MCP Core Specification

Owner: PhantomPM Integrations Team
Last Updated: 2026-02-13
Status: Beta

Goal

Define PHANTOM as a local MCP provider with stable, versioned capabilities for coding agents and IDE clients.

Current Implementation State

1. MCP server package exists at packages/mcp-server.
2. Stdio server path is available via phantom mcp serve.
3. Tool contracts implemented: context.add, context.search, prd.generate, swarm.analyze, bridge.translate_pm_to_dev.
4. Integration data model exists in config layer.

Target State

A local MCP core service with:

1. capability discovery
2. tool execution
3. resource retrieval
4. permission-aware behavior

Initial Capability Surface

Tools:

1. context.add
2. context.search
3. prd.generate
4. swarm.analyze
5. bridge.translate_pm_to_dev

Resources:

1. active project summary
2. installed modules summary
3. capability matrix snapshot

Tool Contract Shape

Each tool request includes:

1. tool
2. arguments
3. request_id

Each response includes:

1. request_id
2. status
3. result
4. errors[] when applicable

Initial Tool Input/Output Contracts

1. context.add
   • input: path, optional mode
   • output: stats with files/types/languages/health
2. context.search
   • input: query, optional limit
   • output: matches[] with path, snippet, and type
3. prd.generate
   • input: title, optional scope, optional constraints[]
   • output: prd_id, markdown, optional output_path
4. swarm.analyze
   • input: question, optional context_scope, optional agents[]
   • output: consensus, confidence, agent_reports[]
5. bridge.translate_pm_to_dev
   • input: pm_intent, optional product_constraints
   • output: technical_tasks[], acceptance_criteria[], risks[]

Permission Model

1. default allow for local read actions
2. explicit policy checks for external side effects
3. denial responses include remediation guidance

Observability

1. request duration metrics
2. error class counts
3. adapter-level success/failure counters

Acceptance Criteria

1. MCP capability listing is deterministic
2. tool responses conform to documented schema
3. permission denials are explicit and auditable