Ask it anything.
Every model thinks. One mind answers.

Master Mind is now exposed as a Model Context Protocol (MCP) compliant tool server, enabling any MCP-compatible AI agent to discover and invoke reviews programmatically via JSON-RPC 2.0.

• Endpoint: POST /api/master-mind/mcp (JSON-RPC 2.0) • GET /api/master-mind/mcp (tool catalog)
• Protocol: MCP 2024-11-05 — supports initialize, tools/list, tools/call
• 5 tools: review_code, check_review_status, keyword_search, create_update_set, list_recent_reviews
• Auth: x-internal-api-key header

A Master Mind Review API is deployed on all 4 ServiceNow instances, bridging directly to Cloud Run via sn_ws.RESTMessageV2.

• POST /review — Submit code for multi-AI review (returns session ID)
• GET /status/{session_id} — Poll review results
• POST /keyword-search — Search instance metadata for related artifacts
• Build Agent skill: Registered on dev392913 so Build Agent can delegate code reviews to Master Mind

Before applying consensus fixes to ServiceNow, Master Mind snapshots the current state of all referenced artifacts via the Table API.

• Auto-snapshot: Detects sys_ids, Class.create names, and name= patterns in your code
• POST /sessions/:id/restore-checkpoint — Roll back to pre-review state
• GET /sessions/:id/checkpoint-diff — Compare snapshot vs current instance state
• Controlled by: SN_CHECKPOINT_ENABLED env var (default: enabled)

After consensus generation for ServiceNow code, Master Mind automatically creates a named update set on the target instance via the Build Agent Update Sets API.

• Naming: MM-<YYYYMMDD>-<descriptor>-<shortId>
• POST /sessions/:id/complete-update-set — Finalize the update set
• Controlled by: SN_AUTO_UPDATE_SET env var (default: enabled)

When Master Mind detects ServiceNow code (weighted pattern scoring, 12 patterns, threshold=3), it auto-queries the Build Agent Keyword Search API on your instance to discover related artifacts.

• 540K+ indexed records: Business rules, script includes, client scripts, and more
• Auto-injected: Discovery results are added to the system prompt for all 6 AI models
• Non-blocking: 15s timeout, fails silently, never blocks execution
• POST /api/master-mind/keyword-search — Standalone test endpoint

All Build Agent operations are tracked in MongoDB via the buildAgentMetrics field on each session.

• Tracked: keywordSearchMs, keywordsExtracted, searchQueriesRun, artifactsFound, updateSetCreated, instanceUrl

Master Mind now automatically detects what your prompt is about and loads the right tools. Ask about stocks — market data tools activate. Ask about ServiceNow — SN tools activate. Ask both — all tools load together.

• 9 tool categories: ServiceNow (10), Market Data (6), News (3), Web (1), Compute (1), Crypto (2), Tech Analysis (6), SEC/Filings (3), Econ Calendar (3) — 35 tools total
• Regex-based routing: Instant, free context detection — no extra LLM call needed
• Cross-domain queries: "What is NVDA at and any BRs referencing NVDA?" loads both market and SN tools simultaneously
• Per-category metrics: The Sources Queried banner shows breakdowns by category with cache hits and timing

Real-time stock market data via Finnhub API — all models get live prices instead of guessing from training data.

• market_get_quote — Live price, change %, high/low, company info (with ticker validation)
• market_get_indices — SPY, QQQ, DIA, IWM, VXX snapshot
• market_get_bars — OHLCV candles for technical analysis (1min to weekly)
• market_clock — Market open/closed status
• market_get_positions / market_get_account — Portfolio data (requires Alpaca config)

• news_company — Recent headlines for a ticker (earnings, upgrades, catalysts)
• news_market_feed — Broad market news (general, forex, crypto, merger)
• news_sentiment — Quantified bullish/bearish sentiment score

• calculate — Position sizing, risk/reward, P&L, percentages via safe math evaluator (never eval())
• web_fetch_url — Extract text from URLs with sanitization (4KB cap, script stripping, prompt-injection hardening)

When connected via the SN Record Context panel, all 6 AI models (Opus, GPT-5.6 Sol, Sonnet, Fable 5, DeepSeek V4 Pro, GLM-5.3) make real-time READ-ONLY queries to your ServiceNow instance before answering. No more guessing from training data.

• Forced first call: Every model is required to query ServiceNow at least once before responding — ensures answers are grounded in live instance data
• 10 available tools: sn_query_record, sn_query_schema, sn_query_table_artifacts, sn_search_by_name, sn_search_artifact, sn_fetch_script_by_name, sn_fetch_script_by_sysid, sn_search_script_body, sn_query_portal, sn_query_catalog
• Per-model metrics: The Sources Queried banner shows total tool calls, SN API requests, data transferred, query time, and cache hits — with a per-model breakdown
• Cross-model caching: When multiple models query the same endpoint, the first call hits the API and subsequent models get cached results — reducing SN load and latency
• Up to 10 rounds: Each model can make up to 10 tool call rounds to build a complete picture before responding

The sn_search_script_body tool searches INSIDE the source code of ServiceNow artifacts — not just by name. AI models use multi-keyword strategies to find relevant scripts:

• Method names: getManager, lookupManager
• GlideRecord references: GlideRecord("sys_user")
• Field accessors: getValue("manager"), current.manager
• Multi-type fan-out: Search across up to 5 artifact types in a single call (e.g., script_include + business_rule + client_script)
• 15 artifact types supported: Script Includes, Business Rules, Client Scripts, Fix Scripts, UI Scripts, Scheduled Jobs, UI Policies, UI Actions, UI Pages, UI Macros, Catalog Client Scripts, Catalog UI Policies, ACLs, Data Policies, Script Actions

When no artifact type filter is selected, the system auto-detects relevant types from your prompt using 15 regex-based detection rules. For example, mentioning "onChange" or "g_form" auto-prioritizes Client Scripts.

• Default: Dropdown defaults to "Auto-Detect from Context" — no manual selection needed
• Manual override: Select a specific type from the SN Artifact Type dropdown to narrow searches
• Smart prioritization: Up to 5 detected types are injected into the system prompt so models search the most relevant artifacts first

A data-driven system that learns from your UAT history to maximize one-shot success probability. Treats your UAT dataset as a two-engine learning system:

• Success Engine: Accepted + uatTurns=0 — exemplar library of what "good" looks like
• Improvement Engine: Accepted + uatTurns>0 — gap library of what was initially missing

Smart Ranking: Accepted UATs are ranked by oneShotScore = 1/(1+turns) — one-shot successes surface first. Top 5 high-signal examples returned instead of 20.

Auto-Tagging: When a session is marked Accepted, topic tags are auto-derived from the content using 20 ServiceNow-aware detection rules (business_rule, client_script, dictionary, glide_api, etc.).

Tier Badges in History:

• 🥇 Gold — One-shot success (0 turns)
• 🥈 Silver — Near one-shot (1–2 turns)
• 🥉 Bronze — Iterated (3+ turns)

Analytics: The cross-ref checkbox shows your one-shot rate percentage alongside the accepted count.

Opt-in toggle that cross-references your new prompt against top-ranked Accepted UAT sessions. AI models receive tier-weighted context — GOLD records (one-shots) are weighted highest, with topic tags and scores for smarter pattern matching.

• Toggle checkbox above the Execute button to enable
• Fetches top 5 Accepted sessions ranked by one-shot score (not recency)
• Each record includes tier label (GOLD/SILVER/BRONZE), score, and topic tags
• AI models are instructed to weight GOLD records most heavily
• Response includes <insights_from_uat> section showing which records were relevant
• Works with both regular execution and Follow Up

Hover over any code block in AI Consensus or model responses to reveal a toolbar with Copy and Download buttons. Download uses language-aware file extensions (.js, .py, .xml, etc.).

Continue the conversation after reviewing an AI Consensus response. Preserves full context across interactions so all AI models understand the prior analysis.

• Initiate anytime: Click the green "Follow Up" button in Results — works on the current session or any session loaded from History
• Context toggle: Checkbox to include or exclude the prior AI Consensus in your follow-up. When included, all AI models receive both the original question and the full consensus text
• Dedicated tab: The Follow Up tab shows a read-only preview of the attached consensus, your follow-up prompt area, and mode selector
• Multi-level: Each follow-up creates a new session that can itself be followed up on — enabling iterative refinement chains
• Threading: Follow-up sessions are linked to their parent via parentSessionId. In History, follow-up sessions display a reply icon
• Mode flexibility: Choose a different execution mode for the follow-up (Single, Quick, or Full) independent of the original session
• Attachments: Attach Local File Paths and drag-drop uploads in the Follow Up tab, same as the main Input tab
• Progress visualization: Follow-up execution shows the same Ready tab with status polling as regular execution

The browser tab title prefixes [LOCAL] on localhost for easy tab identification.

Reference files from your local filesystem by pasting absolute paths. Files are read server-side at execution time and included as attachments.

• Available on localhost only — disabled on production with a clear banner
• Supports images, documents, code files (same types as uploads)
• Up to 10 paths, 25MB per file
• Layered security: path validation, extension allowlist, symlink resolution, blocked directory/file checks, localhost-only enforcement
• Parallel file resolution with partial-success handling
• Privacy: only basenames stored in context/history, never full paths

Paste a folder path and click Add Folder to scan all supported files in that directory. All matching files are added to the file path list automatically.

• Scans the directory for files matching supported extensions (code, docs, images)
• Sensitive files (.env, keys, credentials) are automatically excluded
• Respects the 10-file limit — skips duplicates and stops at the cap
• Available in both Input tab and Follow-Up tab

Single: One selected AI model. Quick (2 Models): Parallel with consensus. Full (6 Models): All models — Claude Opus 5, GPT-5.6 Sol, Claude Sonnet 5, Claude Fable 5, DeepSeek V4 Pro (0813), GLM-5.3. Required before production deployment.

Primary text area for your prompt. Supports up to 100,000 characters. All context is sent to the selected AI models along with any attached files or screenshots.

Capture screenshots during a recording session. Select any screen, window, or application (Zoom, Teams, browser) to capture. Up to 5 screenshots per session, sent to vision-capable models.

Live audio recording with automatic transcription. Speak your prompt and it gets transcribed into the context area. Supports capturing screen while recording.

Drag-and-drop or click to upload files. Supports images (PNG, JPG, GIF, WebP), documents (PDF, TXT, MD, DOCX), code files (JS, PY, XML, JSON, etc.), and more. Up to 5 files, 25MB each.

Pull records from your ServiceNow instance to include as context. Provides real instance data (tables, scripts, schemas) directly to the AI models.

Attach code dependency context so AI models understand your project's tech stack, imports, and architecture when generating solutions.

Define find-and-replace pairs that automatically correct your transcription before submission. Useful for fixing recurring speech-to-text errors (e.g., "G form" → "g_form").

When multiple models respond (Quick or Full mode), Master Mind synthesizes all responses into a single AI Consensus. The consensus identifies agreements, conflicts, and unique insights across models. Includes source verification audit.

Download any individual model response or the consensus as PDF, Markdown (.md), or Word (.docx). Copy to clipboard also available.

All sessions are persisted and searchable. Click any session to reload its context, responses, and consensus. Supports search by title, model, mode, UAT status, and tags.

Add custom tags to any session for organization and filtering. Click tags in History to filter by them. Tags are per-user and searchable. Up to 20 tags per session.

Track the acceptance status of each session (Pending, Ready, In Progress, Accepted, Rejected) and the number of follow-up turns needed. Filter history by UAT status.

One AI model of your choice from the Execution Mode dropdown. Claude Fable 5 bills $10.00 (its API rates are 2× Opus); every other model is $5.00. The badge above tracks your current selection.

Two models respond in parallel — Claude Opus 5 and GPT-5.6 Sol — plus a synthesized AI Consensus.

All six models — Claude Opus 5, GPT-5.6 Sol, Claude Sonnet 5, Claude Fable 5, DeepSeek V4 Pro (0813), and GLM-5.3 — plus the AI Consensus. Required before production deployment.