. Every agent.
A unified coding environment built on T3Code — Claude Code, OpenCode, and Codex converge into a single surface with shared state.
Ruah Guard
Enterprise
Guard
Policy engine for agents.
CLI/core shipped as `@ruah-dev/guard`. Hosted policy + tenant isolation stays enterprise later.

Coming soon

Designed. Shipping when the signal says so.

Ruah Exec
Exec
Typed runtime for agents.
Replaces bash as the agent-to-system interface. Typed file, git, HTTP, and process ops with structured errors.
Ruah Observe
Observe
Traces, costs, replay.
Per-workflow traces, task lineage, and decision snapshots. Exports to OpenTelemetry — no generic obs rebuild.
Ruah Link
Link
Cross-repo orchestration.
Coordinate agents across monorepo boundaries: update API in repo A, client SDK in repo B, docs in repo C.
Ruah Gateway
Enterprise
Gateway
API gateway for agents.
Auth, quotas, audit, tracing, and tenant isolation across MCP, REST, gRPC, WebSocket, and A2A endpoints.
Ruah Sync
Enterprise
Sync
Tickets ⇄ tasks.
Two-way GitHub, Jira, Linear, and Notion sync. Work items become agent tasks; results update tickets.
Ruah Hub
Enterprise
Hub
Registry for agent work.
npm-style registry for tools, prompts, contracts, workflow templates, and skills. Composition, discoverable.
Ruah Watch
Enterprise
Watch
Live workflow UI.
Replay-lite shipped as `@ruah-dev/watch` (static HTML). Live tail + DAG studio stays parked.

Install

One install. The whole toolchain.

Node.js 18+ · Zero runtime dependencies · MIT licensed.

Then this

60-second quickstart Full docs

Thetoolchainfor
AI-native development.

Six tools. One CLI.

CLI Router

Guard

Verify

Optimizer

Watch (Lite)

Eval

Converter

Orchestrator

Six tools. Composed.

Convert any API into agent tools

Isolate every agent in its own worktree

Run the whole workflow from one CLI

An expanding toolchain. Eleven more on the way.

In development

Architect

Optimizer

Code

Guard

Coming soon

Exec

Observe

Link

Gateway

Sync

Hub

Watch

One install. The whole toolchain.

ThetoolchainforAI-native development.

Six tools. One CLI.

CLI Router

Guard

Verify

Optimizer

Watch (Lite)

Eval

Converter

Orchestrator

Six tools. Composed.

Convert any API into agent tools

Isolate every agent in its own worktree

Run the whole workflow from one CLI

An expanding toolchain. Eleven more on the way.

In development

Architect

Optimizer

Code

Guard

Coming soon

Exec

Observe

Link

Gateway

Sync

Hub

Watch

One install. The whole toolchain.

Thetoolchainfor
AI-native development.