Data kept on your machine
The knowledge graph, the search index and the message bus run as services next to the server, on your machine or your host. What leaves is spelled out in the FAQ.
For platform, DevOps and security teams
Project Orchestrator keeps its data local, holds secrets in an encrypted vault and decides what an assistant may reach. This page says what the code proves, what it does not, and where to check.
In the product todayPartly covered
Proofs last checked on 7 Oct 2026
“Where does our data go, and who else sees it?”
An assistant reads your code and your documents. Before you plug it in, you want the list of what stays on the machine and what is sent to a model.
“I do not want API keys pasted into prompts.”
A key typed into a conversation stays in the history and in the logs. You want to decide which assistant gets which secret, for how long.
“What can an agent reach once I connect it?”
Each MCP tool is a way out: a tool that opens sessions, runs plans or reconfigures the server widens the blast radius of one bad prompt.
“Show me what is proven. I will not take a vendor’s word.”
Security claims without a source file are marketing. You want the threat model in plain words, the code that enforces it and the part nobody tested.
The knowledge graph, the search index and the message bus run as services next to the server, on your machine or your host. What leaves is spelled out in the FAQ.
Secrets are encrypted at rest and reach an assistant only through a grant limited to a session or a project, and to a time. The vault is not a sandbox.
Connect external tool servers. By default only read and search tools run; mutating tools are refused until you allow them. You can also list the servers that may be called.
A session opened by an assistant is never more than its parent: same project, same folder, permission mode clamped, one level deep, four at a time. Agent processes start from a clean environment.
You can register another model, allow it project by project and give it a key. It gets a restricted tool list and every send is journalled. Never tried against a real service.
A REST API and WebSocket events drive everything the interface does. /health and /api/version answer without a login. Log level follows RUST_LOG.
One secret, one external tool server, one afternoon. The same task, without and with the vault, the policy and the envelope.
YouYou paste the deploy token into the conversation so the agent can run the script.
AssistantThe token now sits in the transcript, the search index and every copy of the log.
YouYou plug in a GitHub tool server with a write token because it is faster than reading what it exposes.
AssistantThe agent opens helper sessions on its own and nothing says how many, in which folder or with which permissions.
YouYou store the token in the vault and unlock it for the afternoon. You grant it to this project only, with an end date.
AssistantThe agent runs the script through the secret command: the value goes into the command, not into the conversation.
Project OrchestratorIf the value still shows up in the agent’s output, it is replaced before it is stored, indexed or broadcast.
Project OrchestratorThe GitHub tool server answers read and search calls. A call that creates or deletes is refused and says which rule refused it, until you allow mutations.
Project OrchestratorThe helper sessions stay in the parent’s project and folder, with at most the parent’s permissions, one level deep.
The exposure is smaller, bounded in scope and time, and you can read why. An agent that holds the token during the grant could still misuse it.
You say “secrets manager”, “tenant”, “change request”. Project Orchestrator uses its own words, and a workspace is not a tenant boundary.
“the secrets manager, a credential store”
Vault
Where secrets are kept encrypted; an assistant gets one only through a grant limited in scope and time.
“the secrets vault, sharing policy”
Sharing
What leaves this workspace, under which policy, and the secrets assistants may use.
“the MCP server, federation”
MCP server
External tool servers the assistant can call, and the way other tools talk to PO.
“the model endpoint, the LLM backend”
Providers
A model other than Claude Code that you register, allow per project and give a key to.
“agents, clients”
Assistants
A conversation with an AI that works on your project; Today shows which ones are running.
“a tenant boundary”
Workspace
It groups projects. It does not isolate customers from one another: roles and access rights are Planned.
“pipelines, triggers”
Automation
Runs of the assistant across plans, and what starts them.
“a change record”
Decision
A choice that was made, with its rationale and alternatives.
Words this page leaves out : secure, end-to-end, compliant, certified
Secrets are sealed with XChaCha20-Poly1305 under a master key that is stored only wrapped by a key derived from your passphrase with Argon2id. The secret’s name is bound to the ciphertext, so a value copied under another name fails to open.
An unlock lasts at most 12 hours and a grant at most 30 days. Both must hold for an agent to read a value, and the module states that it is not a sandbox.
The default refuses mutating tools (an unclassified tool counts as one), allows 60 calls per minute and per server and can restrict the servers that may be called. Three environment variables loosen it; an unreadable value keeps the strict default.
The parent is read from the signed token, never from the request body. A child keeps the parent’s project and folder, its permission mode is clamped to the parent’s, depth is 1 and a parent has at most 4 live children.
A restricted session sees 22 of the 31 tools and the REST routes behind the others answer 403. The vault, admin, federation and sharing tools are withheld. A session of Claude Code has the full list.
CI runs cargo audit, cargo deny on advisories and licences, clippy with warnings as errors and the tests. The crate declares MIT AND BUSL-1.1 in Cargo.toml. None of this proves how the server behaves at runtime.
Nothing here is hidden: these are the things you might assume and that are not true yet.
The vault is not a sandbox, and an agent is not isolated from the machine it runs on: it works with the permissions you gave it.
Without an auth section, anyone who can reach the port is a user. The example compose files publish service ports and ship example passwords: harden them before exposing a host.
An external MCP server started over stdio inherits the server’s environment, and federation decides read or mutation from the tool name.
Sessions of Claude Code get the full tool profile, the admin tool included. The restricted profile is the default only for other providers.
Providers other than Claude Code, and the instance reached over SSH, have only been tried against test doubles, never a real service or machine.
Remote MCP over HTTP with an OAuth 2.1 server exists in the code but is not yet something to rely on.
No roles and no per-user rights, and no metrics endpoint that we found. Shared and enterprise controls are Planned.
The page on the spawn envelope and the one on providers state what is enforced and what remains open. Decide before you install.
Spawn envelopeUse Homebrew (brew install this-rs/tap/project-orchestrator), the desktop app, or the Docker image with the production compose file. Change the placeholder passwords.
DownloadAdd an auth section to the configuration before anything else, then check that /health answers and that a call without a token is refused. The repository’s authentication guide lists the three sign-in modes.
Open the vault, store a test secret, unlock it for a short time and grant it to one project. Read how it is masked.
Agent runtimeAdd it from the MCP federation screen, keep mutations blocked and list it in MCP_FEDERATION_ALLOWED_SERVERS. Then ask it for a mutating call and read the refusal.
No. An agent that holds a secret can send it somewhere else. The vault limits who receives a secret, for how long and for which project, and masks the value in what is stored. It reduces accidental exposure; it does not contain a hostile agent.
The graph, the index and documents are stored locally. What leaves: the conversation, which goes to the model through Claude Code (Anthropic’s own terms apply), and a periodic check of GitHub releases for updates. Embeddings are computed locally by default; they leave only if you set the embedding provider to http and point it at a remote service. A third-party model is used only after a per-project consent and each send is journalled. A search for analytics or telemetry clients found none; that is an absence of evidence, so read the code if it matters.
Password, Google and any OpenID Connect provider (Keycloak, Okta, Azure AD…), with tokens that last eight hours by default. Without an auth section in the configuration the server is open to one anonymous user, and the example production compose file ships with that section commented out. Roles and per-user access rights are Planned.
For a model other than Claude Code, yes: the restricted profile is the default and is enforced on the server routes. For Claude Code sessions the profile is the full one, including the admin tool. Opening sessions is bounded by the envelope in both cases.
Calls are filtered, but a tool is classified as read or mutating from its name, which is a heuristic. A server started over stdio also inherits the server’s environment, so do not run one you do not trust next to production secrets. Federated tool descriptions come from the remote server.
The project is MIT. Files explicitly marked Enterprise Edition fall under BUSL-1.1 (LICENSE-EE): non-production use is free, production use needs a commercial licence, and the files turn MIT after four years. No file is marked that way in the tree we read. This is a summary, not legal advice: read the licence files.
The repository ships a Docker image, a development compose file and a production one with the image, Neo4j, Meilisearch and NATS. The development file carries example passwords and the production one a placeholder default: set your own, enable auth and do not expose the ports as they are published.
Download
One-click install
There is no Project Orchestrator account to create.
Free · MIT license · Bring your own AISee what it costs
Version 0.0.19 · 13 Nov 2026
The app installs on a computer. Pick the one you use.
Prefer a terminal?
Homebrew · Mac and Linux. Installs orchestrator, orch and mcp_server.
brew install this-rs/tap/project-orchestratorDocker Desktop
Runs the database and the search engine the app relies on. The app starts them for you. Get Docker Desktop
Claude Code
The default engine for chat. The AI itself is not included: with Claude Code, you use your own Claude account, and the app can install it for you. Other engines (Codex, an OpenAI-compatible model, an ACP agent) are available from Settings — newer, in alpha. Get Claude Code
At first launch
A setup wizard in four short parts checks Docker, installs Claude Code if it is missing and downloads the search model.
Runs the two services that hold the project memory (Neo4j and Meilisearch).
Open the downloaded .dmg, drag Docker into Applications, then open Docker once and wait for the whale icon to stop moving.
Or in a terminal
brew install --cask dockerCheck that it works
Each line should print a version number. If you see « command not found », open a new terminal window and try again.
docker --versiondocker compose versionDocker must be running (open it once) before you start Project Orchestrator. Installation page of the publisher
The assistant Project Orchestrator works with. It uses your own Claude account.
Paste the command in a terminal. It installs everything it needs and asks for your password only if required.
Command
curl -fsSL https://claude.ai/install.sh | bashCheck that it works
Each line should print a version number. If you see « command not found », open a new terminal window and try again.
claude --versionRun claude once to sign in. Installation page of the publisher
Free · MIT license · Bring your own AI See what it costs
Try it: start a task in any window
Download
Pick the file for your system above. There is no account to create.
Open
Open the file. A setup wizard checks Docker, installs Claude Code if it is missing and downloads the search model.
Connect your AI
Chat runs on Claude Code by default; other engines are available from Settings (alpha). If it asks you to sign in, type this once in a terminal.
claudemacOS, Apple Silicon
The app for Macs with an M1 chip or later.
Project.Orchestrator_0.0.19_aarch64.dmg · 87.1 MB
macOS, Intel
The app for Macs with an Intel chip.
Project.Orchestrator_0.0.19_x64.dmg · 89.8 MB
macOS, Intel (macOS 10.15 to 13.3)
For Intel Macs on macOS 10.15 to 13.3. Without local embeddings: use an HTTP provider such as Ollama.
Project.Orchestrator_0.0.19_x64-legacy.dmg · 47.7 MB
Linux, AppImage
One file that runs on most Linux distributions (x86_64).
Project.Orchestrator_0.0.19_amd64.AppImage · 163 MB
Debian and Ubuntu, x86_64
Package to install with dpkg or apt.
Project.Orchestrator_0.0.19_amd64.deb · 98.1 MB
Debian and Ubuntu, arm64
Package to install with dpkg or apt, for ARM64 machines.
Project.Orchestrator_0.0.19_arm64.deb · 94.4 MB
Fedora and RHEL, x86_64
Package to install with dnf or rpm.
project-orchestrator-0.0.19-1.x86_64.rpm · 61.5 MB
Fedora and RHEL, arm64
Package to install with dnf or rpm, for ARM64 machines.
project-orchestrator-0.0.19-1.aarch64.rpm · 56.7 MB
Homebrew
Mac and Linux. Installs orchestrator, orch and mcp_server.
brew install this-rs/tap/project-orchestratorInstall script
Mac and Linux. Latest release into ~/.local/bin. Add --no-frontend for the light variant.
curl -fsSL https://raw.githubusercontent.com/this-rs/project-orchestrator/main/install.sh | shPowerShell
Windows. Installs into %LOCALAPPDATA%\orchestrator\bin.
irm https://raw.githubusercontent.com/this-rs/project-orchestrator/main/install.ps1 | iexDocker
Image for amd64 and arm64, with the web interface.
docker pull ghcr.io/this-rs/project-orchestrator:latest# API only, no dashboard: a version tag with the -api suffix# docker pull ghcr.io/this-rs/project-orchestrator:0.0.19-apiorchestrator, orch and mcp_server, with the web interface included. Lighter versions without it are on the release page.
macOS, Apple Silicon
Programs without the app window: orchestrator, orch, mcp_server
orchestrator-full-0.0.19-macos-arm64.tar.gz · 81.9 MB
macOS, Intel
Programs without the app window: orchestrator, orch, mcp_server
orchestrator-full-0.0.19-macos-x86_64.tar.gz · 82.8 MB
Linux, x86_64
Programs without the app window: orchestrator, orch, mcp_server
orchestrator-full-0.0.19-linux-x86_64.tar.gz · 91.1 MB
Linux, arm64
Programs without the app window: orchestrator, orch, mcp_server
orchestrator-full-0.0.19-linux-arm64.tar.gz · 85.6 MB
Windows, x86_64
Programs without the app window: orchestrator.exe, orch.exe, mcp_server.exe
orchestrator-full-0.0.19-windows-x86_64.zip · 92.8 MB
The desktop app starts Neo4j and Meilisearch in Docker for you. With the command line, one compose command does it: follow the quickstart.
More for developersWhat you should see. A list of checks. On the very first launch the Docker lines say “Setup pending”.
What to do. Wait for the setup to open: the start screen only checks what is installed.
What you should see. A green line says “Docker Desktop is running” under two cards, “Docker (recommended)” selected.
What to do. Keep “Docker (recommended)”. The app looks for Docker Desktop. When it says “Docker Desktop is running”, press Next. If it says Docker is missing or not running, use the button next to the message, then wait: the app checks again by itself. The services use the local ports 7474, 7687, 7700 and 4222, and the app itself uses 6600.
What you should see. Three choices (No authentication, Password, OIDC / OAuth), the first one selected.
What to do. On a computer only you use, keep “No authentication” and press Next. On a shared or public network, choose “Password” and set an email and a password: without sign-in, the app’s server accepts connections from your whole network.
What you should see. Claude Code CLI shows a green check once installed (until then: “Claude Code CLI is required” and an install button), and the search model is marked ready.
What to do. If “Claude Code CLI is required” is shown, press “Install Claude Code CLI”. Then let the app download the search model when it offers to; the size is shown on screen. Next unlocks when both are ready. If you have not signed in to Claude yet, see “If it does not work” below.
What you should see. A green “Configuration saved!” message and a “Restart Application” button.
What to do. Press “Generate Config & Save”. When “Configuration saved!” appears, press “Restart Application”. A switch on this screen controls whether the app checks for updates at start-up.
What you should see. A checklist (Docker Desktop, Neo4j, Meilisearch, NATS, Backend server) turning green one line at a time.
What to do. On this second start the app downloads the three images it runs on, then starts them. Later starts reuse the images and are quicker. When every line is green, the main window opens. If a line says “Taking longer than expected”, wait. If a “Continue anyway” button appears, see “If it does not work”.
macOS 11 or later on Apple Silicon (M1 and later). On Intel: macOS 13.4 or later, or macOS 10.15 to 13.3 with the older-macOS file, which has no local embeddings.
Download Docker Desktop from Docker’s page, open the file, drag Docker to Applications, then open it and wait until it says it is running.
Get Docker DesktopDocker Desktop is open and reports that it is running. It is Docker’s own screen, so it is not reproduced here.
Open the .dmg, drag Project Orchestrator to Applications, then open it from there. If macOS says it cannot check the app, we could not verify how the build is signed: allow it in System Settings, Privacy & Security, with “Open Anyway”.
The three problems you are most likely to meet, with what the app shows.
What you see. The app says “Docker Desktop is not running” or “Docker Desktop is required”.
What to do.
What you see. Neo4j or Meilisearch stays “Unhealthy”, or the app says it failed to start services. The services need the local ports 7474, 7687, 7700 and 4222.
What to do.
Only if you are comfortable with a terminal, this shows which program uses the port:
lsof -i :7687What you see. The setup says “Claude Code CLI is required”, or the chat says it has no Claude login.
What to do.
Type this and press Enter:
claudeWindows, 64-bit.
Download Docker Desktop from Docker’s page, run the installer, then start Docker Desktop from the Start menu and wait until it says it is running.
Get Docker DesktopDocker Desktop is open and reports that it is running. It is Docker’s own screen, so it is not reproduced here.
Run the installer and follow the wizard, then start Project Orchestrator from the Start menu. Windows may show “Windows protected your PC” because no signing certificate is set in the build configuration: choose “More info”, then “Run anyway” if you trust the download.
The three problems you are most likely to meet, with what the app shows.
What you see. The app says “Docker Desktop is not running” or “Docker Desktop is required”.
What to do.
What you see. Neo4j or Meilisearch stays “Unhealthy”, or the app says it failed to start services. The services need the local ports 7474, 7687, 7700 and 4222.
What to do.
Only if you are comfortable with a terminal, this shows which program uses the port:
netstat -ano | findstr :7687What you see. The setup says “Claude Code CLI is required”, or the chat says it has no Claude login.
What to do.
Type this and press Enter:
claudeAppImage for Intel/AMD; .deb and .rpm packages for Intel/AMD and ARM64.
Follow Docker’s page for your distribution, then start Docker and wait until it says it is running. The app also recognises Docker installed without the Desktop application.
Get Docker DesktopDocker is running. It is Docker’s own screen, so it is not reproduced here.
For the AppImage: in your file manager, open the file’s properties and allow it to run as a program, then double-click it. For a .deb or .rpm: open it with your software installer, then start Project Orchestrator from your applications menu.
The three problems you are most likely to meet, with what the app shows.
What you see. The app says “Docker Desktop is not running” or “Docker Desktop is required”.
What to do.
What you see. Neo4j or Meilisearch stays “Unhealthy”, or the app says it failed to start services. The services need the local ports 7474, 7687, 7700 and 4222.
What to do.
Only if you are comfortable with a terminal, this shows which program uses the port:
lsof -i :7687What you see. The setup says “Claude Code CLI is required”, or the chat says it has no Claude login.
What to do.
Type this and press Enter:
claude