A hands-on report tests the MCP task primitives introduced in protocol version 2025-11-25: tasks/get, tasks/result, tasks/cancel and tasks/list. In the author’s note-service implementation, task creation returned in about four milliseconds and status updates arrived roughly once per second. The test also identified several operational issues: a client that had not called listTools() could silently fall back to synchronous execution; instance-local stores made tasks unavailable in a stateless setup; InMemoryTaskStore lost task data after a restart; and tasks/cancel changed task status without signaling the running background code. The report concludes that task primitives standardize how long-running tasks are invoked, while applications remain responsible for shared and durable storage, background execution and cancellation handling.
According to iThome Taiwan, RSA Conference research forecasts that AI agents will discover and patch more vulnerabilities in 2027, forcing security leaders to rethink vulnerability management, staffing, GRC, incident response, and recovery. RSAC also expects more stateless systems to move to the cloud, where clean images can replace compromised instances, while stateful systems still require isolation and patching. Its research additionally identifies MCP agent security, model-generated tool calls, model poisoning, and malicious package recommendations as emerging risks.
An engineering team deployed three MCP Servers for ticket queries, release records, and configuration data, after which they ran for two weeks with roughly 200 calls per day, according to a Juejin AI engineering retrospective. The team recommends reserving stdout exclusively for JSON-RPC messages, sending logs to stderr, trimming tool responses, paginating list results, setting timeouts below 30 seconds, and returning structured errors with actionable hints. The article’s deployment figures and before-and-after results are self-reported and have not been independently verified.
TechCrunch AI reports that Nectar has launched Creator Agent, an AI-powered marketing agent for creator discovery, outreach, partnership management, and campaign monitoring. Available globally for Instagram creators and TikTokers, it can search more than 300 million creator profiles, draft personalized messages, analyze content and brand fit, and rank potential partners, according to Nectar. The company also plans to expose the agent through MCP and assistants including Claude, ChatGPT, Gemini, and Copilot.
In a Hacker News Show HN post, the developer describes ZigZag as a visual embroidery editor built in Rust with built-in Model Context Protocol support for AI agents to create and edit designs. The developer says it can turn SVG artwork into stitches, export PES, DST, EXP, JEF and VP3 files, and send designs to a Brother Skitch PP1 over Bluetooth. It includes a visual studio, a command-line tool and an MCP bridge for the same workflow. The release is available for macOS and Windows, according to the post.
AWS Machine Learning Blog published a guide to securing Amazon Bedrock Managed Knowledge Base tools for multi-tenant AI agents. Its architecture uses Amazon Bedrock AgentCore Gateway and Model Context Protocol tools, with trusted code injecting verified user attributes and overriding model-generated search context. Additional AgentCore policies and resource-based controls provide independent safeguards.
Clipto has launched Clipto Studio, an AI memory platform that lets assistants search and analyze recordings of calls, presentations and meetings. SiliconANGLE AI reports that it integrates with ChatGPT, Claude and other assistants through the Model Context Protocol, while proprietary models run locally on Mac and Windows computers. Clipto says the platform indexes speech, speaker voices, facial expressions and on-screen visuals into a searchable dataset.
A commentary on Kingbase’s launch describes KES AI database security as a combination of MCP isolation, row-level access controls, semantic auditing and full-chain LLM output validation. According to Kingbase, an MCP Server between AI clients and KingbaseES means clients do not hold database credentials, while the server validates SQL types and assigns request risk levels before forwarding requests. The report also says the architecture still requires securing the MCP Server, defining AI permissions and protecting audit logs from bypass or tampering. The supplied report does not independently verify these claimed capabilities or provide implementation details or test results.
Adobe is separating Creative Cloud capabilities from traditional app interfaces so users can enter through chatbots or AI agents, with Creative Agent providing a conversational route inside its apps. Firefly serves as an AI-first interface, while Adobe says conventional controls will remain important for professionals who need precision; TechRadar notes that most capability details come from Adobe executives.
Ant Lingguang 2.0 has been released with a focus on helping users complete goals rather than merely generate mini-apps, according to Ant Lingguang CTO Mo Linjian. The company says the assistant can retain context, track tasks, proactively remind users and act with authorization through integrations such as Alipay, email, DingTalk and Feishu, while using MCP and A2A for service connections. TMTPost's interview also highlights unresolved questions around trust, commercialization and conflicts with platform business models.
InfoQ’s 2026 trends report summarizes its editorial team’s and podcast guests’ discussions of emerging AI, ML and data engineering developments and directions to watch over the coming 12 months. It organizes technologies and topics on an adoption curve, including innovator, early-adopter, early-majority and late-majority stages. Topics include agent communication protocols, inference engines, harness engineering, physical AI, AI agents, model and agent evaluation, context engineering, and AI-assisted software development. The report also presents forecasts for 2027.
Google Cloud’s 2027 cybersecurity forecast says AI agents could help security operations centers collect and analyze incident data, prioritize alerts, and summarize investigations. It warns that excessive permissions, shared service accounts, prompt injection, shadow agents, software supply-chain attacks, and theft of AI assets could widen the attack surface. The report recommends independent agent identities, restricted access, operation logging, red-team testing, and human approval for high-risk actions.
AWS outlined an event-driven architecture for synchronizing AWS Agent Registry records with MCP server tools and schemas. AWS CloudTrail captures UpdateAgentRuntime API calls, Amazon EventBridge routes matching events, and AWS Lambda queries the MCP server, compares its tools with the registry, and updates the matching record. The design supports single-account deployments and cross-account event forwarding. Updated records return to DRAFT status and must pass the approval workflow before becoming discoverable again. The sample requires an existing matching registry record, does not automatically register previously unknown servers, has no manual force-sync path, and does not retry when a cold MCP server returns zero tools.
Mastra introduced MCP tasks for MCPServer and MCPClient, allowing long-running tool calls to execute in the background while an agent continues responding. The server returns a task ID that the client polls for working, input_required, completed, failed, or cancelled status, with completed results returned to the agent. The feature requires @mastra/mcp 2.7.0 and @mastra/core 1.81.0 or later.
a16z argues that AI agents are becoming primary internet users and proposes “B2A,” a model in which businesses sell to agents that choose products, subscriptions and spending on users’ behalf. The report maps agent applications and infrastructure, citing Shopify’s eightfold growth in AI-driven store traffic, Human Security’s eightfold rise in agentic browser traffic, Supabase’s AI-created databases and Vercel’s 28-fold increase in agent-triggered deployments. a16z says adoption is measurable, but most business models enabled by the shift remain undefined.
InfoQ 中文 AI&大模型 reports that Manus replaced traditional programming tests with a four-hour task in which candidates use Manus to solve a real problem and explain their AI-assisted workflow. Manus co-founder and chief product officer Tao Zhang says product managers and designers on his team can now write code and submit repository changes, though the company’s job listings retain conventional engineering titles. A tech hiring analysis cited by Lenny Rachitsky adds that U.S. software-engineer hiring has fallen below half its 2022 peak, while FDE openings have increased sharply and AI-related keywords have become more prominent in product and design roles.
The author tested an Express-based MCP Server running an eight-second slow_task during deployment shutdowns. After the server received SIGTERM two seconds into the task, one client waited 58 seconds for an already-dead connection before reporting an MCP timeout, while server.close() preserved active requests but waited for connections rather than task completion. A tested shutdown sequence marked the server unhealthy, stopped accepting traffic, waited for tracked in-flight tasks, and actively closed remaining connections.
An engineering analysis of ETF Agent examines why awaiting cancellation or closure can delay an already known failure, replace its error classification, or affect retry and fallback decisions. Using controlled reproductions and simplified teaching code—not a reported production incident—the analysis covers oversized web responses, HTTP 401 responses from remote MCP services, shared Wind MCP connections, retrieval fallback, and resource ownership. The reported fixes preserve the original failure classification and allow the call to settle without waiting for cleanup. However, a settled result does not prove that remote requests stopped or that all resources were reclaimed. The report also notes that HTTP classification and Host retry rules were tested separately rather than as one end-to-end 401-to-retry chain.
A LangChain case study reports that Workday built an internal platform for developing, deploying, and operating AI agents in production on its own infrastructure. It uses LangGraph for orchestration and LangSmith for tracing, observability, and evaluations. Components include an agent server, data-access controls tied to Workday's audit policies, a collaborative evaluation platform, and a Python SDK. Workday says teams can now run a prototype in a staging environment in under one day, compared with four to five months previously. The team also grew from four people to roughly 30.
Google says its Agent Skills repository has grown to more than 21,000 GitHub stars and over 170 skills from teams across the company, distributed across multiple AI platforms. It has added semantic-versioning checks, evaluation gates, deterministic linters, and an AI review agent to its CI/CD workflow, while requiring human approval before publication. Google says developers can reuse the same gates and review rules in their own skill repositories.
Black Kite’s 2026 AI Supply Chain Report examined 18,160 AI companies worldwide and found gaps in patching, credential protection, email authentication, network security and ransomware resilience. Among the findings, 34.2% of assessed companies had at least one critical vulnerability, while 48.8% of 1,732 logged Model Context Protocol security findings were rated High or Critical. Black Kite recommends continuous monitoring, supplier engagement and auditing AI integrations, components, permissions and data access.
A Juejin AI tutorial shows how to deploy MySQL 8.0 and MySQL MCP Server with Docker Compose, then connect them to WorkBuddy through a token-authenticated MCP endpoint. WorkBuddy uses the mysql_query tool to retrieve database information for asset inventory, health analysis, and a standalone HTML inspection report, while the setup uses a read-only database account and prohibits write operations. The demonstrated report covers assets, data distribution, indexes, diagnostics, and DBA recommendations based on the queried results.
Roblox said more than 80% of the platform's 100 most-played games used AI-assisted development in the past month, while daily cloud asset uploads rose threefold from the prior period. Its prompt-to-game tool Build has attracted more than 500,000 users and produced over 70,000 games during testing, alongside tools such as Studio Assistant and MCP connections to external AI services. 游戏茶馆, writing for 钛媒体, said the wider AI-game industry remains at an exploratory stage, with few examples of scaled production, stable users, and commercial returns.
FormAssembly announced a seven-month platform overhaul that combines form creation, workflow automation, integrations, and compliance management in a no-code environment. The release adds conversational intake agents, MCP support, and the Fai assistant, which can generate forms, approval workflows, and documentation requirements from natural-language descriptions. FormAssembly also says its AI builder can reduce development time from days or weeks to minutes, based on tests that recreated existing customer workflows.
Airbnb said in its third-quarter shareholder letter that it plans to begin offering consumers an AI-native travel agent in the first half of 2027. The company said the travel-focused agent would dynamically combine maps, photos, comparisons, calendars and itineraries, and assist with search, booking, changes, payment issues, refunds, rebooking and 24-hour safety support. Airbnb also said it would remain focused on travel while allowing external services to connect through methods such as Model Context Protocol (MCP).
A Chinese-language commentary explains how to extend Claude Code through Skills, Hooks, and the Model Context Protocol. It describes creating SKILL.md files for reusable project rules and workflows, configuring Hooks to run repeatable actions such as linting after edits, and using MCP servers to connect external services such as databases, APIs, and design files. The article also organizes Claude Code usage into prompt, context, and harness layers. Its configuration details and claimed workflows are not independently verified.
A Juejin author documents building an MCP server on macOS and connecting it to a local Qwen2.5 3B model through Ollama. The implementation exposes read_file, search_in_file, and get_weather tools, then uses a client-side ReAct loop to let the model request tools and incorporate their results. The author also records setup and debugging issues involving Python 3.12, MCP SDK 1.30.0, stdio transport, and MCP Inspector.
The author argues that coding agents and AI assistants use incompatible memory formats, including AGENTS.md, CLAUDE.md, Cursor rules, Anthropic's memory tool, GitHub Copilot memory, and MCP-based stores. Drawing on OpenTelemetry, the article proposes a neutral specification covering memory identity, addressing, lifecycle, provenance, and shared access, while arguing that storage, ranking, embeddings, and interfaces should remain areas of competition. The author, who builds LoreKit, presents its MIT-licensed schema as one input rather than a finished standard.
The open-source project changedetection.io describes itself as a tool for website change detection and monitoring, with notifications via Discord, email, Slack, Telegram and webhooks, and conditional triggers such as firing only when a price crosses a set threshold or a keyword appears. Its description adds AI-powered features: plain-language rules that evaluate each detected diff against a written intent (for example, "notify me only when the price drops below $50"), and AI change summaries that replace a raw diff with a sentence such as "Price dropped from $89.99 to $67.00". Users can bring their own key for models from OpenAI, Gemini, Anthropic and others via LiteLLM, or run models locally with Ollama, vLLM, LM Studio or any OpenAI-compatible endpoint. MCP support is also offered so page changes can be exposed as a resource for questions. The project is self-hostable via Docker or pip, and a paid SaaS subscription is listed at $8.99 per month. The report does not identify a specific release or date for these AI features, and the claims about fewer false positives and per-check cost come from the project itself.
A Reddit user said Claude abruptly refused to use a browser MCP server that had worked for about a year, after the model encountered a human-verification check. The user had built restricted MCP tools that enabled navigation, scrolling, clicking, form filling, button presses, and permitted JavaScript execution.
What problem it solves
Developers using multiple AI command-line tools need a consistent way to manage providers, accounts, prompts, MCP servers, sessions, and configuration. Switching configurations manually across applications can also make it difficult to inspect usage, preserve settings, or use different providers in separate terminal sessions.
Key features
Provider management — Manage API con…
What problem it solves
Users need an AI agent that can carry out tasks beyond producing code suggestions, including research, writing, automation, and data analysis. They also need access through desktop, terminal, and embeddable interfaces while retaining a broad choice of model providers and extensions.
Key features
Multiple ways to run — Use the native desktop application, a full terminal-orien…
What problem it solves
Go developers need packages for constructing MCP clients and servers while connecting them to transports and supporting protocol features such as OAuth. The SDK organizes these capabilities into importable packages and documents how supported MCP specification versions map to them.
Key features
Client and server APIs — The primary package provides APIs for constructing and u…
What problem it solves
Research context is often fragmented across PDF libraries, note applications, and separate model conversations, forcing researchers to move or copy information manually between tools. PDFs can also be difficult for models to interpret reliably, while common AI reading, translation, synchronization, and batch-import workflows require multiple plugins or repetitive manual acti…
What problem it solves
Developers need a unified way to build semantic search applications, ground language models with knowledge sources, and connect models into larger workflows. They also need support for different data types, models, deployment modes, and application interfaces without assembling each capability separately.
Key features
Unified embeddings database — The core component combines…
What problem it solves
Voice input and voice output are usually handled by separate cloud services, which means audio and transcripts leave the machine and the two halves of the voice loop do not connect. Voicebox aims to cover both directions locally: cloning voices from short samples, generating speech, dictating into any text field, and letting MCP-aware agents speak back in a cloned voice.
Key…
What problem it solves
AI coding agents need AWS-specific tools, knowledge, and guardrails to work effectively with AWS services. Teams using different coding agents also need a consistent way to configure AWS API access, guidance, and reusable skills.
Key features
AWS MCP Server — The managed MCP Server gives agents access to AWS APIs, isolated script execution, and current documentation through …
What problem it solves
Jan provides a desktop application for downloading and running LLMs with local control and privacy. It also connects to supported cloud models and exposes an OpenAI-compatible local API for use by other applications.
Key features
Local AI Models — Download and run supported LLMs from HuggingFace on the local machine. The README lists Llama, Gemma, Qwen, and GPT-oss as exampl…
What problem it solves
The server lets an MCP-compatible assistant search and read personal WhatsApp messages, contacts, chats, and message context. It also supports sending messages and media files to individuals or groups while keeping message history in a local SQLite database.
Key features
WhatsApp message and contact search — The MCP tools support searching contacts by name or phone number, l…
What problem it solves
AI assistants need structured access to design-system data and controls in Figma to extract tokens, inspect components, create designs, and debug design-to-code differences. Teams also need deterministic checks for accessibility, design-code parity, and design-system health without imposing a specific framework or house style.
Key features
Design system extraction and token …
What problem it solves
Managing databases across many engines can require separate tools, while many database clients also depend on additional runtimes or bundled browser components. Teams and developers also need a consistent way to work with database connections across desktop, browser, self-hosted, CLI, and AI-agent environments.
Key features
Broad database coverage — Use one tool for more tha…