This document is the release history of Cloud Agents CN.
New
New
This upgrade centers on the core theme of "Cloud Use": Agents are no longer just conversational assistants, but automation units that can securely call cloud APIs with a machine identity and autonomously complete cloud tasks. Together with the diagnostics assistant and Session dynamic Patch, developers see a major efficiency boost across the entire pipeline from assembly to debugging.
Cloud Use—makes "using the cloud" no longer require a person to be present. Assign a task once and close the loop in the cloud, letting Agents use the cloud securely and completely with a machine identity. It integrates the skill marketplace, the MCP credential vault, and Agent assembly into a single Agent.
When a Session gets stuck, errors out, or behaves abnormally, you no longer need to sift through logs to piece together context. The diagnostics assistant helps you complete the entire "locate → attribute → recommend" flow right inside the console.
A running Session can now have its configuration changed directly, without needing to archive the old session and rebuild a new one just to try out an idea.
This upgrade evolves the Agent from a "single cloud execution environment" into an open system with "bring-your-own infrastructure + intranet tool interconnection + programmable environment initialization."
Run Agent tool execution on your own infrastructure while LLM inference stays in the cloud. Your code, file system, and network egress remain entirely within your control, meeting data compliance and intranet access requirements.
Securely and reliably relay MCP servers on your intranet to the Cloud Agent through a Connector, without exposing intranet services to the public internet.
Run a custom initialization script before executing Agent tools, which can be used to install dependencies, preconfigure the environment, fetch resources, or inject credentials. It runs automatically once before the first tool call of each Session.
This upgrade evolves the Agent from "one role in a single conversation" into a complete system with "multi-role collaboration + cross-session memory + scheduled automatic execution."
Multiple Agents can be orchestrated to work together within a single conversation, with each Agent executing in an independent thread and exchanging messages through a Mailbox mechanism. This is well suited to breaking down complex tasks into multi-role pipelines such as "research + coding + testing," with no manual orchestration required.
Custom environment variables (key-value pairs) can be passed in when creating a session, and the Agent can use them directly via
After a session ends, the platform automatically reviews the conversation content and extracts key information to write into the Memory Store. The next time the same Agent starts, existing memory is loaded automatically, achieving continuous context accumulation across sessions—the more you use it, the better the Agent understands your project.
Support for configuring scheduled triggers via cron expressions, as well as manual one-click execution. Provides complete lifecycle management: pause / resume / archive / automatic retry / automatic pause on consecutive failures. This is well suited to periodic scenarios such as daily report generation, scheduled inspections, and data synchronization.
In this upgrade, the Cloud Agents mainline focuses on "smoother Agent input, freer model choice, more controllable tools, and a more stable experience," with a round of concentrated upgrades covering four major themes: GitHub repository integration, new model exposure, custom Agent tool support, and a batch of bug fixes and experience improvements.
Hello, World!
I'm Cloud Agents CN, and I'm glad to meet you.
I'm a fully managed AI Agent cloud service, with a built-in inference-execution engine and tool runtime environment, and I persistently store task conversations and file history. Through an architecture that separates the "brain" from the "hands," I can manage any execution environment while keeping your data secure. Developers only need to call a REST API to quickly build continuously running, self-evolving Agent applications—no need to build your own infrastructure, no need to maintain sandboxes, and no need to worry about model upgrades.
Here's what we can do together.
Create and manage Agents declaratively through the Agent API—configuring models, tools, system instructions, and MCP integrations. Support for version management and rollback; when the platform upgrades models and orchestration strategies, already-integrated applications automatically get stronger with zero code changes.
Long conversations run continuously: continuous execution spanning hours to days, with support for checkpoint recovery and cross-turn persistence of the event stream. The real-time session event stream supports automatic reconnection after network jitter and resumes from the checkpoint, with no refresh needed throughout.
The platform provides a securely isolated Sandbox execution environment with configurable network policies, requiring no self-built infrastructure. It includes 8 out-of-the-box tools: bash, read, write, edit, glob, grep, web_fetch, web_search. When you need more capabilities, you can integrate external MCP Servers for unlimited extensibility.
Through the SSE real-time event stream, every step of the Agent's thinking and every tool invocation is pushed in real time, making it easy to observe, audit, and replay.
0.1.0 (2026-07-07)
Forward Mode General Availability: Delivering Agents to Enterprise End Users
New
- Forward Mode is now generally available: callers only need to provide
template_id+identity_idto deliver a preconfigured Agent to end users - Forward Template and Identity: administrators preset Agent templates and inject personalized configuration for each end user (tools, Skills, files, Memory Store, environment variables), with memory and permissions automatically isolated
- Forward Schedule: supports cron, one-off, and manual triggers, with automatic execution on schedule and results that can be pushed to IM channels
- Forward Channel: built-in IM integrations for DingTalk, Feishu, WeCom, WeChat, and Slack, with support for message send/receive, QR code binding, approvals, and file reception
- Forward Schedule: supports creating, querying, and deleting scheduled tasks in natural language; monitoring-type tasks can be given start and stop times
- Forward Channel: added the Lark channel, covering QR code registration, long-lived session connections, message sending, and streaming card updates
- Forward Channel: available channels and user quotas can be configured per region
- Forward Identity / Channel: after an Identity is deleted or a Template is archived, the Runtime Channel is automatically archived and related sessions are stopped
- Enhanced enterprise integration stability and input validation for Forward mode, lowering the onboarding bar
- Forward Schedule: fair scheduling and automatic recovery on failure, for more stable operation under high concurrency
- Forward Schedule: scheduling parameters support hot updates and take effect without a restart
- Forward Schedule: when creating a task via natural language, users are guided to supply necessary conditions if information is missing
- Forward Channel: strengthened Runtime Channel security and user attribution validation, reducing credential exposure
- Forward Channel: improved approval and streaming card experience, reducing the risk of approvals crossing between concurrent sessions
0.0.8 (2026-07-05)
Event-Driven + Visual Capabilities Make Agents More Proactive
New
- Agents have built-in Image Search (ImageSearch) and Image Generation (ImageGen) tools, so images can be generated or retrieved directly in a conversation
- Support for Webhook event-driven push: lifecycle changes of Agents and Sessions are notified to a developer URL in real time, with no polling required
- When creating an Agent, you can select the model context window size (such as 200K or 1M) to flexibly fit different scenarios
- File uploads support more formats, including images, documents, archives, and audio/video
- The Session input box has been changed to a ChatGPT-style send/stop toggle button, making the operation more intuitive
- Skill / MCP tags in a Session can be clicked to view details
- MCP server configuration adds an "Interaction Policy" option
- The Skill Center supports multi-select + one-click batch import, so Agent capabilities can be assembled quickly
- The Session detail page adds Agent / Environment / Vault tabs, giving a clear view of associated resources at a glance
0.0.7 (2026-06-30)
CloudUse Launches: Let Agents Use the Cloud with a Machine Identity
This upgrade centers on the core theme of "Cloud Use": Agents are no longer just conversational assistants, but automation units that can securely call cloud APIs with a machine identity and autonomously complete cloud tasks. Together with the diagnostics assistant and Session dynamic Patch, developers see a major efficiency boost across the entire pipeline from assembly to debugging.
Cloud Use — Assign Once, Close the Loop in the Cloud
Cloud Use—makes "using the cloud" no longer require a person to be present. Assign a task once and close the loop in the cloud, letting Agents use the cloud securely and completely with a machine identity. It integrates the skill marketplace, the MCP credential vault, and Agent assembly into a single Agent.
- Skill Marketplace Import: Browse the Qoder Skill Center, with support for category filtering, an Alibaba Cloud zone, and multi-select batch import; a persistent bottom action bar lets you confirm and submit at any time
- MCP Credential Management: Pick MCP servers from the Vault credential store, collapse and group them by service, lazy-load on demand, and follow a wizard-guided flow for OAuth onboarding (including a dedicated Alibaba Cloud Cloud Use flow)
- One-Step Assembly: Import skills → configure MCP permissions → create an Agent, all completed within the same interface
Diagnostics Assistant
When a Session gets stuck, errors out, or behaves abnormally, you no longer need to sift through logs to piece together context. The diagnostics assistant helps you complete the entire "locate → attribute → recommend" flow right inside the console.
- Real-Time Diagnostics: Automatically identifies errors, infers root causes, and provides actionable fix recommendations
- One-Click Context Pull: events, turns, and tool_calls are automatically collected as diagnostic input, saving you a manual export
- Shareable Conclusions: Diagnostic results can be exported, copied, and shared, making it easy to paste into a ticket or sync with your team
Session Dynamic Patch
A running Session can now have its configuration changed directly, without needing to archive the old session and rebuild a new one just to try out an idea.
- Instant Config Delivery: Support for PATCH tools and mcp_servers
- Takes Effect on the Next Turn: No need to restart the session or interrupt the current task context
- Typical Scenarios: A/B debugging prompts, temporarily adding a tool to an Agent, switching models to compare results
Improvements
- The MCP tool selector supports collapsible groups, lazy loading, and type filtering, so you can quickly locate items even in long lists
- Skill package downloads have been changed to native browser downloads to avoid cross-origin compatibility issues
- Self-hosted mode no longer forcibly injects workdir, switching to relative-path hints that better match local project conventions
- Improved Deployment creation and runtime experience: prefilled resource configuration and correct display of the initial event block
- The Skill Center import button is now persistently pinned to the bottom action bar, so confirmation actions in multi-select scenarios no longer require scrolling to find the button
- The environment list now shows a "Self-Hosted" tag for self_hosted types, making environment ownership clear at a glance
0.0.6 (2026-06-18)
This upgrade evolves the Agent from a "single cloud execution environment" into an open system with "bring-your-own infrastructure + intranet tool interconnection + programmable environment initialization."
Self-Hosted Sandbox
Run Agent tool execution on your own infrastructure while LLM inference stays in the cloud. Your code, file system, and network egress remain entirely within your control, meeting data compliance and intranet access requirements.
- Create a
self_hostedtype Environment, where the Worker listens for and executes tasks via the Work Queue protocol - Support for the Go SDK, CLI, and direct HTTP integration in any language
MCP Connector
Securely and reliably relay MCP servers on your intranet to the Cloud Agent through a Connector, without exposing intranet services to the public internet.
- Create a Tunnel → start a Connector (SSE reverse connection) → bind a Session to tunnel_id, completed in three steps
- Domain allowlist + IP-level SSRF protection for clear security boundaries
- Horizontal scaling across multiple instances, with a single Tunnel supporting load balancing across multiple Connectors
Custom Initialization Scripts
Run a custom initialization script before executing Agent tools, which can be used to install dependencies, preconfigure the environment, fetch resources, or inject credentials. It runs automatically once before the first tool call of each Session.
Improvements
- Tool execution timing display has been changed to wall-clock anchor-based timing to avoid timing jitter caused by streaming rendering
0.0.5 (2026-06-14)
This upgrade evolves the Agent from "one role in a single conversation" into a complete system with "multi-role collaboration + cross-session memory + scheduled automatic execution."
Multi-Agent Collaboration
Multiple Agents can be orchestrated to work together within a single conversation, with each Agent executing in an independent thread and exchanging messages through a Mailbox mechanism. This is well suited to breaking down complex tasks into multi-role pipelines such as "research + coding + testing," with no manual orchestration required.
Environment Variables
Custom environment variables (key-value pairs) can be passed in when creating a session, and the Agent can use them directly via $ENV_NAME in the Sandbox. This is well suited to injecting runtime configuration such as database connection strings and API Keys, without hardcoding them in code or the system prompt.
Dreaming (Intelligent Memory)
After a session ends, the platform automatically reviews the conversation content and extracts key information to write into the Memory Store. The next time the same Agent starts, existing memory is loaded automatically, achieving continuous context accumulation across sessions—the more you use it, the better the Agent understands your project.
Scheduled Tasks (Deployments)
Support for configuring scheduled triggers via cron expressions, as well as manual one-click execution. Provides complete lifecycle management: pause / resume / archive / automatic retry / automatic pause on consecutive failures. This is well suited to periodic scenarios such as daily report generation, scheduled inspections, and data synchronization.
0.0.4 (2026-06-10)
In this upgrade, the Cloud Agents mainline focuses on "smoother Agent input, freer model choice, more controllable tools, and a more stable experience," with a round of concentrated upgrades covering four major themes: GitHub repository integration, new model exposure, custom Agent tool support, and a batch of bug fixes and experience improvements.
GitHub Repository Integration
- When creating a session, you can directly mount a GitHub repository as a session resource, without first cloning it locally and uploading it—saving a step wherever possible.
- The session creation page adds an inline GitHub Personal Access Token input entry, so private repositories work as soon as you fill it in; the credential is valid only for the current session and no longer depends on a separate credential vault configuration.
New Model Exposure
- Integrated multiple mainstream models such as Qwen, DeepSeek, GLM, Kimi, and MiniMax all at once, uniformly displaying the credit multiplier as
${price_factor}xfor easy selection of the model and cost tier suited to the task.
Custom Agent Tool Support
- Agent configuration supports declaring custom tools of
type: custom, with the execution logic implemented by the client, so that Agent capabilities can be freely extended per business scenario, no longer limited to the built-in toolset. - Tool permissions add a three-tier policy:
always_allowfor automatic pass-through,always_askto request user confirmation before execution, andalways_denyto reject directly—bringing "human-in-the-loop" to life for the first time, so sensitive actions can be forced to require manual confirmation before execution. - MCP tools support whole-server authorization by server prefix, so configuring it once permits all tools under that server—no more checking items one by one.
Bug Fixes & Experience Improvements
- The session event list supports cursor pagination and scroll loading, so events beyond 100 are no longer truncated and lost.
- The file list is displayed in reverse chronological order of creation, with the newest files on top; new "Purpose / Status / File ID" columns and one-click copy were added.
- The SSE real-time event stream supports Last-Event-ID resumable delivery, automatically reconnecting and resuming after network jitter, with no manual refresh needed.
- The 10-second polling of the session list no longer causes the whole page to flicker; the event area now has its own scrollbar and no longer inadvertently scrolls the whole page.
- The skill display in Agent edit mode is now fully consistent with the detail page; when the number of skills reaches the limit of 20, the front end gives a friendly prompt and intercepts early to avoid submission failure.
- Pressing Enter while composing with a Chinese input method no longer sends messages by mistake; the file mount path no longer shows the
/data/datadouble prefix. - Ticket feedback supports image upload and one-click screenshots, and its entry has been moved to the settings menu, making the feedback flow smoother.
- Sessions stuck during cancellation now recover automatically; concurrent resource additions use a
FOR UPDATElock, preventing resources from overwriting each other. - When uploading a Skills zip package, Windows backslash paths are automatically normalized, ensuring consistent cross-platform behavior on Mac / Windows.
0.0.3 (2026-06-09)
Hello, World!
I'm Cloud Agents CN, and I'm glad to meet you.
I'm a fully managed AI Agent cloud service, with a built-in inference-execution engine and tool runtime environment, and I persistently store task conversations and file history. Through an architecture that separates the "brain" from the "hands," I can manage any execution environment while keeping your data secure. Developers only need to call a REST API to quickly build continuously running, self-evolving Agent applications—no need to build your own infrastructure, no need to maintain sandboxes, and no need to worry about model upgrades.
Here's what we can do together.
Declarative Agents—Define Once, Evolve Long-Term
Create and manage Agents declaratively through the Agent API—configuring models, tools, system instructions, and MCP integrations. Support for version management and rollback; when the platform upgrades models and orchestration strategies, already-integrated applications automatically get stronger with zero code changes.
Long-Running Tasks—Run Steadily, Resume Reliably
Long conversations run continuously: continuous execution spanning hours to days, with support for checkpoint recovery and cross-turn persistence of the event stream. The real-time session event stream supports automatic reconnection after network jitter and resumes from the checkpoint, with no refresh needed throughout.
Managed Execution Environment, Out-of-the-Box Toolset
The platform provides a securely isolated Sandbox execution environment with configurable network policies, requiring no self-built infrastructure. It includes 8 out-of-the-box tools: bash, read, write, edit, glob, grep, web_fetch, web_search. When you need more capabilities, you can integrate external MCP Servers for unlimited extensibility.
Fully Observable—Every Step Is Visible
Through the SSE real-time event stream, every step of the Agent's thinking and every tool invocation is pushed in real time, making it easy to observe, audit, and replay.
Complete Resource Management APIs
- File upload and mounting (Files API): provide file context to a Session, and Agent artifacts can also be downloaded via the API
- User credentials (Vaults API): securely manage access credentials, injected on demand at Session runtime
- Agent Skills (Skills API): attach domain expertise to an Agent, making a general-purpose Agent perform more professionally on specific tasks
- Persistent memory (Memory Stores API): let an Agent's learning outcomes and outputs persist across 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 (binding Agent + Environment)
POST /v1/cloud/sessions/{id}/events — Send a message
GET /v1/cloud/sessions/{id}/events/stream — Receive responses via streaming (SSE)
Come on, let's build AI applications a new way together.