@tensorfeed/mcp-server 1.11.1 → 1.13.0

package/CHANGELOG.md

@@ -1,6 +1,25 @@
 # Changelog
 
-All notable changes to the [TensorFeed.ai MCP server](https://github.com/RipperMercs/tensorfeed/tree/main/mcp-server). Free tools work without configuration; premium tools require a bearer token via the `TENSORFEED_TOKEN` env var. Buy credits at [tensorfeed.ai/developers/agent-payments](https://tensorfeed.ai/developers/agent-payments).
+All notable changes to the [TensorFeed.ai MCP server](https://github.com/RipperMercs/tensorfeed-mcp). Free tools work without configuration; premium tools require a bearer token via the `TENSORFEED_TOKEN` env var. Buy credits at [tensorfeed.ai/developers/agent-payments](https://tensorfeed.ai/developers/agent-payments).
+
+## 1.13.0 - 2026-05-04
+
+### Added
+- `get_ai_papers_trending` tool. Daily curated AI/ML research papers ranked by citation count, sourced from Semantic Scholar. Five fan-out queries (LLM, transformer, RLHF, AI agents, diffusion), deduped, top 30. Free.
+- `get_arxiv_recent` tool. Most recent arXiv submissions in cs.AI / cs.LG / cs.CL / cs.CV, top 50 by submission date. Optional `limit` arg for the rendered list. Free.
+- `get_hf_trending` tool. Top 30 most-downloaded models and datasets on Hugging Face. Optional `section` (models / datasets / both) and `limit` args. Free.
+
+All three are backed by the worker endpoints shipped in the same monorepo commit and pass through the 1.12.0 output sanitizer automatically.
+
+## 1.12.0 - 2026-05-04
+
+### Added
+- Defense-in-depth output sanitizer (`src/sanitize.ts`). Every tool's text output passes through a scrub pass before reaching the host LLM (Claude Desktop, Cursor, Cline, Roo, etc): strips ASCII control chars (preserves newline, carriage return, tab), strips bidi control + zero-width spoofing chars, and neutralizes role-confusion markers (ChatML / Llama / Mistral chat tokens, "ignore previous instructions", line-leading "system:" / "developer:" / "assistant:" preambles). Wired via a new `registerTool()` helper that wraps `server.tool()`. The TensorFeed worker already runs the same scrub on its agent-facing endpoints; this layer protects against any new endpoint that might forget it, plus drift between worker rules and what the MCP server should consider safe. No behavior change for legitimate text. No new dependencies.
+
+## 1.11.1 - 2026-05-04
+
+### Changed
+- Repository URL updated to point at the dedicated [`RipperMercs/tensorfeed-mcp`](https://github.com/RipperMercs/tensorfeed-mcp) repo. The MCP server now has its own GitHub home for issues, stars, and discoverability. Source-of-truth for builds remains in the main `tensorfeed` monorepo's `mcp-server/` folder; the standalone repo is mirrored on release. No code changes; no behavior changes; npm install / npx unaffected.
 
 ## 1.11.0 - 2026-04-29
 

package/README.md

@@ -17,13 +17,31 @@ cd mcp-server
 npm run build
 npm publish --access public
 
-# 3. Republish to the official MCP registry
+# 3. Republish to the official MCP registry. The script lives at
+#    repo-root/scripts/, not mcp-server/scripts/, so step back up first.
+cd ..
 .\scripts\mcp-publish.ps1
 
-# 4. Mirror the new version to the standalone repo
-#    (manual until we wire it as a release-time GitHub Action)
+# 4. Mirror to the standalone repo - automated. The
+#    .github/workflows/mirror-mcp-server.yml workflow runs on every
+#    push to main that touches mcp-server/. To trigger a manual sync,
+#    go to the Actions tab and run "Mirror MCP server to standalone
+#    repo" via workflow_dispatch.
 ```
 
+### Automated mirror setup (one-time)
+
+The mirror workflow needs a personal access token with `contents: write`
+permission on `RipperMercs/tensorfeed-mcp`. Set it once:
+
+1. Generate a fine-grained PAT at github.com/settings/personal-access-tokens
+   scoped to the standalone repo only, with Repository permissions
+   `Contents: Read and write` and `Metadata: Read-only`.
+2. Add it to this monorepo as a repository secret named
+   `STANDALONE_REPO_TOKEN` (Settings -> Secrets and variables -> Actions).
+3. The workflow will pick it up on the next push to main that changes
+   `mcp-server/**`, or via the manual "Run workflow" button.
+
 ## Quick links
 
 - **User-facing repo (please star):** https://github.com/RipperMercs/tensorfeed-mcp

package/dist/index.js

@@ -2,6 +2,7 @@
 import { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js';
 import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js';
 import { z } from 'zod';
+import { sanitizeToolResponse } from './sanitize.js';
 const API_BASE = 'https://tensorfeed.ai/api';
 const SDK_VERSION = '1.9.0';
 async function fetchJSON(path, opts = {}) {
@@ -46,8 +47,26 @@ const server = new McpServer({
     name: 'tensorfeed',
     version: SDK_VERSION,
 });
+// ── Defense-in-depth tool wrapper ───────────────────────────────────
+// Every tool's text output passes through sanitizeToolResponse before
+// reaching the host LLM. Strips control chars, bidi/zero-width spoofing,
+// and neutralizes role-confusion tokens (ChatML / Llama / Mistral chat
+// markers, "ignore previous instructions", etc). The TensorFeed worker
+// already runs the same scrub on its agent-facing endpoints; this layer
+// covers any new endpoint that might forget it, plus protects against
+// drift between worker rules and what the MCP server should consider
+// safe. See ./sanitize.ts.
+function registerTool(name, description, paramsSchema, cb) {
+    const wrapped = (async (args, extra) => {
+        const result = await cb(args, extra);
+        return sanitizeToolResponse(result);
+    });
+    // Direct dispatch to the SDK's underlying registration. Do not call
+    // registerTool() here or it will recurse forever.
+    return server.tool(name, description, paramsSchema, wrapped);
+}
 // ── Tool: get_ai_news ───────────────────────────────────────────────
-server.tool('get_ai_news', 'Get the latest AI news articles from TensorFeed.ai. Aggregates from 15+ sources including Anthropic, OpenAI, Google, TechCrunch, The Verge, arXiv, and more.', {
+registerTool('get_ai_news', 'Get the latest AI news articles from TensorFeed.ai. Aggregates from 15+ sources including Anthropic, OpenAI, Google, TechCrunch, The Verge, arXiv, and more.', {
     category: z.string().optional().describe('Filter by category (e.g. "anthropic", "openai", "research", "tools")'),
     limit: z.number().min(1).max(50).optional().describe('Number of articles to return (default 10, max 50)'),
 }, async ({ category, limit }) => {
@@ -62,7 +81,7 @@ server.tool('get_ai_news', 'Get the latest AI news articles from TensorFeed.ai.
     return { content: [{ type: 'text', text: text || 'No articles found.' }] };
 });
 // ── Tool: get_ai_status ─────────────────────────────────────────────
-server.tool('get_ai_status', 'Get real-time operational status of major AI services (Claude, OpenAI, Gemini, Mistral, Cohere, Replicate, Hugging Face).', {}, async () => {
+registerTool('get_ai_status', 'Get real-time operational status of major AI services (Claude, OpenAI, Gemini, Mistral, Cohere, Replicate, Hugging Face).', {}, async () => {
     const data = await fetchJSON('/status');
     const text = data.services
         .map(s => {
@@ -75,7 +94,7 @@ server.tool('get_ai_status', 'Get real-time operational status of major AI servi
     return { content: [{ type: 'text', text: `AI Service Status:\n${text}` }] };
 });
 // ── Tool: is_service_down ───────────────────────────────────────────
-server.tool('is_service_down', 'Check if a specific AI service is currently down or experiencing issues.', {
+registerTool('is_service_down', 'Check if a specific AI service is currently down or experiencing issues.', {
     service: z.string().describe('Service name to check (e.g. "claude", "openai", "gemini", "mistral", "cohere", "hugging face", "replicate")'),
 }, async ({ service }) => {
     const data = await fetchJSON('/status');
@@ -96,7 +115,7 @@ server.tool('is_service_down', 'Check if a specific AI service is currently down
     };
 });
 // ── Tool: get_model_pricing ─────────────────────────────────────────
-server.tool('get_model_pricing', 'Get AI model pricing comparison across all major providers (Anthropic, OpenAI, Google, Meta, Mistral, Cohere). Prices per 1M tokens.', {}, async () => {
+registerTool('get_model_pricing', 'Get AI model pricing comparison across all major providers (Anthropic, OpenAI, Google, Meta, Mistral, Cohere). Prices per 1M tokens.', {}, async () => {
     const data = await fetchJSON('/models');
     const text = data.providers
         .map(p => {
@@ -116,7 +135,7 @@ server.tool('get_model_pricing', 'Get AI model pricing comparison across all maj
     return { content: [{ type: 'text', text: `AI Model Pricing (per 1M tokens):\n\n${text}` }] };
 });
 // ── Tool: get_ai_today ──────────────────────────────────────────────
-server.tool('get_ai_today', 'Get a summary of what happened in AI today. Returns the top stories from the last 24 hours.', {
+registerTool('get_ai_today', 'Get a summary of what happened in AI today. Returns the top stories from the last 24 hours.', {
     limit: z.number().min(1).max(20).optional().describe('Number of stories (default 5)'),
 }, async ({ limit }) => {
     const data = await fetchJSON(`/news?limit=${limit || 5}`);
@@ -126,7 +145,7 @@ server.tool('get_ai_today', 'Get a summary of what happened in AI today. Returns
     return { content: [{ type: 'text', text: `Today in AI:\n\n${text}` }] };
 });
 // ── Tool: get_agent_activity ────────────────────────────────────────
-server.tool('get_agent_activity', 'Get live AI bot traffic on TensorFeed.ai. Returns today\'s bot hit count, the most recent 50 hits with bot name + endpoint + timestamp, and a derived top-bots breakdown. Useful when an agent wants to see which crawlers are pulling AI ecosystem data and how that distribution shifts. Free, no auth.', {}, async () => {
+registerTool('get_agent_activity', 'Get live AI bot traffic on TensorFeed.ai. Returns today\'s bot hit count, the most recent 50 hits with bot name + endpoint + timestamp, and a derived top-bots breakdown. Useful when an agent wants to see which crawlers are pulling AI ecosystem data and how that distribution shifts. Free, no auth.', {}, async () => {
     const data = (await fetchJSON('/agents/activity'));
     const tally = new Map();
     for (const hit of data.recent) {
@@ -158,7 +177,7 @@ server.tool('get_agent_activity', 'Get live AI bot traffic on TensorFeed.ai. Ret
 // PREMIUM TOOLS (require TENSORFEED_TOKEN env var, paid in USDC on Base)
 // ════════════════════════════════════════════════════════════════════
 // ── Tool: get_account_balance ───────────────────────────────────────
-server.tool('get_account_balance', 'Check the credit balance for the configured TensorFeed bearer token. Free, but requires TENSORFEED_TOKEN to be set.', {}, async () => {
+registerTool('get_account_balance', 'Check the credit balance for the configured TensorFeed bearer token. Free, but requires TENSORFEED_TOKEN to be set.', {}, async () => {
     const data = (await fetchJSON('/payment/balance', { auth: true }));
     return {
         content: [
@@ -170,7 +189,7 @@ server.tool('get_account_balance', 'Check the credit balance for the configured
     };
 });
 // ── Tool: get_account_usage ─────────────────────────────────────────
-server.tool('get_account_usage', 'Show per-endpoint usage for the configured TensorFeed token (last 100 calls aggregated). Free, but requires TENSORFEED_TOKEN.', {}, async () => {
+registerTool('get_account_usage', 'Show per-endpoint usage for the configured TensorFeed token (last 100 calls aggregated). Free, but requires TENSORFEED_TOKEN.', {}, async () => {
     const data = (await fetchJSON('/payment/usage', { auth: true }));
     if (data.total_calls === 0) {
         return { content: [{ type: 'text', text: 'No premium API calls on this token yet.' }] };
@@ -189,7 +208,7 @@ server.tool('get_account_usage', 'Show per-endpoint usage for the configured Ten
     };
 });
 // ── Tool: premium_routing (1 credit) ────────────────────────────────
-server.tool('premium_routing', 'Get a ranked list of recommended AI models for a task with full score breakdown (quality, availability, cost, latency). Costs 1 credit.', {
+registerTool('premium_routing', 'Get a ranked list of recommended AI models for a task with full score breakdown (quality, availability, cost, latency). Costs 1 credit.', {
     task: z.enum(['code', 'reasoning', 'creative', 'general']).optional().describe('Task type the model needs to be good at (default: general)'),
     budget: z.number().optional().describe('Max blended USD per 1M tokens'),
     min_quality: z.number().min(0).max(1).optional().describe('Minimum quality score in [0, 1]'),
@@ -217,7 +236,7 @@ server.tool('premium_routing', 'Get a ranked list of recommended AI models for a
     return { content: [{ type: 'text', text: `Routing for "${data.task}":\n\n${list}${billing}` }] };
 });
 // ── Tool: pricing_series (1 credit) ─────────────────────────────────
-server.tool('pricing_series', 'Daily price points for one AI model with min/max/delta summary. Default range = last 30 days, max 90 days. Costs 1 credit.', {
+registerTool('pricing_series', 'Daily price points for one AI model with min/max/delta summary. Default range = last 30 days, max 90 days. Costs 1 credit.', {
     model: z.string().describe('Model id or display name (e.g. "Claude Opus 4.7" or "claude-opus-4-7")'),
     from: z.string().optional().describe('Start date YYYY-MM-DD UTC (default: 30 days ago)'),
     to: z.string().optional().describe('End date YYYY-MM-DD UTC (default: today)'),
@@ -242,7 +261,7 @@ server.tool('pricing_series', 'Daily price points for one AI model with min/max/
     };
 });
 // ── Tool: benchmark_series (1 credit) ───────────────────────────────
-server.tool('benchmark_series', 'Score evolution for a single benchmark on one AI model. Costs 1 credit. Benchmark keys: swe_bench, mmlu_pro, gpqa_diamond, math, human_eval.', {
+registerTool('benchmark_series', 'Score evolution for a single benchmark on one AI model. Costs 1 credit. Benchmark keys: swe_bench, mmlu_pro, gpqa_diamond, math, human_eval.', {
     model: z.string().describe('Model id or display name'),
     benchmark: z.string().describe('Benchmark key (e.g. swe_bench, mmlu_pro, gpqa_diamond, math, human_eval)'),
     from: z.string().optional().describe('Start date YYYY-MM-DD UTC (default: 30 days ago)'),
@@ -268,7 +287,7 @@ server.tool('benchmark_series', 'Score evolution for a single benchmark on one A
     };
 });
 // ── Tool: status_uptime (1 credit) ──────────────────────────────────
-server.tool('status_uptime', 'Daily uptime rollup for one provider with operational/degraded/down day counts and uptime % (degraded counts as half-credit). Costs 1 credit.', {
+registerTool('status_uptime', 'Daily uptime rollup for one provider with operational/degraded/down day counts and uptime % (degraded counts as half-credit). Costs 1 credit.', {
     provider: z.string().describe('Provider name (e.g. anthropic, openai, google)'),
     from: z.string().optional().describe('Start date YYYY-MM-DD UTC (default: 30 days ago)'),
     to: z.string().optional().describe('End date YYYY-MM-DD UTC (default: today)'),
@@ -295,7 +314,7 @@ server.tool('status_uptime', 'Daily uptime rollup for one provider with operatio
     };
 });
 // ── Tool: premium_agents_directory (1 credit) ───────────────────────
-server.tool('premium_agents_directory', 'Enriched AI agents catalog joined with live status, recent news, agent traffic, flagship pricing, and a 0-100 trending_score. Costs 1 credit.', {
+registerTool('premium_agents_directory', 'Enriched AI agents catalog joined with live status, recent news, agent traffic, flagship pricing, and a 0-100 trending_score. Costs 1 credit.', {
     category: z.string().optional().describe('coding, research, general, creative, frameworks'),
     status: z.enum(['operational', 'degraded', 'down', 'unknown']).optional(),
     open_source: z.boolean().optional(),
@@ -328,7 +347,7 @@ server.tool('premium_agents_directory', 'Enriched AI agents catalog joined with
     };
 });
 // ── Tool: whats_new (1 credit) ──────────────────────────────────────
-server.tool('whats_new', 'Agent morning brief: pricing changes, new/removed models, status incidents, and top news from the last 1-7 days. The tool to call when your agent boots up. Costs 1 credit.', {
+registerTool('whats_new', 'Agent morning brief: pricing changes, new/removed models, status incidents, and top news from the last 1-7 days. The tool to call when your agent boots up. Costs 1 credit.', {
     days: z.number().min(1).max(7).optional().describe('Window length in days (default 1)'),
     news_limit: z.number().min(1).max(25).optional().describe('Max news headlines (default 10)'),
 }, async ({ days, news_limit }) => {
@@ -364,7 +383,7 @@ server.tool('whats_new', 'Agent morning brief: pricing changes, new/removed mode
     };
 });
 // ── Tool: compare_models (1 credit) ─────────────────────────────────
-server.tool('compare_models', 'Side-by-side comparison of 2-5 AI models. Returns pricing, benchmarks (normalized to a union of keys with null for missing scores), provider status, and recent news per model, plus rankings (cheapest blended, most context, per-benchmark leaderboard). Costs 1 credit.', {
+registerTool('compare_models', 'Side-by-side comparison of 2-5 AI models. Returns pricing, benchmarks (normalized to a union of keys with null for missing scores), provider status, and recent news per model, plus rankings (cheapest blended, most context, per-benchmark leaderboard). Costs 1 credit.', {
     ids: z.string().describe('Comma-separated list of 2-5 model ids or display names. Examples: "Claude Opus 4.7,GPT-5.5,Gemini 3" or "opus-4-7,gpt-5-5"'),
 }, async ({ ids }) => {
     const data = (await fetchJSON(`/premium/compare/models?ids=${encodeURIComponent(ids)}`, { auth: true }));
@@ -391,7 +410,7 @@ server.tool('compare_models', 'Side-by-side comparison of 2-5 AI models. Returns
     };
 });
 // ── Tool: provider_deepdive (1 credit) ──────────────────────────────
-server.tool('provider_deepdive', 'Everything about an AI provider in one call: live status, all models with pricing + tier + benchmark scores joined in, recent news mentions, and agent traffic. Costs 1 credit. Aggregation IS the value; doing this from free endpoints would take 4 round-trips.', {
+registerTool('provider_deepdive', 'Everything about an AI provider in one call: live status, all models with pricing + tier + benchmark scores joined in, recent news mentions, and agent traffic. Costs 1 credit. Aggregation IS the value; doing this from free endpoints would take 4 round-trips.', {
     provider: z.string().describe('Provider id or display name (case-insensitive). Examples: anthropic, openai, google, mistral, cohere'),
 }, async ({ provider }) => {
     const data = (await fetchJSON(`/premium/providers/${encodeURIComponent(provider)}`, { auth: true }));
@@ -421,7 +440,7 @@ server.tool('provider_deepdive', 'Everything about an AI provider in one call: l
     };
 });
 // ── Tool: cost_projection (1 credit) ────────────────────────────────
-server.tool('cost_projection', 'Project the cost of a token-usage workload across 1-10 AI models. Returns daily/weekly/monthly/yearly totals per model and a ranking by cheapest monthly. Costs 1 credit.', {
+registerTool('cost_projection', 'Project the cost of a token-usage workload across 1-10 AI models. Returns daily/weekly/monthly/yearly totals per model and a ranking by cheapest monthly. Costs 1 credit.', {
     models: z.string().describe('One model or comma-separated list of up to 10 (e.g. "Claude Opus 4.7,GPT-5.5,Gemini 3"). Names or ids both work.'),
     input_tokens_per_day: z.number().min(0).describe('Expected daily input token volume'),
     output_tokens_per_day: z.number().min(0).describe('Expected daily output token volume'),
@@ -456,7 +475,7 @@ server.tool('cost_projection', 'Project the cost of a token-usage workload acros
     };
 });
 // ── Tool: news_search (1 credit) ────────────────────────────────────
-server.tool('news_search', 'Full-text search over the TensorFeed news article corpus with optional date range, provider, and category filters. Relevance scoring with recency boost. Costs 1 credit.', {
+registerTool('news_search', 'Full-text search over the TensorFeed news article corpus with optional date range, provider, and category filters. Relevance scoring with recency boost. Costs 1 credit.', {
     q: z.string().optional().describe('Free-text query, e.g. "claude opus pricing". Omit to browse latest filtered articles.'),
     from: z.string().optional().describe('Start date YYYY-MM-DD UTC (inclusive)'),
     to: z.string().optional().describe('End date YYYY-MM-DD UTC (inclusive end-of-day)'),
@@ -494,7 +513,7 @@ server.tool('news_search', 'Full-text search over the TensorFeed news article co
     };
 });
 // ── Tool: list_watches ──────────────────────────────────────────────
-server.tool('list_watches', 'List the active webhook watches owned by the configured TensorFeed token. Free, requires TENSORFEED_TOKEN.', {}, async () => {
+registerTool('list_watches', 'List the active webhook watches owned by the configured TensorFeed token. Free, requires TENSORFEED_TOKEN.', {}, async () => {
     const data = (await fetchJSON('/premium/watches', { auth: true }));
     if (data.count === 0) {
         return { content: [{ type: 'text', text: 'No active watches.' }] };
@@ -510,7 +529,7 @@ server.tool('list_watches', 'List the active webhook watches owned by the config
     return { content: [{ type: 'text', text: `${data.count} active watches:\n\n${rows}` }] };
 });
 // ── Tool: create_price_watch (1 credit) ─────────────────────────────
-server.tool('create_price_watch', 'Register a webhook watch on a model price change. Costs 1 credit. Watch lives 90 days. Each fire is an HMAC-signed POST to callback_url.', {
+registerTool('create_price_watch', 'Register a webhook watch on a model price change. Costs 1 credit. Watch lives 90 days. Each fire is an HMAC-signed POST to callback_url.', {
     model: z.string().describe('Model name (e.g. "Claude Opus 4.7")'),
     field: z.enum(['inputPrice', 'outputPrice', 'blended']).describe('Which price field to watch'),
     op: z.enum(['lt', 'gt', 'changes']).describe('Trigger: lt = below threshold, gt = above, changes = any change'),
@@ -535,7 +554,7 @@ server.tool('create_price_watch', 'Register a webhook watch on a model price cha
     };
 });
 // ── Tool: create_status_watch (1 credit) ────────────────────────────
-server.tool('create_status_watch', 'Register a webhook watch on a service status transition (e.g. anthropic becomes down). Costs 1 credit. Watch lives 90 days.', {
+registerTool('create_status_watch', 'Register a webhook watch on a service status transition (e.g. anthropic becomes down). Costs 1 credit. Watch lives 90 days.', {
     provider: z.string().describe('Provider name (e.g. anthropic, openai)'),
     op: z.enum(['becomes', 'changes']).describe('becomes = transitions to a specific value; changes = any transition'),
     value: z.enum(['operational', 'degraded', 'down']).optional().describe('Required when op is becomes'),
@@ -559,7 +578,7 @@ server.tool('create_status_watch', 'Register a webhook watch on a service status
     };
 });
 // ── Tool: create_digest_watch (1 credit) ────────────────────────────
-server.tool('create_digest_watch', 'Register a scheduled digest webhook that fires daily or weekly with a curated summary of pricing changes (regardless of whether anything dramatic happened). Costs 1 credit. Watch lives 90 days. Set-and-forget for agents that want a periodic snapshot without subscribing to realtime transitions.', {
+registerTool('create_digest_watch', 'Register a scheduled digest webhook that fires daily or weekly with a curated summary of pricing changes (regardless of whether anything dramatic happened). Costs 1 credit. Watch lives 90 days. Set-and-forget for agents that want a periodic snapshot without subscribing to realtime transitions.', {
     cadence: z.enum(['daily', 'weekly']).describe('How often the digest fires'),
     callback_url: z.string().describe('HTTPS URL to POST to when the digest fires'),
     secret: z.string().optional().describe('Optional HMAC shared secret'),
@@ -581,14 +600,14 @@ server.tool('create_digest_watch', 'Register a scheduled digest webhook that fir
     };
 });
 // ── Tool: delete_watch ──────────────────────────────────────────────
-server.tool('delete_watch', 'Delete one of your active webhook watches by id. Free, requires TENSORFEED_TOKEN.', {
+registerTool('delete_watch', 'Delete one of your active webhook watches by id. Free, requires TENSORFEED_TOKEN.', {
     watch_id: z.string().describe('The wat_... id from create_price_watch / create_status_watch / list_watches'),
 }, async ({ watch_id }) => {
     await fetchJSON(`/premium/watches/${watch_id}`, { method: 'DELETE', auth: true });
     return { content: [{ type: 'text', text: `Deleted watch ${watch_id}.` }] };
 });
 // ── Tool: probe_latest (free) ───────────────────────────────────────
-server.tool('probe_latest', 'Last 24 hours of measured LLM endpoint latency and availability per provider (Anthropic, OpenAI, Google, Mistral, Cohere). TensorFeed pings each provider\'s chat completion endpoint every 15 min and records time-to-first-byte, total response time, and HTTP status. Returns per-provider success rate and ttfb/total p50/p95/p99 latency. The data is unique because we measure it ourselves, not self-reported by the providers. Useful when an agent needs to pick a provider whose SLA you can verify, or to detect ongoing incidents before they hit a status page. Free, no auth.', {}, async () => {
+registerTool('probe_latest', 'Last 24 hours of measured LLM endpoint latency and availability per provider (Anthropic, OpenAI, Google, Mistral, Cohere). TensorFeed pings each provider\'s chat completion endpoint every 15 min and records time-to-first-byte, total response time, and HTTP status. Returns per-provider success rate and ttfb/total p50/p95/p99 latency. The data is unique because we measure it ourselves, not self-reported by the providers. Useful when an agent needs to pick a provider whose SLA you can verify, or to detect ongoing incidents before they hit a status page. Free, no auth.', {}, async () => {
     const data = (await fetchJSON('/probe/latest'));
     const s = data.summary;
     if (s.providers.length === 0) {
@@ -614,7 +633,7 @@ server.tool('probe_latest', 'Last 24 hours of measured LLM endpoint latency and
     };
 });
 // ── Tool: probe_series (1 credit) ───────────────────────────────────
-server.tool('probe_series', 'Daily SLA series for one LLM provider, measured by TensorFeed. Returns per-day count, success rate, ttfb p50/p95/p99, total p50/p95/p99, and incident-hour count across the requested window. Provider status pages are politically managed; this is the measured truth. Pairs naturally with premium_routing for picking a model whose SLA you can verify. Costs 1 credit.', {
+registerTool('probe_series', 'Daily SLA series for one LLM provider, measured by TensorFeed. Returns per-day count, success rate, ttfb p50/p95/p99, total p50/p95/p99, and incident-hour count across the requested window. Provider status pages are politically managed; this is the measured truth. Pairs naturally with premium_routing for picking a model whose SLA you can verify. Costs 1 credit.', {
     provider: z.enum(['anthropic', 'openai', 'google', 'mistral', 'cohere']).describe('LLM provider key'),
     from: z.string().optional().describe('Inclusive start YYYY-MM-DD (default: 30 days before to)'),
     to: z.string().optional().describe('Inclusive end YYYY-MM-DD (default: today UTC)'),
@@ -645,7 +664,7 @@ server.tool('probe_series', 'Daily SLA series for one LLM provider, measured by
     };
 });
 // ── Tool: mcp_registry_snapshot (free) ──────────────────────────────
-server.tool('mcp_registry_snapshot', 'Today\'s summary of the official Model Context Protocol server registry. Returns total servers, by-status breakdown, top namespaces, and 1-day deltas (newly added, reactivated, deprecated). Captured daily at 9:30 AM UTC from registry.modelcontextprotocol.io. Useful when an agent wants to see how the MCP ecosystem is growing or detect freshly-deprecated servers it may be using. Free, no auth.', {}, async () => {
+registerTool('mcp_registry_snapshot', 'Today\'s summary of the official Model Context Protocol server registry. Returns total servers, by-status breakdown, top namespaces, and 1-day deltas (newly added, reactivated, deprecated). Captured daily at 9:30 AM UTC from registry.modelcontextprotocol.io. Useful when an agent wants to see how the MCP ecosystem is growing or detect freshly-deprecated servers it may be using. Free, no auth.', {}, async () => {
     const data = (await fetchJSON('/mcp/registry/snapshot'));
     const s = data.summary;
     const statusLine = Object.entries(s.by_status)
@@ -676,7 +695,7 @@ server.tool('mcp_registry_snapshot', 'Today\'s summary of the official Model Con
     };
 });
 // ── Tool: mcp_registry_series (1 credit) ────────────────────────────
-server.tool('mcp_registry_series', 'Multi-day series of MCP server registry growth and churn. Returns per-day total servers, active count, and daily added/removed counts across the requested window. The registry itself is open data, but a 30/90-day trend requires daily capture started weeks ago, which TensorFeed has been running. Costs 1 credit.', {
+registerTool('mcp_registry_series', 'Multi-day series of MCP server registry growth and churn. Returns per-day total servers, active count, and daily added/removed counts across the requested window. The registry itself is open data, but a 30/90-day trend requires daily capture started weeks ago, which TensorFeed has been running. Costs 1 credit.', {
     from: z.string().optional().describe('Inclusive start date YYYY-MM-DD (default: 30 days before to)'),
     to: z.string().optional().describe('Inclusive end date YYYY-MM-DD (default: today UTC)'),
 }, async ({ from, to }) => {
@@ -705,6 +724,85 @@ server.tool('mcp_registry_series', 'Multi-day series of MCP server registry grow
         ],
     };
 });
+// ── Tool: get_ai_papers_trending (free) ─────────────────────────────
+registerTool('get_ai_papers_trending', 'Daily curated AI/ML research papers ranked by citation count. Sourced from Semantic Scholar across five fan-out queries (large language model, transformer, RLHF, AI agents, diffusion model), deduped by paperId, top 30 returned. Each paper carries title, abstract, authors, year, venue, citation count, arxivId, doi, and fields of study. Refreshed daily at 11:00 UTC. Use this when an agent wants the citation-ranked top of the AI research literature. Free, no auth.', {}, async () => {
+    const data = (await fetchJSON('/papers/ai-trending'));
+    const s = data.snapshot;
+    const lines = s.papers.map((p, i) => {
+        const authors = p.authors.slice(0, 3).join(', ') + (p.authors.length > 3 ? ' et al.' : '');
+        const year = p.year ? ` (${p.year})` : '';
+        const venue = p.venue ? ` ${p.venue}.` : '';
+        const arxiv = p.arxivId ? ` arXiv:${p.arxivId}` : '';
+        const link = p.url ? `\n     ${p.url}` : '';
+        return `${i + 1}. ${p.title}${year}\n     ${authors}.${venue} Citations: ${p.citationCount}.${arxiv}${link}`;
+    }).join('\n\n');
+    const venues = s.summary.top_venues.slice(0, 5).map(v => `${v.venue} (${v.count})`).join(', ');
+    return {
+        content: [
+            {
+                type: 'text',
+                text: `AI Papers, citation-ranked. ${s.date}, ${s.total_papers} papers.\n` +
+                    (venues ? `Top venues: ${venues}\n\n` : '\n') +
+                    lines +
+                    `\n\nSource: Semantic Scholar Graph API. Captured ${s.capturedAt}.`,
+            },
+        ],
+    };
+});
+// ── Tool: get_arxiv_recent (free) ───────────────────────────────────
+registerTool('get_arxiv_recent', 'Most recent arXiv submissions in cs.AI / cs.LG / cs.CL / cs.CV. Single Atom API call, deduped by arxivId, top 50 by submission date. The firehose pair to get_ai_papers_trending: this shows what just dropped, ai_papers_trending ranks by citation count. Each paper carries arxivId, title, abstract, authors, primary category, publication date, html and pdf URLs. Refreshed daily at 11:30 UTC. Free, no auth.', {
+    limit: z.number().min(1).max(50).optional().describe('Max papers to include in the rendered list (default 25, max 50)'),
+}, async ({ limit }) => {
+    const data = (await fetchJSON('/papers/arxiv-recent'));
+    const s = data.snapshot;
+    const cap = Math.min(limit ?? 25, s.papers.length);
+    const lines = s.papers.slice(0, cap).map((p, i) => {
+        const authors = p.authors.slice(0, 3).join(', ') + (p.authors.length > 3 ? ' et al.' : '');
+        const cat = p.primaryCategory ? ` [${p.primaryCategory}]` : '';
+        const date = p.publishedAt ? ` ${p.publishedAt.slice(0, 10)}` : '';
+        return `${i + 1}. ${p.title}${cat}${date}\n     ${authors}\n     ${p.htmlUrl}`;
+    }).join('\n\n');
+    return {
+        content: [
+            {
+                type: 'text',
+                text: `arXiv recent submissions. ${s.date}, showing ${cap} of ${s.total_papers}.\n` +
+                    `Categories: ${s.categories_queried.join(', ')}\n\n` +
+                    lines +
+                    `\n\nSource: arxiv.org. Captured ${s.capturedAt}.`,
+            },
+        ],
+    };
+});
+// ── Tool: get_hf_trending (free) ────────────────────────────────────
+registerTool('get_hf_trending', 'Top 30 most-downloaded models and top 30 most-downloaded datasets on Hugging Face. Captured daily at 12:00 UTC against the public HF API (no auth). Each entry carries id, downloads, likes, pipeline_tag (models), tags, lastModified, private, gated. Use this when an agent wants the current shape of the open model/dataset ecosystem. Once enough daily snapshots accumulate, day-over-day download deltas become a real trending signal. Free, no auth.', {
+    section: z.enum(['models', 'datasets', 'both']).optional().describe('Which to include in the rendered output (default both)'),
+    limit: z.number().min(1).max(30).optional().describe('Max items per section (default 15, max 30)'),
+}, async ({ section, limit }) => {
+    const data = (await fetchJSON('/hf/trending'));
+    const s = data.snapshot;
+    const sec = section ?? 'both';
+    const n = Math.min(limit ?? 15, 30);
+    const fmtCount = (x) => x >= 1_000_000 ? `${(x / 1_000_000).toFixed(1)}M` : x >= 1_000 ? `${(x / 1_000).toFixed(1)}K` : String(x);
+    const modelLines = s.models.items.slice(0, n).map((m, i) => {
+        const tag = m.pipeline_tag ? ` [${m.pipeline_tag}]` : '';
+        return `${i + 1}. ${m.id}${tag}  ${fmtCount(m.downloads)} downloads, ${fmtCount(m.likes)} likes`;
+    }).join('\n');
+    const datasetLines = s.datasets.items.slice(0, n).map((d, i) => `${i + 1}. ${d.id}  ${fmtCount(d.downloads)} downloads, ${fmtCount(d.likes)} likes`).join('\n');
+    const pipelines = s.summary.top_pipeline_tags.slice(0, 5).map(t => `${t.tag} (${t.count})`).join(', ');
+    const namespaces = s.summary.top_namespaces.slice(0, 5).map(n => `${n.namespace} (${n.count})`).join(', ');
+    let text = `Hugging Face top-downloaded. ${s.date}.\n`;
+    if (pipelines)
+        text += `Top pipeline tags (models): ${pipelines}\n`;
+    if (namespaces)
+        text += `Top namespaces: ${namespaces}\n`;
+    if (sec === 'models' || sec === 'both')
+        text += `\nModels:\n${modelLines}\n`;
+    if (sec === 'datasets' || sec === 'both')
+        text += `\nDatasets:\n${datasetLines}\n`;
+    text += `\nSource: huggingface.co. Captured ${s.capturedAt}.`;
+    return { content: [{ type: 'text', text }] };
+});
 // ── Start ───────────────────────────────────────────────────────────
 async function main() {
     const transport = new StdioServerTransport();

package/dist/sanitize.d.ts

@@ -0,0 +1,40 @@
+/**
+ * Defense-in-depth output scrub for tool responses.
+ *
+ * The MCP server returns text from the TensorFeed worker into a host LLM
+ * (Claude Desktop, Cursor, Cline, Roo, etc). Any text we forward becomes
+ * untrusted input to that downstream model. The worker already runs
+ * sanitize.ts on its agent-facing endpoints, but this layer protects
+ * against:
+ *
+ *   1. New worker endpoints that forget the worker-side scrub
+ *   2. Future tools that aggregate or reformat upstream text
+ *   3. Drift between the worker's sanitization rules and what the MCP
+ *      server should consider safe
+ *
+ * The rule set mirrors worker/src/sanitize.ts:
+ *   - Strip ASCII control chars (preserves \n \r \t)
+ *   - Strip bidi controls and zero-width chars used in spoofing
+ *   - Neutralize role-confusion markers (ChatML / Llama / Mistral chat
+ *     tokens, jailbreak preamble lines, "ignore previous instructions")
+ *   - Collapse runs of whitespace
+ *
+ * No length cap is applied here. Tool output is conversational and may
+ * legitimately be tens of KB; upstream KV size limits already bound it.
+ */
+/**
+ * Scrub a single string for return to a host LLM. Idempotent and pure.
+ */
+export declare function sanitizeForLLM(input: string): string;
+/**
+ * Recursively scrub the `text` field of every text-content item in an
+ * MCP tool response. Non-text items (resources, images) pass through
+ * untouched. Returns a new content array; the input is not mutated.
+ */
+export declare function sanitizeToolResponse<T extends {
+    content: Array<{
+        type: string;
+        text?: string;
+        [k: string]: unknown;
+    }>;
+}>(response: T): T;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tensorfeed/mcp-server",
-  "version": "1.11.1",
+  "version": "1.13.0",
   "mcpName": "ai.tensorfeed/mcp-server",
   "description": "MCP server for TensorFeed.ai - AI news, status, model data, and premium endpoints (routing, cost projection, forecast, history series, news search, webhook watches with digest tier, enriched directory) for AI agents",
   "main": "dist/index.js",

package/server.json

@@ -3,7 +3,7 @@
   "name": "ai.tensorfeed/mcp-server",
   "title": "TensorFeed",
   "description": "AI news, model pricing, service status, and machine-payable premium tools. AFTA-certified.",
-  "version": "1.11.0",
+  "version": "1.13.0",
   "websiteUrl": "https://tensorfeed.ai/developers/agent-payments",
   "repository": {
     "url": "https://github.com/RipperMercs/tensorfeed-mcp",
@@ -14,7 +14,7 @@
       "registryType": "npm",
       "registryBaseUrl": "https://registry.npmjs.org",
       "identifier": "@tensorfeed/mcp-server",
-      "version": "1.11.0",
+      "version": "1.13.0",
       "runtimeHint": "npx",
       "transport": {
         "type": "stdio"
@@ -44,7 +44,10 @@
         "get_ai_today",
         "get_agent_activity",
         "mcp_registry_snapshot",
-        "probe_latest"
+        "probe_latest",
+        "get_ai_papers_trending",
+        "get_arxiv_recent",
+        "get_hf_trending"
       ],
       "premiumTools": [
         "premium_routing",