CodeBuddy (Tencent Cloud) (AI IDE Platform Market Research)

1. Introduction

CodeBuddy is the AI coding product family launched by Tencent Cloud and the representative sample of the "domestic major-vendor ecosystem route" in this group. Unlike tools such as Cursor and Claude Code that enter with a single form factor, CodeBuddy's core narrative is full-form coverage: the plugin (IDE Plugin), the standalone IDE, and the CLI (CodeBuddy Code) share a single account, model, and quota system across all three ends; downward it connects through Tencent Cloud's R&D infrastructure (CloudBase, CloudStudio, EdgeOne, CODING, etc.), and upward it provides private deployment capabilities for customers in regulated industries such as finance.

From the perspective of AI Harness, CodeBuddy's differentiated value is concentrated at two ends. The first is the L6 governance and security layer — private deployment, enterprise-dedicated VPC, and a login system that connects to enterprise identity sources make it one of the few AI coding products that can be placed directly on procurement lists in highly regulated industries. The second is enterprise knowledge integration — through a knowledge base and multi-layer RAG, in-house standards and architecture constraints are injected into the model context, forming an enterprise adoption methodology of "model layer — plugin layer — operations layer". Its point technologies (such as rule files and MCP support) mostly follow existing industry designs; its competitiveness mainly comes from ecosystem integration and enterprise service capabilities.

1.1. Developer and Product Evolution

The product evolution shows a clear three-step path: first the plugin completes the foundation of completion and conversation, then the Craft agent fills in agent capabilities, and finally the IDE and CLI complete the form-factor loop. The official side analogizes the three forms to different "driving modes": the plugin is a manual transmission (fine-grained control), the IDE is an automatic transmission (easy and convenient), and the CLI is track mode (max-speed efficiency).

1.2. Positioning and Product Portfolio

CodeBuddy's official positioning is "the intelligent programming partner of the AI era" (Code Where You Create), while at the enterprise level it is phrased as "an enterprise-grade software engineering AI development efficiency and full-link solution". The product portfolio includes:

The latest version number is fast-changing information; at the time of writing no verifiable current unified version number was obtained, so it is recorded as [To be filled]. Verifiable version facts on the CLI side: the early public beta versions were the v1.0.0 / v1.1.0 series (around 2025-09); these versions did not yet support /agents, /hooks, /ide, background tasks, or MCP over the HTTP/SSE protocols (, based on early hands-on testing; entries such as /hooks are already visible in later version command lists, and the extent of closed gaps should be confirmed against the latest official documentation).

1.3. Pricing System

Official international-site pricing (Tencent Cloud International price details page):

Domestic version: during the public beta, individual users can use the full product line for free and seamlessly call domestic large models such as DeepSeek (per the official campaign statement; the end of the free period follows official announcements). A third-party summary also claims the enterprise edition is ¥158 / person / month (minimum 10 seats, including dedicated VPC deployment), marked [To be verified]. Private deployment pricing requires consulting the official side.

Engineering implication of the billing structure: CodeBuddy uses a credits-based rather than a pure subscription model, and the complexity of agent tasks maps directly to credit consumption. The team edition's shared public credit pool and top-up pack design indicate that the vendor expects "agent usage to grow with team maturity", which is consistent with this group's industry judgment in the README of a "shift from seat-based to consumption-based billing".

2. Glossary

3. Feature Description

3.1. Plugin Form

The plugin is the form factor with which CodeBuddy started. Core capabilities:

Craft's context system: @Add can select files / folders, file changes, the knowledge base, and the last command executed in the terminal as conversation context; conversations support exporting, editing, and deleting history. Craft supports three hosts: VS Code, JetBrains IDEs, and WeChat DevTools.

3.2. CodeBuddy IDE

The standalone IDE (based on the VS Code technology stack) targets the full "from idea to deployment" flow. A typical capability chain:

1. Design and requirements phase: conversational requirements Q&A and decomposition; inputting a target description auto-generates specification documents such as PRDs; generates UI design mockups from natural language; importing a Figma design auto-generates page code (DOM-level precise debugging and real-time sync).
2. Development and coding phase: the Craft agent performs multi-file generation and rewriting; preset component libraries (TDesign, MUI, Shadcn) with natural-language adjustment of style and layout; integrates BaaS services such as CloudBase and Supabase to automatically complete database and user authentication setup.
3. Deployment and validation phase: one-click deployment to a sandbox environment via CloudStudio, generating a shareable link; can also publish to EdgeOne Pages.

The IDE and CLI share an account and quota system; during the international public beta, the IDE and CLI share credits.

3.3. CodeBuddy Code CLI

CodeBuddy Code is an AI CLI tool for professional engineers (released 2025-09-09), installed with npm install -g @tencent-ai/codebuddy-code (requires Node.js 18.20+; a native installer beta that does not require Node.js is also available), on macOS / Linux / Windows. Key capabilities:

There are four login paths: the China site (copilot.tencent.com, domestic models), the international site (codebuddy.ai, international models), enterprise-dedicated / private deployment (connecting a self-hosted enterprise service address), and iOA (Tencent's internal employee zero-trust login). This design is a direct embodiment of "serving different compliance domains with the same Harness".

4. Platform Architecture

Fig. 4-1 | CodeBuddy Platform Overall Architecture (Full-Form Coverage × Dual-Track Models × Tencent Cloud Ecosystem)

Diagram note: drawn based on the platform architecture section of this article.

4.1. Overall Architecture

4.2. Model and Ecosystem Integration

CodeBuddy's model strategy is a "dual-track" approach: the domestic track is dominated by DeepSeek and Hunyuan (early hands-on testing of the domestic CLI shows DeepSeek-V3.1 and DeepSeek-V3 tiers), while the international track offers tiers such as the GPT-5 series, the Gemini 2.5 series, and Claude Sonnet 4, and supports /model switching and enterprise custom model integration (an enterprise can configure internally authorized models or APIs, with keys managed through environment variables or an enterprise configuration center rather than stored in plaintext).

Ecosystem integration is its structural difference relative to other standalone IDEs: AI capabilities are embedded in Tencent Cloud's R&D pipeline — designs come from Figma, backends land on CloudBase / Supabase, deployment goes through CloudStudio / EdgeOne, R&D collaboration connects with the CODING system, and WeChat DevTools covers the Mini Program ecosystem. For R&D organizations already using Tencent Cloud, this constitutes an integration path that requires "no infrastructure change"; for non-Tencent-Cloud users, the value of these integrations is limited.

5. Harness Design

5.1. L1 Context Engineering Layer

CodeBuddy's context engineering consists of three mechanisms:

Project guide file (CODEBUDDY.md). The /init command scans the codebase and auto-generates a project guide covering project structure, technology stack, and working conventions. The official positioning of /init reveals an important design judgment: prebuilding a project knowledge graph is better than rescanning on every session — it both improves comprehension accuracy and saves the token overhead of repeated scans (the official side claims a 30%–50% reduction in context token overhead). This line of thinking shares the same origin as the rule-file logic of Cursor / Claude Code in this group's earlier documents, but elevates "generating a memory file" to a strongly recommended initialization step.

Rules directory (.codebuddy/rules). Rule files maintained within the project (e.g. .codebuddy/rules/backend.md) pin down team standards such as framework versions, layering conventions, documentation update requirements, and sensitive-information bans. Rules are referenced during generation so output conforms to team standards. This belongs to the same lineage as the .cursor/rules and .claude/rules already analyzed by this group; CodeBuddy has not disclosed fine-grained trigger semantics like Cursor's alwaysApply/globs, so its granularity is coarser.

Multi-layer RAG and knowledge base. Enterprise materials disclose three types of retrieval: Infra RAG (infrastructure and runtime environment knowledge), CodeBase RAG (repository-level code retrieval), and Architect RAG (architecture standards and project constraint retrieval); on the user side, @Add injects knowledge base documents, file changes, and terminal output into the conversation context. The knowledge base mechanism is CodeBuddy's main channel for enterprise knowledge integration — in-house standards documents need not be rewritten as rule files; they can be registered directly into the knowledge base to participate in retrieval.

Summary: L1's maturity sits in the "complete mechanisms but coarse disclosure granularity" tier. Details such as rule loading priority, compression trigger conditions, and context budget allocation have not been disclosed at the documentation level and are marked as information gaps.

5.2. L2 Tool and Execution Layer

• Built-in tools: file editing, command execution, code retrieval; inside the CLI, a ! prefix enters Bash mode to directly execute shell commands.
• MCP: the first code assistant in China to support the MCP protocol (official claim); on the CLI side codebuddy mcp list/add manages servers, and a project-level .mcp.json can be shared with the repository. Early versions supported only stdio transport and did not support HTTP/SSE remote MCP (, early hands-on testing).
• Browser and preview: the IDE includes real-time preview, supporting DOM-level debugging and real-time sync from design to code.
• Deployment toolchain: one-click deployment to CloudStudio / EdgeOne Pages; BaaS integration automatically completes database and authentication setup.
• Sandbox: mainly local execution; no official disclosure of an OS-level sandbox (similar to Seatbelt / Landlock) was found, marked as a gap. The official stance on --dangerously-skip-permissions is "only recommended for sandboxes without network access", i.e. it delegates the sandboxing responsibility to the user.

5.3. L3 Orchestration and Control Layer

Craft's three-toggle model is the core of CodeBuddy's orchestration design: generate a plan (Plan mode, first decomposes the task and clarifies requirements), auto-run (automatically executes tasks per the plan), and auto-modify files (authorizes file generation and rewriting). The three toggles are independently configurable and on by default — users can combine tiers such as "plan only, no action", "plan and execute but do not modify files", and "fully automatic". Compared to Codex CLI's two-dimensional sandbox × approval model, this is coarser-grained but covers the same intent: separating the authorization of the three things "plan — execute — write".

On the CLI side this corresponds to --permission-mode (default / acceptEdits / bypassPermissions / plan). Early versions did not support subagent configuration (/agents) or hooks (/hooks), the latter already becoming visible in later version command lists (). On the IDE side, task decomposition with step-by-step execution and design mode (exporting code from design mockups) constitute lightweight orchestration aimed at non-professional users.

5.4. L4 Memory and State Layer

• Global memory: ~/.codebuddy/CODEBUDDY.md stores personal long-term preferences; the project-root CODEBUDDY.md stores team-shared conventions.
• Session resumption: --continue / --resume restore historical sessions; /compact compresses history while keeping a summary.
• Historical conversation management: the plugin and IDE support viewing, exporting, editing, and deleting historical sessions.
• Config layering: settings.json is split into three layers — global (~/.codebuddy/settings.json), project (.codebuddy/settings.json, can be committed and shared), and local (.codebuddy/settings.local.json, not committed) — with priority local > project > user.

Gap: no official disclosure was found of a workspace snapshot / checkpoint rollback mechanism, marked "no result". Compared with Cline's shadow Git checkpoints and Windsurf's named checkpoints, this is a clear gap in L4; current rollback mainly relies on the user's own Git discipline.

5.5. L5 Evaluation and Observability Layer

• Cost observability: the CLI /cost shows the current session's token consumption and duration; /status shows model and connection status; the international site provides credit consumption queries.
• Code review: /review reviews changes and /pr-comments brings in GitHub PR comment context; the plugin has a built-in code review agent and commit-branch code scanning (generating scan reports and fixing vulnerabilities).
• Organization-level metrics: CodeBuddy's distinctive trait is extending L5 from the tool layer to the operations layer. The China Merchants Bank case discloses a "three-layer structure" methodology — model layer (choose well, determines the product ceiling), plugin layer (use well, multi-endpoint adaptation), operations layer (iterate well, build a metrics system, and continuously evolve the model through operational data and Badcase fixes). Its metrics system includes completion adoption rate, AI code generation share, and defect density, supported by internal promotion and incentive mechanisms.

Gap: no public trace tracking, regression eval set, or mechanized benchmark evaluation was found; all official quantitative data is operational-statistics-based, marked as information gaps.

5.6. L6 Governance and Security Layer

L6 is where CodeBuddy shows its greatest structural difference relative to independent AI IDEs:

The counterpart at the methodology layer is enterprise rule integration: through .codebuddy/rules and the knowledge base, compliance red lines such as "sensitive information should not be printed directly in business code" are solidified into generation constraints. Compared with Cursor's MDM-level Hooks and Claude Code's managed policy files, CodeBuddy's mechanical enforcement is weaker (no disclosed hook-interception system), but its deployment-form compliance capability (private deployment, VPC, identity source integration) is among the most complete tiers of any platform this group has analyzed.

Against this group's README security baseline (filesystem isolation and network isolation are both indispensable, credentials are hidden by default, and destructive commands are hard-blocked), only the allowlist mechanism in CodeBuddy's public materials can partially assume the third responsibility; the first two rely on the deployment form as a safety net.

5.7. Six-Layer Capability Summary

6. Case Studies

6.1. Tencent Internal Large-Scale Practice

The following data comes from Tencent Global Digital Ecosystem Conference materials (reposted via Tencent Cloud Developer Community articles; a third-party repost of official figures, marked overall as [To be verified]):

6.2. Onewo (Public Cloud SaaS)

Background: limited budget, needing assessable efficiency gains in real business scenarios. Approach: public cloud out-of-the-box, quick promotion on a 10-minute hands-on basis, a two-week deep evaluation of the flagship edition (producing Goodcase / Badcase reports), and promotion with recognition-based incentives modeled on Tencent's internal practice. Results (official figures): 300+ seats licensed, 240+ daily active users (80% adoption); AI code generation share over 35% across the board, adoption rate ~30%; coding time shortened ~50% for the same workload; 85% positive user reviews and 93% agreeing that efficiency was significantly improved.

6.3. China Merchants Bank (Private Deployment)

Background: its self-developed IDE plugin faced four difficulties — many plugin types, complex versions, non-converging model fine-tuning, and a lack of a metrics system. Approach: established a co-innovation project team for the AI code assistant with Tencent Cloud, collaborating deeply across plugin development, model evaluation, model training, and operations promotion, building a three-layer closed loop of "model layer (choose well) — plugin layer (use well) — operations layer (iterate well)". Results (official figures): 8,000+ internal users; code completion adoption rate of 28%; AI code generation share 35%+; and the enterprise's own capability to continuously train and optimize models.

Methodological value of the cases: the CMB (China Merchants Bank) case is the only one in this group to date that fully discloses an "enterprise-side Harness operations system" — it shows that the bottleneck for AI coding adoption in regulated industries is not model selection but the building of a closed loop of metric measurement and Badcase feedback. This corroborates this group's README thesis #3 (self-assessment cannot serve as an acceptance basis): the operations-layer metrics system is precisely the engineering means of replacing subjective feel with comparable measurement.

It should also be noted that the Chuang Meng Tiandi case (gaming industry; 75% of engineers using it, per-person defect density down ~5%, AI-assisted generation rate 30%, efficiency up 12%+) is likewise from official materials. This group found no independent third-party quantitative evaluation of CodeBuddy.

7. Summary

Strengths:

1. Full-form coverage with a unified account system: three ends (plugin, IDE, CLI) share models and quotas; it is the only vendor in China providing all three forms simultaneously (official claim), and organizations can assign forms by role rather than replacing tools across the board.
2. Complete L6 deployment-form compliance capability: private deployment, dedicated VPC, enterprise identity source with iOA zero-trust integration, domestic/international domain separation, and a replaceable model track make it one of the few AI coding products that can enter procurement in highly regulated industries.
3. Clear enterprise knowledge integration channel: the knowledge base + three types of RAG inject in-house standards into context, and combined with an operations-layer metrics system form a replicable adoption methodology (the CMB case).
4. Domestic model track has no network barriers: the DeepSeek and Hunyuan dual-engine avoids cross-border network and data-export issues.

Limitations:

1. Low disclosure of Harness details: rule trigger semantics, context budget, compression strategy, and sandbox mechanism all lack documentation-level disclosure, and no public implementation of mechanized enforcement (hook interception, fail-closed) was found.
2. L4 gap: no checkpoint rollback mechanism is disclosed; recovery from mid-task errors in long tasks relies on the user's Git discipline.
3. All quantitative data is official: figures such as internal adoption rate and Bug rate reduction have no independent third-party verification.
4. Ecosystem value is tied to Tencent Cloud: for non-Tencent-Cloud users, the marginal value of deployment, BaaS, and R&D collaboration integrations is limited.

Applicability boundary and selection advice: For R&D organizations already using Tencent Cloud, or those that must privately deploy due to data compliance constraints, CodeBuddy is an unavoidable option in the domestic market; its plugin form factor can serve as a low-friction completion and review layer over existing IDEs, and the CLI suits CI and batch tasks. Individual developers pursuing editor-level agent experiences or early availability of frontier models will still find Cursor, Claude Code, and Codex CLI more mature choices. Teams that need the domestic model track but prefer open, auditable implementations can evaluate against this group's paper #09 on Cline (BYOK integration of domestic models, Apache 2.0 auditable). Whichever path is chosen, the ordering the CMB case prescribes — "build measurement first, then grant autonomy" — should be the execution prerequisite for enterprise adoption.

Information Gap Statement

1. Current version number: The latest stable version numbers for the plugin, IDE, and CLI are not published on the official search page; marked [To be filled].
2. Pricing discrepancy: The enterprise edition at ¥158 / seat / month (minimum 10 seats, dedicated VPC) is a third-party aggregated figure not verified on the official pricing page; marked [To be verified]. The end date of the domestic individual free trial has not been disclosed.
3. Early CLI capability limits: Statements that /agents, /hooks, and MCP over HTTP/SSE are unsupported come from third-party testing around 2025-09 (v1.1.0); whether the current version has closed these gaps has not been verified item by item, marked [To be verified].
4. MCP Server landscape: The official MCP Server list (Design / Security / CODING / T-Sec, etc.) comes from enterprise conference materials; the public documentation does not describe each one in detail, marked [To be verified].
5. Quantitative case data: All effectiveness figures for Tencent internal use, Onewo, Chuang Meng Tiandi, and China Merchants Bank are from official materials (third-party reposts) with no independent third-party verification; the whole set is marked [To be verified].
6. Sandbox and hooks: No official documentation was found for OS-level sandboxing, lifecycle hooks, or audit log export; marked "no result".
7. WorkBuddy naming: No public product page from Tencent was found for the name "WorkBuddy" mentioned in the task instructions; this piece is written according to the official on-sale brand CodeBuddy. If WorkBuddy is an internal or other product-line name, the relevant content remains to be supplemented.

8. References

1. CodeBuddy official site (domestic version) — Tencent Cloud, 2026. https://www.codebuddy.cn/
2. CodeBuddy international site pricing details — Tencent Cloud International, 2026. https://intl.cloud.tencent.com/zh/document/product/1256/77270
3. Craft agent usage documentation — Tencent Cloud CodeBuddy official documentation, 2025. https://www.codebuddy.cn/docs/plugin/%E6%93%8D%E4%BD%9C%E6%8C%87%E5%8D%97/Craft%20%E6%99%BA%E8%83%BD%E4%BD%93/Craft%20%E7%9A%84%E4%BD%BF%E7%94%A8
4. CodeBuddy Code CLI Quick Start Guide — CodeBuddy official documentation, 2025—2026. https://www.codebuddy.ai/docs/cli/quickstart
5. Tencent Cloud CodeBuddy: overview of the enterprise-grade software engineering AI development efficiency and full-link solution — Tencent Cloud Developer Community (repost of Tencent Global Digital Ecosystem Conference materials), 2026. https://cloud.tencent.com/developer/article/2677065
6. CodeBuddy Code CLI installation, configuration, and hands-on testing — Juejin community, 2025-09. https://juejin.cn/post/7572714389943500852
7. Tencent releases CodeBuddy Code CLI and IDE international beta — Hong Kong programming society (repost of IT Home coverage), 2025-09-09. https://www.hkprog.org/?p=3993/
8. CodeBuddy installation and configuration tutorial (VS Code / IDEA / CLI) — AI Stack Nav, 2026. https://en.aistacknav.com/?p=3479
9. @tencent-ai/codebuddy-code — npm, 2025—2026. https://www.npmjs.com/package/@tencent-ai/codebuddy-code
10. Tencent Cloud CodeBuddy international product page — Tencent Cloud, 2025—2026. https://www.codebuddy.ai/