/** * AGENTSCOPE-SIFT — MCP Server * * "SIFT gives investigators tools. AGENTSCOPE gives the AI an evidence trail." * * Every tool call passes through: * BOB Plasma Filter → SENTINEL → SealForge WORM chain * * The agent physically cannot: * - Run destructive commands (no write tools exposed) * - Skip evidence validation (BOB gates every result) * - Produce an unproven finding (every finding sealed to a tool call) */ import { Server } from '@modelcontextprotocol/sdk/server/index.js' import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js' import { CallToolRequestSchema, ListToolsRequestSchema } from '@modelcontextprotocol/sdk/types.js' import { z } from 'zod' import { bob } from './pipeline/bob.js' import { sentinel } from './pipeline/sentinel.js' import { sealEntry, initChain } from './pipeline/seal.js' import * as SIFT from './tools/sift.js' initChain() const server = new Server( { name: 'agentscope-sift', version: '1.0.0' }, { capabilities: { tools: {} } } ) // ── Tool registry ───────────────────────────────────────────────────────────── const TOOLS = [ { name: 'list_processes', description: 'List running processes from a memory image (volatility3 windows.pslist)', inputSchema: { type: 'object', properties: { memory_image: { type: 'string', description: 'Path to memory dump (.raw, .mem, .dmp)' } }, required: ['memory_image'] }, }, { name: 'scan_network_connections', description: 'Extract network connections from memory (volatility3 windows.netstat)', inputSchema: { type: 'object', properties: { memory_image: { type: 'string' } }, required: ['memory_image'] }, }, { name: 'dump_cmdline', description: 'Extract command-line arguments of all processes from memory', inputSchema: { type: 'object', properties: { memory_image: { type: 'string' } }, required: ['memory_image'] }, }, { name: 'find_injected_code', description: 'Detect code injection and suspicious memory regions (volatility3 malfind)', inputSchema: { type: 'object', properties: { memory_image: { type: 'string' } }, required: ['memory_image'] }, }, { name: 'get_registry_hives', description: 'List Windows registry hives found in memory', inputSchema: { type: 'object', properties: { memory_image: { type: 'string' } }, required: ['memory_image'] }, }, { name: 'list_disk_partitions', description: 'List partitions in a disk image (mmls)', inputSchema: { type: 'object', properties: { image_path: { type: 'string' } }, required: ['image_path'] }, }, { name: 'list_files', description: 'List files on a disk image filesystem including deleted files (fls)', inputSchema: { type: 'object', properties: { image_path: { type: 'string' }, offset: { type: 'number' } }, required: ['image_path'] }, }, { name: 'get_filesystem_info', description: 'Get filesystem metadata from disk image (fsstat)', inputSchema: { type: 'object', properties: { image_path: { type: 'string' }, offset: { type: 'number' } }, required: ['image_path'] }, }, { name: 'scan_with_yara', description: 'Scan a file or directory with YARA malware rules', inputSchema: { type: 'object', properties: { rules_file: { type: 'string' }, target_path: { type: 'string' } }, required: ['rules_file', 'target_path'] }, }, { name: 'extract_strings', description: 'Extract printable strings from a binary — finds IOCs, URLs, registry keys', inputSchema: { type: 'object', properties: { file_path: { type: 'string' }, min_length: { type: 'number' } }, required: ['file_path'] }, }, { name: 'hash_file', description: 'Compute SHA-256 and MD5 hashes of a file for evidence integrity', inputSchema: { type: 'object', properties: { file_path: { type: 'string' } }, required: ['file_path'] }, }, { name: 'get_file_metadata', description: 'Extract file type and metadata (exiftool)', inputSchema: { type: 'object', properties: { file_path: { type: 'string' } }, required: ['file_path'] }, }, { name: 'analyze_pcap', description: 'Analyze a network capture — DNS queries, HTTP requests (tshark)', inputSchema: { type: 'object', properties: { pcap_file: { type: 'string' } }, required: ['pcap_file'] }, }, { name: 'verify_chain', description: 'Verify the WORM evidence chain has not been tampered with', inputSchema: { type: 'object', properties: {}, required: [] }, }, ] server.setRequestHandler(ListToolsRequestSchema, async () => ({ tools: TOOLS })) // ── Tool dispatcher ─────────────────────────────────────────────────────────── server.setRequestHandler(CallToolRequestSchema, async (req) => { const { name, arguments: args } = req.params const a = (args ?? {}) as Record // Seal the call intent before executing sealEntry('TOOL_CALL', { tool: name, args: a }, name) let rawResult: object try { rawResult = await dispatch(name, a) } catch (err: any) { sealEntry('BLOCK', { tool: name, reason: err.message }, name) return { content: [{ type: 'text', text: `[BLOCKED] ${err.message}` }], isError: true } } const rawStr = JSON.stringify(rawResult) // BOB Plasma Filter — catches adversarial tool output const bobResult = bob(rawStr) if (!bobResult.pass) { sealEntry('BLOCK', { tool: name, blocked_at: bobResult.blocked_at, violation: bobResult.plasma.violation ?? bobResult.coherence.message, }, name) return { content: [{ type: 'text', text: `[BOB BLOCKED at ${bobResult.blocked_at}] Tool output failed integrity check. Finding discarded.` }], isError: true, } } // SENTINEL zero-trust gate const sentinelResult = sentinel(name, rawStr) if (sentinelResult.verdict !== 'APPROVED') { sealEntry('BLOCK', { tool: name, verdict: sentinelResult.verdict, finding: sentinelResult.finding, }, name) return { content: [{ type: 'text', text: `[SENTINEL ${sentinelResult.verdict}] ${sentinelResult.finding ?? 'trust gate failed'}` }], isError: true, } } // Seal the approved result const entry = sealEntry('FINDING', { tool: name, trust_score: sentinelResult.trust_score, result_preview: rawStr.slice(0, 200), }, name) const output = { ...rawResult, _agentscope: { seal: entry.seal, seq: entry.seq, trust_score: sentinelResult.trust_score, verdict: sentinelResult.verdict, entropy: bobResult.coherence.entropy, } } return { content: [{ type: 'text', text: JSON.stringify(output, null, 2) }] } }) async function dispatch(name: string, a: Record): Promise { switch (name) { case 'list_processes': return SIFT.list_processes(String(a.memory_image)) case 'scan_network_connections': return SIFT.scan_network_connections(String(a.memory_image)) case 'dump_cmdline': return SIFT.dump_cmdline(String(a.memory_image)) case 'find_injected_code': return SIFT.find_injected_code(String(a.memory_image)) case 'get_registry_hives': return SIFT.get_registry_hives(String(a.memory_image)) case 'list_disk_partitions': return SIFT.list_disk_partitions(String(a.image_path)) case 'list_files': return SIFT.list_files(String(a.image_path), a.offset as number | undefined) case 'get_filesystem_info': return SIFT.get_filesystem_info(String(a.image_path), a.offset as number | undefined) case 'scan_with_yara': return SIFT.scan_with_yara(String(a.rules_file), String(a.target_path)) case 'extract_strings': return SIFT.extract_strings(String(a.file_path), a.min_length as number | undefined) case 'hash_file': return SIFT.hash_file(String(a.file_path)) case 'get_file_metadata': return SIFT.get_file_metadata(String(a.file_path)) case 'analyze_pcap': return SIFT.analyze_pcap(String(a.pcap_file)) case 'verify_chain': { const { verifyChain } = await import('./pipeline/seal.js') return verifyChain() } default: throw new Error(`Unknown tool: ${name}`) } } // ── Start ───────────────────────────────────────────────────────────────────── const transport = new StdioServerTransport() await server.connect(transport) console.error('[AGENTSCOPE-SIFT] MCP server running — BOB + SENTINEL + SealForge active')