@xjtlumedia/context-first-mcp-server

Core library powering the Context-First MCP ecosystem — 37 tools across 7 layers for context health, memory, reasoning, and truthfulness.

Looking for the ready-to-run MCP server? Use context-first-mcp via npx context-first-mcp — zero install needed.

This package is the shared core library consumed by the stdio and remote servers. Use it if you are building your own MCP transport layer or embedding Context-First tools in a larger application.

Installation

npm install @xjtlumedia/context-first-mcp-server
# or
pnpm add @xjtlumedia/context-first-mcp-server

Node ≥ 18 required.

What This Package Exports

createServer(options?)

Creates a fully configured McpServer instance with all 37 tools registered, MCP prompts attached, and server-level usage instructions injected (visible to LLMs during handshake).

import { createServer } from "@xjtlumedia/context-first-mcp-server";

const server = createServer({ name: "my-context-first-server", version: "1.0.0" });

// Connect to any MCP transport
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
const transport = new StdioServerTransport();
await server.connect(transport);

State & Session

import { SessionStore, SiloManager } from "@xjtlumedia/context-first-mcp-server";

const store = new SessionStore();       // Per-session key-value + history
const silos = new SiloManager(store);   // Multi-agent quarantine silos

Tool Catalog (TF-IDF Discovery)

import { ToolCatalog, TfIdfIndexer } from "@xjtlumedia/context-first-mcp-server";

const catalog = new ToolCatalog();
// catalog.registerBatch([...]) — register tools for semantic routing
// catalog.query("search query", topK) — returns relevant tools only

Persistent Memory

import { UnifiedMemoryManager } from "@xjtlumedia/context-first-mcp-server";

Type Exports

All state, registry, and memory types are re-exported:

import type {
  // State types
  SessionState, HistoryEntry, GroundTruth,
  // Registry types
  ToolEntry,
  // Memory types
  MemoryEpisode, MemoryTier,
} from "@xjtlumedia/context-first-mcp-server";

Tool Layers

| Layer | Tools | Description | |-------|-------|-------------| | Orchestrator | context_loop | 8-stage pipeline: ingest → recap → conflict → ambiguity → entropy → abstention → discovery → synthesis | | Layer 1 · Context Health | 8 tools | recap_conversation, detect_conflicts, check_ambiguity, verify_execution, entropy_monitor, abstention_check, detect_drift, check_depth | | Layer 1b · State | 4 tools | get_state, set_state, clear_state, get_history_summary | | Layer 2 · Sandbox | 3 tools | discover_tools, quarantine_context, merge_quarantine | | Layer 3 · Memory | 6 tools | memory_store, memory_recall, memory_compact, memory_graph, memory_inspect, memory_curate | | Layer 4 · Reasoning | 5 tools | inftythink_reason, coconut_reason, extracot_compress, mindevolution_solve, kagthinker_solve | | Layer 5 · Truthfulness | 7 tools | probe_internal_state, detect_truth_direction, ncb_check, check_logical_consistency, verify_first, ioe_self_correct, self_critique | | Research & Export | 2 tools | research_pipeline, export_research_files |

Architecture

This package is the shared core consumed by both transport layers:

@xjtlumedia/context-first-mcp-server   ← this package
├── All 37 tool implementations
├── SessionStore · SiloManager
├── UnifiedMemoryManager
├── ToolCatalog · TfIdfIndexer
├── Engine: loop-freshness, NLP utils
└── Research pipeline (autonomous file writing)

        ↑ imported by
        │
        ├── context-first-mcp         (npx / stdio transport)
        └── @xjtlumedia/context-first-remote-server  (Vercel / HTTP transport)

Usage Protocol

When createServer() connects to an LLM-facing client, the CONTEXT_FIRST_INSTRUCTIONS are injected at handshake time via ServerOptions.instructions. The LLM receives:

• Mandatory context_loop call cadence
• Research workflow: interleave one search → one research_pipeline(gather) → repeat
• Memory preservation guide (memory_store / memory_recall)
• Autonomous file writing: pipeline writes *.batch-N.topic.md, *.analysis.md, *.synthesis.md to disk without LLM cooperation

Research Pipeline — Autonomous File Writing

research_pipeline({ action: "init",     topic, outputDir? }) → state initialized, outputDir confirmed
research_pipeline({ action: "gather",   content, topic })   → batch-N.topic.md written to disk
research_pipeline({ action: "analyze",  problem })          → analysis.md written to disk
research_pipeline({ action: "verify",   claims })           → health gate (non-blocking)
research_pipeline({ action: "finalize" })                   → synthesis.md + all batch files on disk

outputDir defaults to ./context-first-research-output/ when not provided. Set it explicitly to control the output location.

Development

git clone https://github.com/XJTLUmedia/Context-First-MCP.git
cd Context-First-MCP
pnpm install

# Build this package only
pnpm --filter @xjtlumedia/context-first-mcp-server build

# Run tests
pnpm --filter @xjtlumedia/context-first-mcp-server test

# Type-check only
pnpm --filter @xjtlumedia/context-first-mcp-server exec tsc --noEmit

Related Packages

| Package | Description | |---------|-------------| | context-first-mcp | npx-ready stdio server — the easiest way to run Context-First | | @xjtlumedia/context-first-remote-server | Vercel Streamable HTTP transport |

Full documentation, Claude Desktop / Cursor / VS Code setup, and the tool reference are in the main README.

License

MIT — © XJTLUmedia