Cloud Agents CN release notes-Qoder CN Series(Lingma)-阿里云帮助中心

0.0.7 (2026-06-30)

Cloud Use: Agent as a machine identity

This update centers on Cloud Use, transforming an Agent from a simple conversational assistant into an autonomous unit that securely calls cloud APIs. Coupled with the Diagnostic Assistant and dynamic session patching, you can now streamline your entire workflow from configuration to debugging.

Cloud Use: Assign once, close the loop in the cloud

Cloud Use eliminates the need for human intervention in cloud operations. Delegate a task, and the Agent securely manages it on the cloud as a machine identity. This feature integrates the Skill Marketplace, MCP credential management, and Agent configuration into a single, streamlined process.

Import from Skill Marketplace: Browse the Qoder Skill Center, filter skills by category, access the Alibaba Cloud zone, and import multiple skills in bulk. A persistent bottom action bar allows you to confirm and submit at any time.
MCP Credential Management: Select MCP servers from the Vault Credential Store. Servers are grouped by service, loaded on demand, and accessible through a guided OAuth flow that includes a dedicated process for Alibaba Cloud Use.
One-step configuration: Import skills, configure MCP permissions, and create an Agent, all within a single interface.

Diagnostic assistant

When a session hangs, throws an error, or behaves unexpectedly, you no longer need to manually parse logs to understand the context. The Diagnostic Assistant identifies problems, determines root causes, and provides actionable suggestions, all within the console.

Real-time Diagnosis: Automatically identifies errors, infers the root cause, and provides actionable recommendations for fixes.
One-click Context Fetch: Automatically collects events, turns, and tool_calls to use as diagnostic input, eliminating the need for manual exports.
Shareable Results: Export, copy, and share diagnostic results to easily include them in a support ticket or share with your team.

Dynamic session patching

You can now modify the configuration of a running session directly. This eliminates the need to archive an old session and create a new one to test a change.

Instant Configuration Updates: Supports patching tools and mcp_servers.
Effective in the Next Turn: Changes apply in the next turn without restarting the session or interrupting the current task context.
Typical Use Cases: A/B testing prompts, temporarily adding a tool to an Agent, or switching models to compare performance.

Enhancements

The MCP Tool Selector now supports collapsible groups, lazy loading, and filtering by type, making it easier to find tools in long lists.
Skill pack downloads now use the browser's native download functionality to avoid cross-origin compatibility issues.
Self-hosted mode no longer forces workdir injection. Instead, it provides a relative path hint to better align with local development practices.
Streamlined the creation and execution of a Deployment. Resource configurations are now pre-filled, and the initial event block displays correctly.
The import button in the Skill Center is now persistent in the bottom action bar, so you no longer need to scroll to find it when selecting multiple skills.
The Environment List now displays a "Self-hosted" label for self_hosted environments to clarify ownership at a glance.

0.0.6 (2026-06-18)

This update transforms the Agent from a single cloud-based execution environment into an open ecosystem that supports self-hosted infrastructure, internal tool connectivity, and programmable environment initialization.

Self-hosted sandbox

Run Agent tool executions on your own infrastructure while keeping LLM inference in the cloud. Your code, file system, and network egress remain entirely within your control to meet data compliance and internal network access requirements.

Create a self_hosted environment, where a worker listens for and executes tasks through a work queue protocol.
Supports the Go SDK, CLI, and direct HTTP integration from any language.

MCP connector

Use the MCP Connector to securely and reliably expose MCP servers from your internal network to a cloud-based Agent without exposing your internal services to the public internet.

A three-step process: Create a tunnel, start the connector (using a reverse Server-Sent Events connection), and bind the session to the tunnel_id.
Features a domain allowlist and IP-level Server-Side Request Forgery (SSRF) protection for a clear security boundary.
Supports horizontal scaling with multiple instances. A single tunnel can handle load balancing across multiple connectors.

Custom initialization script

Use this script to install dependencies, pre-configure the environment, fetch resources, or inject credentials. The script runs automatically once before the first tool call in each session.

Enhancements

Tool execution duration is now measured based on wall-clock time to prevent timing jumps caused by stream rendering.

0.0.5 (2026-06-14)

This release evolves the Agent from a single-role, single-conversation entity into a complete system that supports multi-role collaboration, cross-session memory, and scheduled automatic execution.

Multi-agent collaboration

You can orchestrate multiple Agents to work together in a single session. Each Agent runs in an independent thread and communicates through a mailbox mechanism. This is ideal for breaking down complex tasks into a multi-role pipeline, such as research, coding, and testing, without manual orchestration.

Environment variables

You can now pass custom environment variables (as key-value pairs) when creating a session. The Agent can access these directly in the sandbox using $ENV_NAME. This is useful for injecting runtime configurations like database connection strings or API keys without hardcoding them in your code or system prompts.

Dreaming (intelligent memory)

After a session ends, the platform automatically reviews the conversation, extracts key information, and writes it to a Memory Store. The next time the same Agent starts, it automatically loads this memory, allowing it to learn and build context across sessions. The more you use it, the more your Agent understands your project.

Scheduled tasks (Deployment)

You can configure a Deployment to trigger on a schedule using cron expressions or run it manually with a single click. The feature provides full lifecycle management, including pause, resume, archive, automatic retries, and auto-pause on consecutive failures. It is ideal for recurring tasks like generating daily reports, scheduled inspections, or data synchronization.

0.0.4 (2026-06-10)

This release of Cloud Agents focuses on streamlining Agent inputs, providing more model flexibility, giving you more control over tools, and improving overall stability. It includes four main updates: GitHub repository integration, new model availability, custom Agent tool support, and a series of bug fixes and experience enhancements.

GitHub repository integration

You can now mount a GitHub repository directly as a resource when creating a session, eliminating the need to clone it locally and upload it first.
The Create Session page now includes an inline input for a GitHub Personal Access Token. This allows you to use private repositories on the fly for the current session only, without needing to configure a separate credential store.

New model availability

Access a wide range of leading models, including Qwen, DeepSeek, GLM, Kimi, and MiniMax. A unified cost factor, displayed as ${price_factor}x, helps you choose the right model and price point for your task.

Custom Agent tool support

Agent configurations now support declaring custom tools with type: custom. The client implements the execution logic, so you can extend an Agent's capabilities for specific business scenarios beyond the built-in toolset.
Tool permissions now include three policies: always_allow for automatic approval, always_ask to prompt for user confirmation before execution, and always_deny to reject the action. This introduces a human-in-the-loop workflow, enabling mandatory manual confirmation for sensitive operations.
You can now authorize all MCP tools under a specific server prefix at once, eliminating the need to select them individually.

Fixes and enhancements

The Session Event List now supports cursor-based pagination and infinite scrolling, ensuring that event lists longer than 100 entries are no longer truncated.
The File List is now sorted by creation time in descending order, placing the newest files at the top. Added new columns for "Purpose", "Status", and "File ID", along with a one-click "Copy" button.
The real-time Server-Sent Events (SSE) stream now supports breakpoint resume using Last-Event-ID. It automatically reconnects and resumes from where it left off after a network interruption, eliminating the need for a manual refresh.
The 10-second polling on the Session List no longer causes the entire page to flash. The Events area now has its own scrollbar, preventing accidental scrolling of the entire page.
The skill display in Agent Edit Mode is now consistent with the skill details page. A user-friendly message now appears and blocks submission if the number of skills reaches the maximum limit of 20 to prevent submission failures.
Pressing Enter while composing text with a Chinese Input Method Editor (IME) no longer accidentally sends a message. The file mount path no longer contains the duplicate /data/data prefix.
The Submit Ticket feature now supports image uploads and a one-click screenshot tool. The entry point has been moved to the Settings menu for a more convenient feedback process.
Sessions that get stuck during a cancellation process now recover automatically. Concurrent resource additions are now protected by a FOR UPDATE lock to prevent overwrites.
When uploading a skill as a ZIP file, Windows backslash paths are now automatically normalized to ensure consistent behavior across macOS and Windows platforms.

0.0.3 (2026-06-09)

Hello, World!

We are Cloud Agents CN, and we're excited to meet you.

Cloud Agents CN is a fully managed cloud service for AI Agents. It includes a built-in inference and execution engine, a tool runtime environment, and persistent storage for task conversations and file history. By decoupling the "brain" (reasoning) from the "hands" (execution), the service allows you to manage any execution environment while ensuring data security. Developers can use REST APIs to quickly build continuously running, self-evolving Agent applications—without managing infrastructure, maintaining a sandbox, or worrying about model upgrades.

Here is what we can build together.

Declare Agents once, evolve them over time

Use the Agent API to create and manage Agents declaratively—configure models, tools, system instructions, and MCP integrations. Versioning and rollback are supported. As the platform upgrades models and orchestration strategies, your integrated applications improve automatically with no code changes required.

Reliable and resumable long-running tasks

Run continuous, long-running sessions that can last for hours or even days. The service supports breakpoint resume and persists the event stream across turns. The real-time event stream automatically reconnects and resumes after a network interruption, all without requiring a page refresh.

Managed execution environment with out-of-the-box tools

The platform provides a secure and isolated sandbox execution environment with configurable network policies, so you do not need to build your own infrastructure. It includes eight out-of-the-box tools: bash, read, write, edit, glob, grep, web_fetch, and web_search. For more capabilities, you can connect to an external MCP Server for unlimited extensibility.

Full observability into every step

The real-time SSE stream delivers every thought process and tool call from the Agent in real time, allowing you to observe, audit, and replay its actions.

Complete resource management APIs

File Upload and Mounting (Files API): Provide file context to a session. Agent-generated artifacts can also be downloaded via the API.
User Credentials (Vaults API): Securely manage access credentials, which are injected on-demand into the session at runtime.
Agent Skills (Skills API): Add domain-specific knowledge to your Agent so a general-purpose Agent can perform like an expert on specific tasks.
Persistent Memory (Memory Stores API): Persist an Agent's learnings and outputs across multiple sessions.

Core API endpoints

POST /v1/cloud/agents — Create an Agent
POST /v1/cloud/environments — Create an execution environment
POST /v1/cloud/sessions — Create a session (binds an Agent and an environment)
POST /v1/cloud/sessions/{id}/events — Send a message
GET /v1/cloud/sessions/{id}/events/stream — Receive a streaming response (SSE)

Let's build AI applications in a new way.