Features
What Project Orchestrator does
Plans, tasks, notes, decisions and specialized assistants, kept together in one memory for each project. See what Project Orchestrator does for you.
Projects, with or without code
A project holds its own plans, tasks, notes and decisions. It can be a codebase, or a piece of work with no code, such as an event or a budget.
- Shows all your projects on one screen, Today: what waits for your answer, what to resume, what moves on its own.
- Groups several projects in a workspace that shares context and objectives.
- A project with no code is created by asking an assistant for now. The project form in the app does not offer the choice yet.
Under the hood
A project has a profile: software is a codebase that is synced and watched, work is documents, plans, tasks and notes with no source tree.
- project · create
- workspace · list
- project · get_roadmap
Try: filter by type, then open a project.
Objectives, plans and tasks
An objective is what you aim for. A plan says how to get there, as tasks and steps. PO shows what is done, what is blocked and what comes next.
- Tracks the status of each step, task and plan, with the progress added up.
- Shows the longest chain of tasks, so you know what sets the end date.
- Shows the same plan as a map of links or as a board with columns.
- In the chat, type # to point at a plan, a task, a note, a decision or an RFC, or drag a card into the conversation: the agent sees exactly what you mean. Behind a feature flag for now.
Under the hood
The longest chain is the critical path, computed from the dependency graph with plan(action: "get_critical_path").
- plan · get_critical_path
- task · add_dependencies
- step · update
Organize the team weekend
In progressTry: click a task to move it forward and watch the progress.
Assistants that work together
PO groups the tasks that do not depend on each other and gives each one to its own assistant. Four tasks run at a time by default. PO checks the result before the next group starts.
- Shows live which tasks have started and which are finished.
- Stops a run at any time, and keeps the list of past runs with their status.
- Starts a plan on a schedule or when something happens, with nobody at the keyboard.
- A schedule, webhook or event trigger on a plan now starts a run of it, one run per signal and with a cooldown. Reacting to CI, releases and pull requests is still in progress.
- Stop reaches the tools still running on the native engine, and a cancel that fails is shown as failed.
Under the hood
plan(action: "run") starts the runner on its own git branch, from a project folder. The build check runs for Rust, Node and Go projects.
- plan · get_waves
- plan · run
- plan · run_status
Try: press Run to play the waves one after the other.
Notes and decisions
Notes and decisions belong to your project, its tasks and its files. Each one has a status (active, needs review, stale or obsolete), so assistants know what to trust.
- Saves a decision with its reason and what it affects, and replaces it when it changes.
- Finds a note by meaning, not only by exact words, and lists the notes that need a review.
- A proposal puts a change up for review before it is decided.
Under the hood
Notes and decisions are nodes in a Neo4j graph. Notes used together are linked by synapses that grow stronger, and a search follows those links.
- note · search_semantic
- decision · add_affects
- note · get_needing_review
Try: search the notes, then accept or reject the proposed decision.
Documents
Documents are filed under a project. The original file stays on your computer, and PO cuts the text into passages that remember where they came from.
- Upload a file to a project, open the original again later, or delete it.
- Reads the text of PDF and text files.
Under the hood
Originals are stored on local disk, addressed by their SHA-256. The text is cut into chunks that keep their position in the source.
Try: filter by type or change the order.
Specialist assistants
A persona is an assistant specialized in one area of your work. A skill is a group of related notes that the assistant switches on by itself when the topic comes up.
- Picks the persona that fits the task.
- A persona sets a time limit, a cost limit and the model for its tasks.
- A skill goes from emerging to active, and to dormant after 30 days without use. Skills can be exported and imported.
Under the hood
Skills are detected from clusters of notes (Louvain community detection). Each one has an energy and a cohesion score. The persona builder offers code-based modes and a manual mode.
- persona · get
- skill · activate
- skill · export
Try: use a skill, or let a week pass.
Repeatable procedures
A protocol is a procedure with a start, steps and an end. The assistant follows it in order, and each run is recorded.
- Define the steps and what moves a run from one step to the next, then start a run and follow it.
- Link a protocol to a skill, so it starts when that skill applies.
- A run can wait for a person to review before it goes on.
Under the hood
A protocol is a finite state machine: start, intermediate and terminal states, and transitions fired by triggers. plan-runner-reviewed adds an awaiting_review state.
- protocol · add_state
- protocol · start_run
- protocol · link_to_skill
Try: advance the run, or make the gate fail.
Your choice of AI, or let PO choose
One menu in the chat sets it all. Switch Auto on and PO picks the model for each task or turn; switch it off and tick the models you allow: one ticked model is strict, several is mixed. Claude Code is built in, and you can add an OpenAI-compatible endpoint such as DeepSeek, Codex, an ACP agent such as opencode, or a Claude Code on another machine reached over SSH. You allow each one project by project.
- The menu is there at any time in a conversation, including one already started, and the choice stays local to that conversation. A choice you make yourself is never replaced by PO.
- In Auto, PO weighs tools, context size, budget and consent, records why it chose, and learns from the results. A shadow report shows what it would have chosen and what that would have saved, before you let it decide.
- A conversation can move to another provider, in either direction, and keep its history. In Auto, PO can move it before a turn.
- The assistant that delegates a task can name the provider and the model for that task.
- A key is never typed into the form: it is kept in the vault and granted to one provider. If the vault is locked, you unlock it from the provider picker without leaving the chat.
- Costs are counted in two ways: money really spent, and the notional value of a subscription or a free model. An unknown price is never shown as zero.
- Early: the engines beyond Claude Code are not all measured against the real services yet. Codex is reached for real (its own health check, the machine’s login) and an ACP agent gets a system prompt; a full turn over Codex or opencode is still unmeasured. A Claude Code on another machine does not get PO tools in its session.
- A conversation timeline shows what happened, turn by turn, so you can pick up the thread.
- Approve a tool call once, for the session, or always: PO keeps and checks the lasting rule itself, and the button shows exactly the call it covers.
- On the native engine, Codex and ACP, a turn cut short by a server restart resumes by itself and a tool is never replayed. A Claude Code turn is closed with a visible error, and your next message picks the conversation up.
- Images you attach reach the model on both engines, Claude Code and the native one. On the native engine, each model's own vision ability decides.
- A native model can ask you a question in the middle of a task, or hand part of it to a sub-agent, as Claude Code does.
Under the hood
Instance kinds: openai_compatible (presets deepseek, vllm, ollama, llama_server, nim), anthropic (the Anthropic Messages API, its credential kept in the vault), codex, acp and claude_code_remote. They run on the agent engine, with a consent tied to the origin and the credential reference. A conversation carries its own routing mode (primary, mixed or full for Auto). Resolution order: session, request, task, run, project role, global role, default, Claude Code.
- plan · delegate_task
Claude Code: I split the weekend into four tasks. Nothing is booked yet.
Try: choose the default assistant, allow or block a provider.
Reachable from anywhere, on your phone too
PO can run on a server of yours and be reached through a web address of your choosing. You sign in with your usual account, and the interface adapts to a phone screen, so you can check progress or answer a request away from your desk.
- Sign in with your existing account (Google, Microsoft, Okta and other OpenID Connect providers), or with a password.
- You can limit access to specific e-mail addresses.
- The interface is built to adapt to a phone screen.
- Turn sign-in on before opening PO to the internet: without it, anyone who can reach the address can use it.
- The interface is translated into 13 languages, screen group by screen group; the remaining screens follow.
Under the hood
OpenID Connect client (auth/oidc.rs), e-mail allow-list (allowed_emails), password login and JWT sessions. With no authentication configured, the API accepts every request.
Plan the team weekend and find a venue.
Three venues fit 12 people. The lake house is the cheapest. Shall I book it?
Try: switch the device, sign out, answer from your phone.
For developers
PO also reads your code
PO turns each repository into a graph of files, functions, types and calls, and answers who calls a function and what a change affects. Every concept is a tool for Claude Code, Cursor or any MCP client.
Under the hood
- Parses Rust, TypeScript, Python, Go, Java, C and C++, Ruby, PHP, Kotlin, Swift, Scala, Zig, Dart and more.
- Searches code through Meilisearch, across one project or a whole workspace.
- Connects other MCP servers through mcp_federation.
- Graph content put in a prompt is fenced as data the model is told not to obey as instructions.
Tree-sitter parses the code into a Neo4j graph, and a file watcher keeps it in sync as you edit.
- code · get_call_graph
- code · analyze_impact
- code · find_references
- mcp_federation · list
Get started
Give your agents a memory and a plan
Install the orchestrator, connect your assistant over MCP, and let your first agent read the project before it writes a line.
The app is free. You only pay for the AI you already use, such as your own Claude Code subscription or a key for another model. See what it costs
Or start from the terminal
brew install this-rs/tap/project-orchestrator