Orchevia app icon: orbit-and-nodes mark on a blue-violet gradient Orchevia
● Free & open source · macOS & Linux

Your AI coding CLIs, as reviewable pipelines.

Orchevia runs Claude Code, Codex, git, gh and your test commands as a visual workflow on your own machine. Plan, approve, implement, test and cross-review — with the subscriptions you already have.

Download Orchevia ▶ See a run ★ Star on GitHub

No account · No API keys · No telemetry · Apache-2.0

Orchevia showing a finished run on the canvas: a Jira ticket is planned by Claude, approved by a human, implemented, tested, reviewed in parallel by Claude and Codex, fixed once and opened as a draft PR. The side panel shows the Codex review findings.
A finished run, shown read-only on the canvas: every step, its output and its findings stay attached to the graph that produced them.
🔑 Your subscriptionsAgent steps call your signed-in claude and codex. No keys, no Orchevia bill.
🤝 Claude + Codex togetherOne implements, the other reviews — in parallel, with P0–P3 findings.
🧭 PredictableA fixed graph with approvals, bounded loops and timeouts. Nothing runs until you press Run.
🔒 Local-firstWorkflows and history in local SQLite. No accounts, no hosted runner, no telemetry.

Try it here

Watch a ticket become a reviewed draft PR

This is the run from the screenshot above, replayed in your browser. Claude plans, the run waits for you, Claude implements, tests run, Claude and Codex review in parallel, a P1 finding loops back once for a fix, and a draft PR opens.

Ready
Startidle
Start
Ticket · MCPidle
Load Jira ticket
Agent · Claudeidle
Plan
Human approvalidle
Approve plan
Agent · Claudeidle
Implement
Commandidle
pnpm test
Agent · Claudeidle
Claude review
Agent · Codexidle
Codex review
Agent · loopidle
↺ Fix findings
Conditionidle
P0/P1 found?
GitHub · ghidle
Open draft PR
Endidle
End
Press ▶ Run demo to start. The run pauses at the approval gate until you approve.

Simulated in your browser — nothing is installed or sent anywhere. In the app, each step runs your real CLI and keeps its full output.

Workflow generator

Describe the task. Get a graph you can check.

✦ Generate asks your own Claude Code or Codex CLI to draft the workflow. Orchevia validates and compiles the draft into a preview — it never runs it.

  1. Open Workflows → ✦ Generate. Pick Claude Code or Codex and, optionally, a repository: the AI is asked to inspect it without changing anything, so the draft uses your real paths and conventions. It runs with that profile's permissions.
  2. Write the goal like a short spec. Name the agents, the commands that prove it works, where you approve, and which writes are allowed.
  3. Review the preview. Steps, risk categories, setup items and every compiler repair are listed. Ask AI to fix sends blocking problems back for another round.
  4. Accept & edit, then Dry run. The workflow opens on the canvas; test its routing with fixtures before running real tools. 
Read Jira ticket {{input.ticket}} and plan it with Claude, then pause for my approval. Implement it on branch {{input.branch}}, run pnpm test, and have Claude and Codex review the diff in parallel. Fix P0/P1 findings (at most twice), then open a draft PR.

Generation prefers native Agent, Git, GitHub, Ticket, MCP, Condition and Approval nodes; Command nodes cover your own tools. Instructions such as “CLI only” win. ✦ Assist improves or rearranges an existing workflow through the same reviewed preview.

The Generate workflow dialog: a goal describing a ticket-to-PR flow, a repository directory, and a preview listing 12 steps, 4 risk categories, one setup item, the compiler notes and the proposed steps.
The preview lists what must be set up before running and every repair the compiler made.

Product tour

Everything a run touched, in one place

Real screens from the desktop app. Edit on the canvas, follow each run step by step, and see what your CLIs reported.

The workflow editor with the Implement agent node selected; the inspector shows its agent profile, prompt, required tools and session group.
  • Ten node kinds: Start, Command, AI Agent, Condition, Human Approval, Git, GitHub, Ticket, MCP Tool and End.
  • An inspector for every node: profile, prompt with {{variables}}, required tools, session reuse, timeouts and loop bounds.
  • Validation before you run: badges on the canvas, and Simple JSON that stays in sync with the graph.
The Runs view: a succeeded run with a timeline of 16 step attempts, including parallel Claude and Codex reviews, and a report of the final Codex review.
  • A timeline on one time axis, so parallel reviews and loop visits are easy to read.
  • Output and findings per step, with whether the agent resumed its earlier context.
  • Read-only run snapshots on the canvas, and deliberate recovery of failed runs from their original inputs.
The Usage and costs page: a daily chart of agent attempts over 30 days and a breakdown table by provider and model.
  • Agent attempts, tokens and tool calls by workflow, profile, provider and model.
  • Honest money lanes: CLI-reported API estimates stay apart from the subscription fees you enter.
  • Budgets per run: cap agent attempts or pause at a reported spend estimate.
The Agents page: Claude Code and Codex detected as installed, their known models, and agent profiles for each role.
  • Profiles per role: planning, implementation, fast and deep review, generation — each with its own CLI, model and effort.
  • Any CLI: Claude Code, Codex, or a custom binary that takes the prompt last.
  • Detected on your login shell's PATH, including nvm, fnm, Homebrew and ~/.local/bin.

Recipes

Start with a workflow you already do

Build a small manual flow first. Add integrations and scheduling once its steps work in your repository. Nine built-in templates help you start.

Plan, implement, test, review

Plan → Approve → Implement → Test → Review

Turn a change request into an inspected plan, then implementation and an independent review, using your project's real test command.

Ticket to draft pull request

Ticket → Plan → Approve → Implement → Test → Draft PR

Load a Jira ticket through your MCP server, then use native Git and GitHub operations for the handoffs. Every write step stays visible.

A repeatable PR review

Read PR → Parallel reviews → Summarize → Decide

Give each reviewer its own profile, compare their findings, and add a local watcher once the flow is proven.

Capabilities

Built for real repositories

Real control flow

Constrained conditions (no scripting), parallel branches with joins, failure edges, bounded retries and guarded review/fix loops.

Context that carries over

Agents can resume their own session within a run and receive bounded handoffs from earlier steps, while reviewers stay independent.

Git, GitHub, Jira, MCP

Safe Git operations (dirty states refused), GitHub through your own gh, and Jira or any tool through MCP servers you control.

Local automation

Durable schedules, PR review and feedback watchers with bot-loop guards, an inbox, and opt-in background mode.

Test before you run

Dry runs replay node results from fixtures so you can check routing without executing a single tool.

Reuse and memory

Templates with typed inputs, JSON import/export, and a project playbook of accepted commands and conventions.

Authoring

A graph for inspection. JSON for reuse.

Simple JSON uses readable node keys and portable profile roles; Advanced JSON keeps the full runtime format. Choose a native node when it supports the work and a Command node for your own tools.

Which node to use
Work Node What you supply
Plan, implement or review AI Agent Profile, prompt and repository
Repository and PR operations Git / GitHub Supported operation, repository and authenticated tools
Ticket or integration data Ticket / MCP Tool Enabled MCP server, discovered tool and arguments
Tests and local developer tools Command Executable and separate arguments; a shell only when explicit
A decision or human checkpoint Condition / Human Approval Constrained rule or approval message

Safety

Local-first, with permissions you can read

Orchevia passes permission flags on the command line and never edits your Claude Code or Codex settings. You decide once, in setup or Environment, whether agents work without asking; an Agent Profile can still pick its own level:

Use the CLI's settings

No flag is added; your own Claude Code or Codex configuration decides what a step may do. What every profile uses until you choose.

Allowed tools --allowedTools

An explicit Claude Code allow list. Nodes declare the tools they need, and imports and runs warn about anything the profile doesn't grant.

Full access

Bypasses permission checks, for every agent or one profile. Never chosen for you, and a schedule that uses it needs its own trust opt-in.

Refusals fail the step

When headless Claude Code refuses a tool, the step fails with the exact allow rules that would admit it instead of a silent “success”. Codex and Claude text-output profiles report no refusals, so these can't be caught.

Explicit execution

Opening, importing or previewing never runs anything. Cancel and timeouts stop the step's process group on Linux and macOS.

Your data stays home

Workflows and runs live in local SQLite. Your AI CLIs, GitHub and MCP servers may still contact their own services.

Getting started

Your first workflow in 5 minutes

  1. Install and open Orchevia. The setup wizard detects git, gh, claude and codex and creates recommended agent profiles.
  2. Workflows → Samples… → Plan → Implement → Review → Create. A ready-made graph appears on the canvas.
  3. Select any node to set its profile, prompt or timeout. Canvas badges flag problems before you run.
  4. Press ▶ Run. Approve the paused gate, follow each step in Runs, cancel at any time.
  5. Go further: ✦ Generate from a goal, save Templates, or schedule runs under Automations.

Start without any AI

The getting-started guide begins with a one-step echo workflow, then adds an agent, tests, an approval gate and reuse.

Headless Claude Code refuses tools its settings don't allow. If a step reports tools refused, add the listed rules to the profile in one click.

Download

Install Orchevia

Free and open source. The links below always point to the latest release. Windows is not supported yet.

macOS Apple Silicon Orchevia-macos-arm64.dmg macOS Intel Orchevia-macos-x86_64.dmg Linux AppImage · x86_64 Orchevia-linux-x86_64.AppImage Linux Debian / Ubuntu · x86_64 Orchevia-linux-x86_64.deb

Verify your download

Save SHA256SUMS next to the installer, keep the file names, and look for OK: 

# Linux
sha256sum -c SHA256SUMS --ignore-missing
# macOS
shasum -a 256 -c SHA256SUMS --ignore-missing

First launch

macOS: builds are not signed or notarized yet. Run the command below, or open the app once and choose Open Anyway in System Settings → Privacy & Security. 

xattr -d com.apple.quarantine /Applications/Orchevia.app

Debian and Ubuntu: add the signed apt repository, and sudo apt upgrade keeps Orchevia up to date: 

curl -fsSL https://iazaran.github.io/orchevia/apt/orchevia-archive-keyring.gpg | sudo tee /usr/share/keyrings/orchevia-archive-keyring.gpg > /dev/null
echo "deb [arch=amd64 signed-by=/usr/share/keyrings/orchevia-archive-keyring.gpg] https://iazaran.github.io/orchevia/apt stable main" | sudo tee /etc/apt/sources.list.d/orchevia.list
sudo apt update && sudo apt install orchevia

Other Linux: sudo apt install ./Orchevia-linux-x86_64.deb, or chmod +x the AppImage and run it.

Build from source

Contributors can build from source with Node.js 24+, pnpm 10+, Rust stable and the platform dependencies in the development guide. pnpm dev opens a browser demo with a mock backend; pnpm tauri dev runs real workflows in the desktop app. 

git clone https://github.com/iazaran/orchevia.git
cd orchevia
pnpm install
pnpm tauri dev

FAQ

Before you start

Do I need an Orchevia account or a model API key?

No. Orchevia orchestrates CLIs you install and sign in to yourself, so AI steps use your existing Claude or ChatGPT/Codex plan and its limits. A command-only workflow needs no AI CLI at all.

What does it cost?

Orchevia is free and open source (Apache-2.0). Usage of your AI CLIs is billed by their providers as usual. Runs show the tokens and any dollar estimate a CLI reports; those estimates are not your subscription bill or remaining quota.

Does local-first mean everything stays offline?

Workflows and run history live in local SQLite, and Orchevia has no telemetry or hosted worker. Orchevia itself only asks GitHub for its latest release, and you can turn that off. Your AI CLIs, GitHub operations and MCP integrations may contact their own remote services.

Can I use my own CLI or my repository's commands?

Yes. Add a custom CLI profile, or use a Command node with an executable and separate arguments. Templates are language neutral; you supply the commands, models and conventions that work in your project.

What happens when a step fails or I cancel?

Inspect per-step errors and partial output in Runs. Workflows can define retries and failure paths. Cancelling stops the active step and its process group on Linux and macOS; it does not undo earlier file edits, commits or remote writes, so review side effects before rerunning.

Will schedules run with the app closed?

Schedules and watchers need the app running. Background mode keeps it running when the window closes. Quitting the app or shutting down the machine stops local automation; there is no cloud worker.

Which platforms are supported?

Linux x86_64 and macOS on Apple Silicon and Intel. Windows is not supported yet. macOS installers are not signed or notarized. Steps run with your user permissions and are not sandboxed.

Help shape Orchevia

Orchevia is a young, maintainer-run project. A star helps other developers find it; a workflow recipe, a setup fix or a bug report with a redacted log makes it better for everyone.

★ Star on GitHub Contributor guide Report an issue