Skip to content

Use cases

It plugs into your tools and remembers what works

Project Orchestrator connects to your tools, follows procedures you write, and keeps what it learns. Here is what that looks like on real work, for developers and for everyone else.

Each capability says where it stands: available, in progress or planned.

  • Available
  • In progress
  • Planned

Plugs into anything

Plugs into what you already use

If a tool speaks MCP, an assistant can call it through Project Orchestrator. If a procedure fits states and triggers, it can be a protocol. Each entry says what works today.

ProtocolsHooksMCP serversAI enginesCodeKnowledge graph

Try: press a plug to see what it does today.

  • Available

    Any procedure you can write as states and triggers: a release, a review, a triage. Every run is recorded.

  • Available

    Actions that run when a task, step or plan changes status, and hooks inside the assistant's own tool loop.

  • Available

    MCP servers over stdio, SSE or streamable HTTP. Assistants call them by name, and calls that change things are blocked by default.

  • In progress

    Claude Code is the default. The native engine (any OpenAI-compatible model), Codex, ACP agents such as opencode and Claude Code over SSH can run an agent too, and a conversation can move from one engine to another. These engines are alpha: tested against stand-ins, not the real services yet.

  • Available

    Seventeen languages parsed into a graph: who calls what, what a change touches, how the project is laid out.

Project Orchestrator is an MCP server too: Claude Code or any MCP client reaches it through 33 tools.

What makes it learn

What the system keeps, and how

No magic here. Each card says what happens, when it happens, and what stays manual. Open it to read the mechanism.

  • Procedures as state machines

    Available

    A protocol is a state machine: states, and transitions that each need a named trigger. Every run is recorded with the states it went through.

    Mechanism and limits

    What happens

    A run starts by hand, on an event (after a project sync or a skill import) or on a schedule (hourly, daily or weekly). A runner can drive it to its end, and a state that needs judgment is handed to a Claude session.

    What stays manual

    Guards written on a transition are notes for the assistant today: they are not evaluated. Only the Proposal procedure is created by PO itself; the others you compose, or take from the guide.

  • Proposals, from draft to implemented

    Available

    A Proposal is a note of type RFC whose life is a state machine of nine states, from draft to implemented, rejected or superseded.

    Mechanism and limits

    What happens

    Writing the Proposal starts its run. Each step (propose, submit for review, accept, start planning, start work, complete) is a named trigger that you or the assistant fires, and the run keeps the history.

    What stays manual

    Linking a Proposal to a plan is a call the assistant makes. Accepting a Proposal does not create the plan's tasks by itself.

  • A memory that ages

    Available

    Notes and decisions carry an importance, an energy and a freshness, so what is old and unused loses weight instead of piling up.

    Mechanism and limits

    What happens

    Background checks run on their own: freshness every hour, consolidation every 2 hours (notes nobody uses are archived, not deleted), energy and link decay every 6 hours. A project idle for 14 days is frozen, so a break does not erase it.

    What stays manual

    A note flagged as needing review is not rewritten for you. Someone confirms it, corrects it or retires it.

  • Available

    Notes close in meaning are linked automatically. Each time notes are used together the link between them gets stronger, and links nobody uses fade.

    Mechanism and limits

    What happens

    A new note is compared with the others and linked to its close neighbours. Notes handed to the assistant, searched or touched by a commit gain energy and stronger links, a few hundredths per use and capped. Unused links lose weight every 6 hours and disappear under a floor, and a job rebuilds the ones that are still similar.

    What stays manual

    Automatic linking needs an embedding provider: without one, notes are linked by hand only. A search follows links two steps from its first matches, not further.

  • Skills and personas

    Available

    A skill is a cluster of related notes that is handed to the assistant when the work touches its topic. A persona is an assistant specialised on one area, with its own limits.

    Mechanism and limits

    What happens

    Clusters are detected every day from the links between notes (at least 15 notes in the project, 3 in a cluster). A skill goes from emerging to active once its notes are close and have been used a couple of times, and to dormant after 30 days unused. Skills and personas can be exported and imported.

    What stays manual

    Building a persona from the project is a call you or an assistant makes: nothing schedules it. The system does not yet learn which skills turned out useful.

  • The improvement loop

    In progress

    When a protocol run or a plan ends, PO records an episode: what started it, the states it went through and what it produced. From three similar episodes it can write notes and add a check to a procedure.

    Mechanism and limits

    What happens

    Episodes are collected on their own. The analysis needs at least three similar ones; it writes at most 10 notes per pass and 50 per project, tagged auto-learned, skips near-duplicates, and records each change to a procedure as a Decision.

    What stays manual

    It only learns from runs that went through protocols or the plan runner. Nothing measures yet that results improve, so read the auto-learned notes. The reflection steps in a protocol are instructions: the assistant may or may not follow them.

  • Hooks on every status change

    Available

    A lifecycle hook runs an action when a task, step, plan or objective changes status: complete the steps below it, create a note, raise an alert or start a protocol.

    Mechanism and limits

    What happens

    Hooks are stored per project or for all, and run by priority; a failing hook is logged and the others still run. One built-in hook completes a task's steps when the task completes. Inside the assistant's own tool loop, hooks before and after each tool call, and before the context is compacted, hand it the right skills and notes.

    What stays manual

    Hooks fire on status changes only. An action that calls an outside tool is not implemented yet, so it is not offered.

  • Trajectories and neural routing

    In progress

    PO records the path of each session, which tools were called and how it went, and can search past paths that look like a new request.

    Mechanism and limits

    What happens

    Recording is on by default and covers tool calls over MCP and in chat. Neural routing finds the nearest past paths for a request and logs them for the dashboard.

    What stays manual

    The routing result does not change yet which tools or prompts the assistant gets, and no training loop runs in the product. Today you query the paths yourself, with search similar or stats.

Use cases

Six jobs, step by step

Three for developers, three for everyone else. Each one shows the situation before, what Project Orchestrator does, and the result.

For developersAvailable

A red CI, triaged without starting from zero

A pipeline fails on main. You want the cause, a fix, and a trace for next time.

Before

Each red build opens a new conversation. The assistant reads the same logs again and does not know this test failed the same way last month.

Protocol · ci-triage
  1. reproduce
  2. locate
  3. fix
  4. verify
Read the logs of the failing run
Completed
ci::get_run_logsread-only
What does the failing change touch?
In progressP8
code · analyze_impact
session-expiry test fails when the clock moves
Active
gotchasession.rs

The cause is saved as a gotcha note linked to the file. The next time an assistant touches that file, the note is handed back before it edits.

Try: go through the steps, or compare before and after.

Step 3: Protocol · ci-triage

What Project Orchestrator does

  1. You connect your CI's MCP server. The assistant can read runs and logs through it; calls that change things stay blocked unless you allow them.
  2. A protocol you composed, ci-triage, walks the run through reproduce, locate, fix and verify. Each move needs a named trigger, and code · analyze_impact lists what the failing change touches.
  3. The cause is saved as a gotcha note linked to the file. The next time an assistant touches that file, the note is handed back before it edits.

Result

A fix task with its steps, a recorded run you can reopen, and a note that outlives the conversation.

What stays manual

You write the ci-triage protocol once (the guide has templates) and you connect the CI server. PO does not watch your CI by itself.

Built on MCP servers · Protocols · Code graph · Ageing memory

For everyoneAvailable

A team offsite, with a dozen threads in one plan

Venue, catering, invitations, trains for 24 people. Each one is a conversation that can get lost.

Before

You juggle conversations with assistants and a spreadsheet. You forget who is waiting for what, and what was already decided.

Plans

Organise the autumn team offsite

In progress
1/4
Venue and booking
In progressP8
Catering quotes
In progressP8
Guest invitations
PendingP5
Hold the offsite in November rather than October
Accepted
venue, invitations

Each assistant that works on the plan finds the earlier decisions and notes, such as diets and budget per person, instead of asking again.

Try: go through the steps, or compare before and after.

Step 3: Plans

What Project Orchestrator does

  1. You give one goal. PO turns it into a plan of tasks with dependencies: venue and catering quotes start together, invitations wait for the dates.
  2. Choices become Decisions, with the reason. An assistant that needs your permission or an answer asks, and it shows in Today.
  3. Each assistant that works on the plan finds the earlier decisions and notes, such as diets and budget per person, instead of asking again.

Result

One screen shows what is done, what runs and what waits for you, and the choice of November is written down once.

What stays manual

PO plans and keeps the thread. Booking a venue still needs a tool you connect, or you doing it.

Built on Plans and decisions · Ageing memory

For developersAvailable

From a Proposal to a plan and its tasks

A change is discussed, accepted, then built. The decision should not live in a chat that scrolled away.

Before

"We should move the session store to Redis" is said in a conversation, then lost. Weeks later nobody knows whether it was accepted.

Proposals
  1. draft
  2. proposed
  3. under review
  4. accepted
  5. planning
  6. in progress
  7. implemented
Move the session store to Redis
Accepted
RFC
Migrate session store to Redis
Draft
linked plan4 tasks

Once it is accepted, the assistant creates a plan and links it to the Proposal. Tasks get dependencies, so independent ones run together.

Try: go through the steps, or compare before and after.

Step 2: Proposals

What Project Orchestrator does

  1. You write the Proposal, an RFC note. Its run starts at draft, and every move (propose, submit for review, accept) is a named trigger.
  2. Once it is accepted, the assistant creates a plan and links it to the Proposal. Tasks get dependencies, so independent ones run together.
  3. As work starts and ends, the Proposal moves to in progress, then implemented, and its history shows each step.

Result

Anyone can see whether the change was accepted, why, and which plan carries it out.

What stays manual

The link between a Proposal and its plan is a call the assistant makes: accepting a Proposal does not create tasks by itself.

Built on Proposals · Plans and decisions · Protocols

For everyoneAvailable

A funding application, section by section

Guidelines, a budget, letters of support and a deadline. Nothing here is code.

Before

The call's rules sit in a PDF, the figures in a spreadsheet, and the assistant forgets both between conversations.

Notes
The call closes on 15 March
Accepted
timeline, every section
Apply to the regional call
Accepted
call-guidelines.pdf
The budget cap is 50,000
Needs review
budget.xlsx
Letters of support: two received, one missing
Active

Notes age. One that was not confirmed for long enough is flagged as needing review, so an old budget cap is checked before it reaches the text.

Try: go through the steps, or compare before and after.

Step 3: Notes

What Project Orchestrator does

  1. You create a project without code and give the goal. PO plans one task per section of the application, with the deadline as the objective.
  2. Rules and figures become notes and decisions: "the call closes on 15 March", "apply to the regional call".
  3. Notes age. One that was not confirmed for long enough is flagged as needing review, so an old budget cap is checked before it reaches the text.

Result

A checklist by section, the facts the assistant relies on, and a flag on those that may be out of date.

What stays manual

Someone confirms or corrects the flagged notes. PO flags what is old; it does not know which figure is still true.

Built on Plans and decisions · Ageing memory

For developersIn progress

A release cycle that starts by itself

The last task is done, and the release checklist should begin without anyone having to remember it.

Before

The checklist lives in someone's head. Steps get skipped when the week is busy, and each release runs a little differently.

Protocol · release-checklist
  1. freeze
  2. verify
  3. tag
  4. announce
Release 0.9 plan completed
Completed
on completed: start_protocol
release-checklist
Running
verify
The changelog check keeps failing at the verify state
Active
auto-learned

Each run is recorded as an episode. After three similar episodes, PO can write notes about what keeps failing and add a check to the procedure.

Try: go through the steps, or compare before and after.

Step 3: Protocol · release-checklist

What Project Orchestrator does

  1. You compose a release-checklist protocol: freeze, verify, tag, announce.
  2. A lifecycle hook starts it when the release plan becomes completed. Another built-in hook already completes a task's steps when the task completes.
  3. Each run is recorded as an episode. After three similar episodes, PO can write notes about what keeps failing and add a check to the procedure.

Result

The checklist starts by itself, every run leaves a history, and repeated failures become notes.

What stays manual

The first two steps are available now. The last one is in progress: it only learns from runs that go through protocols, and nothing measures yet that releases get better. Read the auto-learned notes.

Built on Hooks · Protocols · Improvement loop

For everyoneAvailable

A trip for four, where each trip starts from the last

Flights, lodging, an itinerary, and preferences you do not want to repeat.

Before

Each trip is planned from scratch: tabs, messages, the same questions about food, dates and budget.

Skills
travel-preferences
Two vegetarian meals · no night trains · a budget per person
Emerging
EnergyMediumCohesionMedium
airline-loyalty
Dormant
EnergyLow
Plan the trip to Lisbon in May
In progress
2/4

With enough notes on the topic, a skill forms from the cluster and is handed to the assistant when you talk about a trip. Unused for 30 days, it goes dormant.

Try: go through the steps, or compare before and after.

Step 3: Skills

What Project Orchestrator does

  1. You give the goal and the dates. PO plans the tasks: flights and lodging can be searched together, the itinerary waits for both.
  2. Preferences you state become notes: two vegetarian meals, no night trains. Notes used together get linked, and the links get stronger each time.
  3. With enough notes on the topic, a skill forms from the cluster and is handed to the assistant when you talk about a trip. Unused for 30 days, it goes dormant.

Result

The next trip begins with what the last one taught, and preferences you stop using fade.

What stays manual

It takes a project with enough notes (at least 15) and an embedding provider for the automatic links. PO plans and remembers; it does not book anything.

Built on Plans and decisions · Links · Skills

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