puzzlebox
!puzzlebox
Coordinating agents with state machines
An MCP server that hosts finite state machines as dynamic resources that clients can subscribe to and be updated when their state changes.
What problem does puzzlebox address?
Collaboration and coordination are related but different problems. Collaboration is good for non-trivial yet relatively simple tasks. To tackle long horizon efforts, puzzlebox solves for coordination.
Teams need coordination
Marshalling multiple agents toward a big goal is tougher than just breaking down a request into tasks, assigning them to available agents and enabling collaboration between them.
Just as a few agents can collaborate to complete a small project, several teams of process-aware agents need to operate within distinct project phases to tackle long horizon efforts.
Consider enterprise-level software development processes:
-
A large software project typically moves through a multi-step, occasionally backtracking path from inception to design to building to testing to documentation to marketing to production.
-
Different teams are focused on different aspects over time, informed by what's gone before and with an eye toward an ever-changing goal that is refined according to lessons learned.
-
Even within a phase, teams may cycle through their own phases, like Agile sprints. A certain amount of work is scoped for the sprint, the team works on their parts, and at the end of the sprint they decide what to tackle next. It accepts that each sprint could change the course of future development. These cycles can also be represented as puzzles.
With puzzlebox, members of agentic teams can be made process-aware, but the process itself is not subject to hallucination.
Scenario: Teams passing the torch
Three agents are working. The current state of their shared puzzle is "Specification".
- Agent 1 is specifying the domain language.
- Agent 2 is defining project scope.
- Agent 3 is producing the specification document.
- The agents collaborate to reach the final specification document.
- Once the spec is done, Agent 3 initiates a transition to "Design" state.
- First, the spec is checked by an exit guard (i.e., LLM sampling) for completeness.
- If problems are found, the state transition is canceled and the team continues.
- If acceptable, the state changes to "Design".
- The "Specification" agents are monitoring the puzzle and should clock out now.
- Their long (and expensive) contexts have been distilled into the specification.
- The "Design" team picks from here, with the spec as a resource and their contexts fresh and role-specific.
What is a puzzle?
A puzzle is a finite state machine. It's just easier to say, write, and think about.
A stateful thing you can act upon
Imagine the Rubik's Cube puzzle. It has 43 quintillion states, and to transition between them, you act upon it by rotating the intersecting planes of the mechanism.
Properties of a puzzle
- A finite number of discrete states, e.g., "Series Concept and Tone", "World Building", "Arc Plotting", "Episode Planning", "Plotline Blending", "Episode Outline", "Script Writing" etc.
- Each state may have any number of actions (including 0) that initiate transition to another state.
- There is an initial state.
- There is a current state that may differ after actions have been performed on the puzzle.
- Transitions can be canceled by state exit and enter guards, e.g., Consult LLM via client sampling request.
A Simple Example
{
"initialState": "LOBBY",
"states": {
"LOBBY": {
"name": "LOBBY",
"actions": {
"START_GAME": { "name": "START_GAME", "targetState": "PLAYING" }
}
},
"PLAYING": {
"name": "PLAYING",
"actions": {
"END_GAME": { "name": "END_GAME", "targetState": "GAME_OVER" }
}
},
"GAME_OVER": {
"name": "GAME_OVER",
"actions": {
"RESTART": { "name": "RESTART", "targetState": "PLAYING" }
}
}
}
}
What is puzzlebox?
Most MCP servers have a one-to-one relationship with the client. Puzzlebox is different.
Many clients sharing dynamic resources
Puzzlebox is an MCP Server implementation that:
- Supports multiple client connections that can create and monitor shared, dynamic resources.
- Manages puzzle instances
- Exposes tools for:
- Adding puzzles
- Getting a snapshot of the state and available actions for a given puzzle in the box
- Performing actions on a given puzzle in the box that trigger state transitions
- Exposes registered puzzles as resources
- Clients can use the
Puzzle Snapshot resource template to fetch the resource by ID
- Resource URI is
puzzlebox:/puzzle/{puzzleId}
- Clients can subscribe/unsubscribe to individual resource URIs
How It Works
- Clients connect to a puzzlebox SSE server.
- Clients register puzzles with the server.
- Clients can subscribe to a given puzzle to receive updates when its state changes.
- Clients perform actions on puzzles that may change their state and available actions.
- The puzzlebox server ensures that any attempted action is valid for the current state of the given puzzle.
- If an action is valid, a transition to the target state is initiated.
- During transition, optional exit and enter guards may send sampling requests to the client, the results of which could lead to cancellation of the transition (think acceptance testing by stakeholders)
- If guards pass, the state transition completes.
- When a client receives a resource updated notification, they can either read the resource or use the
get_puzzle_snapshot tool to get the current state and available actions.
- Clients update their UI based on the new state.
MCP Tools
These functions are exposed to the agents for managing puzzles.
⚙️ add_puzzle
Add a new instance of a puzzle (finite state machine).
- Inputs: None
- Returns: JSON object with boolean
success and puzzleId
⚙️ get_puzzle_snapshot
Get a snapshot of a puzzle (its current state and available actions).
- Inputs:
puzzleId
- Returns: JSON object with
currentState and availableActions array
- Note: MCP clients that don't support resource subscriptions can poll this tool to watch for state changes.
⚙️ perform_action_on_puzzle
Perform an action on a puzzle (attempt a state transition).
- Inputs:
puzzleId and actionName
- Returns: JSON object with
currentState and availableActions array
⚙️ count_puzzles
Get the count of registered puzzles
- Inputs: None
- Returns: JSON object with current
count of registered puzzles
Local Setup
Running locally requires Node and npm be installed. Then follow these steps...
Install Dependencies
cd /path/to/puzzlebox/npm install
Build
npm run build- Builds the MCP server runtime at
/dist/index.js
Start
npm run start- Launches an SSE-based/MCP server on port
:3001 with endpoint /sse
- MUST BE LAUNCHED BEFORE RUNNING INSPECTOR
Inspector
Format
npm run format- Runs
prettier on the code, adjusting formatting
Typecheck
npm run typecheck- Runs
tsc with args to check and report type issues
Lint
npm run lint- Runs
eslint to non-destructively check for and report syntax problems
LintFix
npm run lint:fix- Runs
eslint to check for and fix syntax problems
Test
npm run test- Run the unit tests
Screenshots
These screenshots show the various MCP tools and resources implemented by the sever.
Testing of the server was done with the official reference client - the MCP Inspector.
0 - List Tools
!0. list_tools
1 - Add Puzzle
!1. add_puzzle
2 - Get Puzzle Snapshot (Initial State)
!2. get_puzzle_snapshot
3 - Perform Action On Puzzle
!3. perform_action_on_puzzle
4 - Get Puzzle Snapshot (New State)
!4. get_puzzle_snapshot
5 - Perform Action On Puzzle
!5. perform_action_on_puzzle
6 - Get Puzzle Snapshot (Another New State)
!6. get_puzzle_snapshot
7 - List Resources
!7. list resources
8 - Resource Template
!8. resource_template
9 - Unsubscribed Resource
!9. unsubscribed resource
10 - Subscribed Resource
!10. unsubscribed resource
11 - Resource Updated Notification
!11. subscribed resource updated
