ADR-0000094: Approval of Anthropic Claude Code for AI Coding Agents Whitelist¶
Bidirectional Sync Enabled
This page is synced with GitLab repository source (adrs/adr-94/adr-94-approval-of-claude-code-for-ai-coding-whitelist.md). You can edit this page directly in Confluence - changes will be automatically pulled back to the repository during the next sync. For sync details and configuration, see the README in the repository. Last sync: 2026-05-22 13:58:33
- Type: Technology Decision
- Status: READY FOR REVIEW
- Dependency: ADR-0000065: Decision of use of AI Coding Assistants
- Derived from: ADR-0000065
- Scope: ENTERPRISE
- Owning architect: @Pavel Bach
1.1 Context and Problem Statement¶
1.1.1 What triggered this ADR?¶
- ADR-0000065 established the governance framework for AI Coding Assistants at company, selecting Option 3: Tactical usage of chosen AI coding assistants. This framework mandates a whitelist of approved AI coding tools maintained under the AI Coding Agents whitelist Confluence page.
- Currently, only Anthropic Claude Sonnet (under commercial/Pro plan terms) is an approved and evaluated entry on the whitelist. The whitelist page lists GitHub Copilot (as "Visual studio copilot") as a candidate that has not yet been formally evaluated.
- Engineers have requested access to additional AI coding tools - specifically GitHub Copilot - due to its native IDE integration, inline code completion, and multi-IDE support that go beyond what the current Claude tooling (browser chat and Claude Code CLI) can offer. The market for AI coding assistants has matured significantly since ADR-0000065, with multiple enterprise-grade tools now available: GitHub Copilot, Claude Code, JetBrains AI Assistant, Amazon Q Developer, Google Gemini Code Assist, and others.
- Without a formal evaluation and decision, engineers cannot use any of these tools, blocking potential productivity gains - especially for inline code completion, multi-file refactoring, and agentic coding workflows.
1.1.2 Who is affected¶
- R&D teams - direct users of the tool
- CTO/CISO / CEO - risk and compliance owners
- IT / Service Desk - provisioning and account management
- Legal / Compliance - contractual and regulatory oversight
1.1.3 Considered scenarios¶
- A developer uses an approved AI coding assistant in their IDE (VS Code or JetBrains) for inline code completion, refactoring, and documentation generation.
- A developer uses the AI assistant's chat panel within the IDE to ask questions about the codebase and generate code snippets.
- A developer uses the AI assistant in agent mode to drive larger changes from natural-language intent ("vibe coding") - the agent autonomously edits multiple files, runs builds and tests, and iterates until the change is green, with the developer steering and reviewing rather than typing the code line by line.
- Generated code is reviewed via merge request and scanned with ScanCode Toolkit before merging to the release branch.
- The AI assistant is used only on developer machines - no CI/CD integration, no autonomous cloud agents.
1.1.4 Out of scope¶
- Autonomous cloud agents - features like GitHub Copilot Cloud Agent or similar autonomous code generation are excluded from this approval.
- CI/CD integration - the approved tool must not be integrated into GitLab pipelines or CI/CD automation.
- Non-coding tasks - usage for customer communication, contract review, or marketing is not covered.
- Replacing Claude - this ADR does not replace or reduce Claude's whitelist status; both can coexist. Anthropic Claude is already approved as a reasoning/chat and agentic coding tool (including Claude Code CLI, which can be used in the terminal and VS Code) and is increasingly available as a model inside other IDE tools (e.g., GitHub Copilot can use Claude as its underlying model).
- Standalone AI IDEs - tools like Cursor and Windsurf that require abandoning existing IDE investments are excluded. company engineers use VS Code and JetBrains IntelliJ; the selected tool must integrate with these existing IDEs.
- Any future advanced integration (e.g., MCP servers, browser extensions) requires a separate ADR.
1.1.5 Problem Statement¶
Which AI coding assistant should be added to the AI Coding Agents whitelist under the governance framework established by ADR-0000065, to provide company engineers with IDE-integrated code completion, refactoring assistance, and agentic coding capabilities for a pilot group (<=10 engineers)?
1.2 Market Overview - AI Coding Assistants¶
The following market overview evaluates the enterprise-grade AI coding assistants available as of May 2026. Standalone AI IDEs (Cursor, Windsurf) are excluded because they require abandoning existing IDE investments. Anthropic Claude (including Claude Code) is already approved on the AI Coding Agents whitelist managed under ADR-0000065, covering browser chat, terminal, and VS Code usage. Claude Code is nevertheless evaluated as a separate option below to provide a complete comparison against the other tools. A self-hosted option built from publicly available open-weight models is also evaluated (Option 6), because company's data-sovereignty profile and OVH Cloud migration make on-premises inference a strategically relevant alternative.
Market evolution as of May 2026 (all prices below verified against vendor pages on 2026-05-22). The field has moved fast since ADR-0000065 and the pricing in earlier drafts of this ADR is now stale.
The most consequential changes: 1. GitHub Copilot is moving from "premium request" billing to usage-based "AI Credits" billing on 1 June 2026, and since 24 April 2026 the individual plans (Free / Pro / Pro+) train on user interaction data by default - only the Business and Enterprise plans remain excluded, which reinforces this ADR's Business/Enterprise-only constraint. 2. Amazon Q Developer now runs on Anthropic Claude (Sonnet 4.6) via Bedrock - it is no longer locked to AWS-proprietary models. 3. JetBrains AI rebranded, made BYOK generally available (OpenAI/Anthropic plus local models via Ollama and LM Studio) and now offers fully air-gapped on-premises AI Enterprise via IDE Services 2026.0. 4. Google Gemini Code Assist is sunsetting its individual free tier and the Gemini CLI on 18 June 2026, pushing individuals to the free agent-first Antigravity IDE (a VS Code fork); Gemini 3.5 and Antigravity 2.0 were announced at Google I/O 2026. 5. Open-weight code models (Qwen3-Coder, GLM-4.7, DeepSeek-V3.2, Devstral) have matured to the point that a self-hosted, on-premises coding assistant is now a credible option, evaluated below.
These changes are reflected in the tables below. A notable cross-cutting consequence: usage-based, token-metered pricing is now the norm across GitHub Copilot (AI Credits), Anthropic Claude (Team usage limits) and JetBrains AI (credits), so "metered usage beyond an included allotment" is a shared industry reality managed via spend controls - not a differentiator between these tools. The differentiator is the predictability of the entry tier and whether the best model is consumed directly from its maker or through a reseller (see Decision Outcome).
1.2.1 Pricing Comparison¶
Prices verified against the vendors' public pricing pages on 2026-05-22. Annual-commit prices are shown where the vendor offers a discount over monthly billing. Where a vendor's pricing is mid-transition (notably GitHub Copilot) or being retired (Gemini individual tier), this is noted.
| Tool | Business / Team Tier | Enterprise Tier | Notes |
|---|---|---|---|
| GitHub Copilot | $19/user/month (Business) | $39/user/month (Enterprise) | Premium requests included (300/mo Business, 1000/mo Enterprise). Moving to usage-based "AI Credits" billing on 1 June 2026 - each paid plan includes a credit balance equal to its fee; code completions stay free; overage metered. Individual plans (Pro $10/mo, Pro+ $39/mo) train on user data by default since 24 Apr 2026 and are NOT approved for company. |
| JetBrains AI | AI Pro ~$10/user/mo (individual)/~$20 (business); AI Ultimate $30/$60 | AI Enterprise - custom (contact sales), per-seat quota "on par with Ultimate or higher" | $-denominated AI-credit model; code completion and context chat are unlimited even on Free. IDE licence sold separately (AI Pro bundled with All Products Pack). BYOK GA (OpenAI/Anthropic/Ollama/LM Studio); AI Enterprise supports air-gapped on-premises via IDE Services 2026.0 (ships JetBrains Mellum for completion). |
| Amazon Q Developer | Free tier (perpetual): 50 agentic requests/mo | $19/user/month (Pro) | Now runs on Anthropic Claude (Sonnet 4.6) via Bedrock no longer AWS-proprietary-only. $0.003/line code-transformation overage. Pro content excluded from training; Free content trains unless opted out. |
| Claude (Anthropic) | Team Standard $20/seat/mo annual ($25 monthly, 5-seat min.); Premium $100/seat/mo annual ($125 monthly) | Enterprise $20/seat/mo annual + API usage pay-as-you-go (20-seat min. self-serve, 50-seat sales-assisted) | Claude Code is included on every Team seat (Standard and Premium) and on Enterprise; SSO/SCIM/audit logs at Enterprise. Models: Opus 4.7, Sonnet 4.6, Haiku 4.5. Consumer plans (Pro $100/mo, Max from $100/mo) are out of scope for company use. |
| Google Gemini Code Assist | Standard ~$19/user/mo annual (~$22.80 monthly) | Enterprise ~$45/user/mo annual (~$54 monthly) | Gemini 3 family, 1M-token context window (Gemini 3.5 announced at I/O 2026). Agent Mode + MCP GA. Free individual tier and Gemini CLI are sunset on 18 June 2026 (individuals migrate to the free Antigravity IDE). |
| Self-hosted (open models) | Laptop variant ~$2.5-4K one-time per machine (devs need laptops anyway); Azure-hosted = metered GPU opex (scale-to-zero) | On-prem server ~$35-45K (single H100-80 GB); $80K+ for multi-GPU frontier serving | Apache-2.0/ MIT open-weight models (Qwen3-Coder, DeepSeek-V3.2, GLM-4.x, Devstral, Granite, StarCoder2). No per-seat/token fees; data stays on company-controlled hardware (or its EU Azure tenant). Three deployment variants - see Option 6. |
For a 10-engineer pilot, indicative monthly cost: Copilot Business ~$190, Copilot Enterprise ~$390, Claude Team Standard ~$200 (annual) / ~$250 (monthly), Claude Team Premium ~$1,000, Claude Enterprise ~$200 + uncapped API usage (note: 20-seat minimum, so a 10-seat pilot would use Team, not Enterprise), JetBrains AI Pro ~$200 / AI Ultimate ~$600, Amazon Q Pro ~$190, Gemini Standard ~$190. Self-hosted: ~$0/month in licences against a one-time ~$10-45K hardware spend plus recurring ML-ops effort.
1.2.2 IDE Support¶
| Tool | VS Code | JetBrains IDEs | Visual Studio | Other IDEs |
|---|---|---|---|---|
| GitHub Copilot | Native first-party (same company) | Plugin | Plugin | Eclipse, Xcode, Neovim |
| JetBrains AI | CLI only (Junie CLI; no native plugin) | Native first-party | No | Android Studio |
| Amazon Q Developer | Extension | Extension | Extension | Eclipse, CLI |
| Claude Code | Extension | Plugin (Beta, 2026) | No | Terminal CLI (primary interface) |
| Google Gemini Code Assist | Extension; Antigravity (VS Code fork) | Extension | No | Cloud Shell, Firebase, Antigravity CLI |
| Self-hosted (open models) | Continue.dev, Tabby, Cline/Roo Code | Continue.dev, Tabby | No | Aider, opencode, Mistral Vibe CLI (terminal) |
company engineers use both VS Code and JetBrains IntelliJ. GitHub Copilot and JetBrains AI each have first-party native support in one of these IDEs and plugin support in the other - note that JetBrains AI's VS Code reach is now CLI-only (Junie CLI), there is no native VS Code plugin. Claude Code supports both VS Code (extension) and JetBrains (Beta plugin), but neither is a first-party native integration. For the self-hosted option, the open-source front-ends Continue.dev and Tabby cover both VS Code and JetBrains with inline completion plus chat, connecting to any OpenAI-compatible endpoint.
1.2.3 IDE Integration Depth¶
The term "IDE integration" is often used loosely. The following table breaks down what it means concretely for each tool, distinguishing between superficial plugin support and deep native integration.
| Capability | GitHub Copilot | Claude Code | JetBrains AI | Amazon Q | Gemini Code Assist | Self-hosted (open models) |
|---|---|---|---|---|---|---|
| Inline code completion (as-you-type suggestions) | Yes ghost text, multi-line, context-aware | No | Yes ghost text, multi-line, context-aware | Yes | Yes | Yes - via Continue.dev / Tabby (Qwen2.5-Coder, StarCoder2) |
| Next Edit Suggestions (predicts where to edit next) | Yes (GA in VS Code; preview in JetBrains) | No | Yes | No | No | No |
| Chat panel in IDE | Native panel with @ commands | Extension panel (VS Code; JetBrains Beta) | Native panel | Extension panel | Extension panel | Continue.dev / Cline / Roo (VS Code + JetBrains) |
| Agent mode (autonomous multi-file editing, runs terminal, auto-fixes) | Yes native in VS Code and JetBrains | Yes - terminal-first, strong autonomous capabilities | Yes Junie agent (GA); Junie CLI (Beta) (SWE-bench leader) | Yes - agentic mode | Yes GA on Gemini 3; Antigravity | Yes Cline / Roo / Aider (quality depends on model) |
| Multi-file edit mode | Copilot Edits in VS Code, JetBrains, Visual Studio | Yes-strong agentic multi-file editing | Yes - via Junie | Limited | Limited | Yes - Cline / Roo / Aider |
| Terminal integration (AI-assisted command suggestions) | Yes explain errors, suggest commands | Primary interface - Claude Code is terminal-first | Junie CLI | Separate Q CLI | Separate Gemini / Antigravity CLI | Aider / opencode / Mistral Vibe CLI |
| Code review/PR integration | AI code review + PR summaries on GitHub PRs (GitHub only; core GA, some sub features preview) | /review, gh / glab CLI or Git MCP; first-party GitHub Action (GA, v1); GitLab CI integration (beta, GitLab maintained) | No (IDE-only) | GitHub / GitLab integration (preview) | No | Via gh / glab CLI or MCP (Aider / opencode); no managed service |
| Custom instructions (repo/org level configuration) | Repo, personal, org, and path specific .github/copilot-instructions.md | CLAUDE.md (+ AGENTS.md) at enterprise/managed, user, project, and directory levels (+ skills, subagents, hooks) | Limited (prompt templates) | Limited | No | Continue / Cline rules files (repo level) |
| Model choice | 15+ models: OpenAI GPT-5.x, Anthropic Claude (Opus 4.x/Sonnet 4.5/4.6, Haiku 4.5), Google Gemini 3.x | Claude only (Opus 4.7, Sonnet 4.6, Haiku 4.5); routable via Bedrock / Vertex / Foundry | OpenAI, Anthropic, Google, xAI + BYOK + local models | Anthropic Claude (Sonnet 4.6) via Bedrock | Google Gemini 3 family only | Any open weight model (Qwen, DeepSeek, GLM, Devstral, Granite, Llama) |
| MCP (Model Context Protocol) support | Yes - including admin managed MCP servers / org registry | Yes - MCP client and server (claude mcp serve); managed allow/deny | Yes (GA) | Yes (CLI + IDE) | Yes (GA) | Yes - Continue / Cline / Roo support MCP |
| Admin policy management | Org + Enterprise policies, audit logs, content exclusion | Managed settings policy (MDM/GPO/Ansible), permission allow/deny, SSO/SCIM (Enterprise), managed MCP, OpenTelemetry, audit/compliance API | SSO, SCIM, AI audit logs (Enterprise) | AWS IAM based | Google Cloud IAM | Fully self managed (own deployment, network, logging) |
| Anthropic Claude inside Copilot | Selectable Claude models + a GitHub-built Claude agent (Claude Agent SDK, public preview; github.com & VS Code only) | N/A - Claude Code is a separate Anthropic product, not runnable inside Copilot | No | No | No | No |
Key insight: "Deep IDE integration" means the AI assistant is not just a sidebar chat panel, but is woven into the entire development workflow: inline completions while typing, predicting the next edit location, running commands in the terminal, reviewing pull requests, and operating autonomously across multiple files. GitHub Copilot provides the broadest set of these capabilities across both VS Code (native) and JetBrains (plugin). JetBrains AI provides comparable depth but only natively in JetBrains IDEs (its VS Code reach is now CLI-only). Claude Code covers both VS Code and JetBrains via plugins (JetBrains still Beta) and excels at agentic and terminal workflows, but still lacks inline completions. The self-hosted stack (Continue.dev / Tabby for completion + Cline/Roo for agent) reaches both IDEs and is the only option whose data never leaves company infrastructure, at the cost of self-managed operations and a model quality ceiling.
1.2.4 Data Privacy and Training Guarantees¶
All commercial enterprise-tier tools provide contractual guarantees that customer code is not used for training, so data privacy is not a differentiator among them at the enterprise tier. The two notable caveats are that GitHub Copilot's individual plans now train on user data by default (reinforcing the Business/Enterprise-only constraint) and that Amazon Q's Free tier trains unless opted out. The self-hosted option is categorically stronger: because inference runs on company-owned infrastructure, no source code or prompt ever reaches a third party - the decisive advantage for data sovereignty.
1.3 Decision Outcome¶
Approve Anthropic Claude Code (Option 2) for inclusion in the AI Coding Agents whitelist.
Claude Code is selected as the AI coding assistant to add to the whitelist. The decision rests on a fair, fact-checked comparison (see Decision Drivers): Claude Code matches GitHub Copilot on 9 of 11 quality attributes with zero negatives, and the two attributes where Copilot still leads - inline ghost text completions and a non-Beta JetBrains plugin - are addressed via a strategic action item below.
1.3.1 Why Claude Code?¶
- Direct access to the leading model: Claude Code runs Anthropic's native Claude models (Opus 4.7, Sonnet 4.6), which are the top-ranked models for agentic coding (Sonnet 4.6 leads SWE-bench Verified). GitHub Copilot owns no frontier model; it brokers OpenAI, Anthropic and Google models. Choosing Claude Code goes direct to the model maker rather than through a middleman, and Claude Code is co-designed with the models it runs.
- Direct connection and cost clarity: Direct access to the best model, with a predictable entry tier - all hosted tools meter heavy usage beyond an included allotment (see Rational Cost), so that is not the differentiator. What favours Claude Code for using the leading model is twofold: its entry tier (Team Standard, ~$200/mo for 10, comparable to Copilot Business ~$190/mo) includes generous Claude usage, and you pay Anthropic directly rather than consuming Claude Opus through a reseller. From 1 June 2026 Copilot's most-advanced models (Claude Opus, GPT-5.x, Gemini Pro) draw down "AI Credits" at their token rate on top of the seat fee - paying a broker to reach the same Anthropic model.
- Best-in-class agentic editing + already approved: Claude Code is the strongest option for autonomous multi-file editing (plan mode, subagents, background tasks), terminal-first, with a native VS Code panel and rich enterprise governance. It is already available to company engineers under the existing Claude whitelist approval (managed under ADR-0000065).
Approved under the same constraints and governance model established by ADR-0000065: * Account type: Claude Team (Standard or Premium) or Enterprise (mandatory - consumer Pro/Max plans are NOT approved for company source code) * Users: Limited pilot group (<=10 engineers), access granted via Service Desk request with CTO approval * Usage: Terminal and IDE-only (CLI, VS Code, JetBrains); no CI/CD pipelines, no autonomous cloud agents * Code scanning: Mandatory ScanCode Toolkit scan of AI-generated code before merge to release branch, results attached to Jira ticket * Review: Human peer review via merge request required before merge * Sandbox: Minimal permissions enforced; permission allow/deny rules and managed-settings policy applied; no access to production systems or secrets
1.3.2 Action Items¶
| Area | Action | Link |
|---|---|---|
| IT / Service Desk | Create provisioning workflow for Claude Team (or Enterprise) seats; assign to the pilot group | |
| Compliance | Update AI Coding Agents whitelist Confluence page to include Anthropic Claude Code with evaluation results | AI Coding Agents whitelist |
| CISO / IT | Apply managed-settings policy (permission allow/deny, managed MCP, content/secret exclusions); enable OpenTelemetry usage/cost export | |
| R&D | Communicate approved usage guidelines and constraints to pilot group | |
| CISO | Confirm Anthropic Commercial Terms / DPA (eff. 1 Jan 2026) and ZDR (if required) alignment with company data handling policy | |
| CTO / CFO | Set Claude Team/Enterprise spend controls and monitor metered usage beyond plan limits during the pilot | |
| Architecture / R&D | Run the self-hosted developer laptop PoC (Apache-2.0/MIT models, Continue.dev / Tabby) to cover Claude Code's inline completion gap and trial full data sovereignty; later compare an on-prem GPU server vs an OVH-hosted deployment (preferred over Azure for cloud alignment) | |
| R&D | Periodic review after pilot phase - evaluate productivity impact and risk incidents |
1.4 Decision Drivers¶
Quality attributes and stakeholders are aligned with the governance framework from ADR-0000065.
Options Marked: * decided for: ✅ * not decided for: ❌
| Quality Attribute | Stakeholder | Context / Decision Driver | AS-IS ❌ | O1: Copilot ❌ | O2: Claude Code ✅ | O3: JetBrains AI ❌ | O4: Amazon Q ❌ | O5: Gemini ❌ | O6: Self hosted ❌ | O7: None ❌ |
|---|---|---|---|---|---|---|---|---|---|---|
| Risk (Input - Confidentiality) | CISO, CTO | Source code must not be used for model training or permanently stored by the provider | ☀️ | ☀️ | ☀️ | ☀️ | ☀️ | ☀️ | ☀️ | ☀️ |
| Risk (Output - License Contamination) | CEO, CISO, R&D | AI generated code must not contain copyleft/copyrighted material | neutral | neutral | neutral | neutral | neutral | neutral | neutral | ☀️ |
| IDE Integration Depth | CTO, R&D | Tool must provide inline completions, agent mode, and terminal integration across both IDEs used at company | neutral | ☀️ | neutral | neutral | neutral | neutral | neutral | ️ |
| Model Flexibility | CTO, R&D | Tool should support multiple AI models, including already approved Claude | neutral | ☀️ | neutral | ☀️ | neutral | neutral | ☀️ | ️ |
| Efficiency of organization | CTO, R&D | Measurable productivity gains for coding, refactoring, and documentation tasks | neutral | ☀️ | ☀️ | ☀️ | ☀️ | ☀️ | ☀️ | neutral |
| Compliance | CEO, CISO | Enterprise Usage must comply with EU AI Act, GDPR, ISO/IEC 42001 requirements | ☀️ | ☀️ | ☀️ | ☀️ | ☀️ | ☀️ | ☀️ | ☀️ |
| Cost | CTO, CFO | Tool cost must be within acceptable limits for the pilot scope (<=10 users) | ☀️ | ☀️ | ☀️ | neutral | ☀️ | ☀️ | neutral | ☀️ |
| Cloud Vendor Neutrality | CTO | Tool must not create dependency on a specific cloud provider | ☀️ | ☀️ | ☀️ | ☀️ | ️ | ️ | ☀️ | ☀️ |
| Security | CISO, R&D, Infra | No integration with CI/CD, production systems, or secrets management | ☀️ | ☀️ | ☀️ | ☀️ | ☀️ | ☀️ | ☀️ | ☀️ |
| Maintainability / Serviceability | CTO, Cloud Ops, R&D, Infra | Tool should require minimal in-house operational burden (patching, model/version updates, serving uptime) - maintenance vendor managed | ☀️ | ☀️ | ☀️ | ☀️ | ☀️ | ☀️ | ️ | ☀️ |
| Evolvability | CTO, R&D | Tool should align with company's strategic direction for developer experience | neutral | ☀️ | ☀️ | ☀️ | neutral | ️ | ☀️ | ️ |
Rating legend: ️ = negative impact / ☀️ = positive impact / neutral = no significant impact
1.4.1 Rational Risk¶
Description: The risk of exposing proprietary IP through AI coding assistant usage, covering both input risk (source code confidentiality / model training) and output risk (license contamination from generated code). * Ideal: Copilot Enterprise provides contractual guarantees that no customer code is used for training; code suggestions are filtered for known license violations; mandatory ScanCode scanning catches remaining issues before merge. * Worst: Source code is ingested and used to train models accessible to competitors; AI generated code contains GPL-licensed snippets that go undetected, forcing involuntary open-sourcing of proprietary codebase.
1.4.2 Rational Compliance¶
Description: Regulatory and standards compliance under EU AI Act, GDPR, and ISO/IEC 42001 for the use of AI tools processing source code. * Ideal: Enterprise DPA is signed; no DPIA required (no personal data processed); usage is documented under AIMS; audit trail exists via Jira tickets and ScanCode reports. * Worst: Usage of a consumer-grade account without DPA leads to regulatory findings; lack of documentation results in ISO audit non-conformity.
1.4.3 Rational Cost¶
Description: Total cost of ownership for the pilot, including licence fees and administration.
1.4.4 Rational IDE Integration¶
Description: The depth and breadth of the tool's integration into the developer's IDE workflow, covering inline completions, chat, agent mode, terminal assistance, and custom configuration. * Ideal: The tool provides inline code completions as the developer types, predicts the next edit location, operates autonomously across multiple files (agent mode), assists in the terminal, supports custom project-level instructions, and works natively in both VS Code and JetBrains. The developer never needs to leave the IDE or copy-paste code from a browser. * Worst: The tool offers no inline completions and only Beta JetBrains support. Developers working in IntelliJ can use the Claude Code CLI (and its Beta plugin) in a terminal alongside their IDE, but have no inline completions or Next Edit Suggestions in-editor. Those using VS Code get Claude Code in the terminal and via the extension but still lack as-you-type ghost text suggestions - the most frequent AI-assisted interaction in day-to-day work.
1.4.5 Rational Maintainability / Serviceability¶
Description: The internal resource footprint required to operate, support, and keep the solution current.
1.5 Pros and Cons of the Options¶
1.5.1 AS-IS¶
Maintain the current whitelist with only Claude Sonnet (browser-based chat) approved. Do not add any IDE-integrated AI coding assistant.
Capabilities of AS-IS (Claude Code CLI): * Terminal-first agentic workflow with autonomous multi-file editing * VS Code extension with chat and agentic capabilities * MCP support for external tool integration * Custom project instructions via CLAUDE.md files * Can be used in tmux/SSH sessions for uninterrupted remote work
Remaining limitations of AS-IS: * No inline code completions while typing (ghost text) - the most frequently used AI assist feature * No Next Edit Suggestions (predicting where to edit next) * JetBrains integration is Beta only - the Claude Code JetBrains plugin runs the CLI in the integrated terminal but is not a first-party native integration; no inline completions in IntelliJ * Claude models only * Risk (Input): ☀️ No additional risk introduced. * Risk (Output): ☀️ No additional output risk. * IDE Integration: ️ No improvement. Engineers continue with Claude Code (terminal / VS Code; JetBrains Beta) but lack inline completions and Next Edit Suggestions. * Model Flexibility: ️ Only Claude models. No access to GPT, Gemini, or other models within the IDE. * Efficiency: neutral Status quo. Productivity gains from IDE-integrated AI tools are forfeited. * Compliance: ☀️ No additional compliance burden. * Cost: ☀️ No additional licence cost. * Cloud Vendor Neutrality: ☀️ No new dependencies. * Security: ☀️ No additional attack surface. * Maintainability / Serviceability: ☀️ Nothing new to operate or maintain. * Evolvability: ️ Limits tooling options to Claude-only. Engineers may use unapproved tools informally (shadow IT risk). No inline completions puts company behind industry standard.
1.5.2 Option 1: GitHub Copilot Enterprise¶
Add GitHub Copilot Enterprise as an approved assistant layer. * Model flexibility: ☀️ Copilot supports 15+ underlying models, including Anthropic Claude (Opus 4.7, Sonnet 4.6 - already approved at company), OpenAI GPT-5.x and Google Gemini 3.x. Engineers can select the approved Claude model inside Copilot, combining it with Copilot's IDE integration layer. Copilot additionally offers a GitHub-built Claude agent (Claude Agent SDK, public preview); this is the Claude model/agent, not Anthropic's standalone Claude Code product. The investment in the approved Claude model is preserved within Copilot. * Custom instructions: ☀️ company can define coding conventions, architecture rules, and security guidelines in .github/copilot-instructions.md files at the repo level. This ensures AI-generated code follows company standards (we already have these files in our repositories). * MCP support: ☀️ Model Context Protocol allows Copilot to connect to external tools and data sources. This is relevant for future integration with company's internal tooling. * IDE Integration: ☀️ Completions, agent mode, terminal, MCP, custom instructions all GA. * Efficiency: ☀️ Inline completions eliminate copy-paste. Agent mode enables multi-file refactoring. Terminal integration assists with commands. * Compliance: ☀️ Enterprise DPA. No DPIA required. Aligned with ADR-0000065. * Cost: ☀️ $190-$390/month base for 10 users, within budget. Premium-model use is metered via "AI Credits" from 1 June 2026 (as with all hosted tools - see Rational Cost). * Cloud Vendor Neutrality: ☀️ GitHub is a development platform, not a cloud infrastructure provider. No AWS/GCP lock-in. * Security: ☀️ IDE-only, no CI/CD, no production access. Sandbox restrictions enforced. * Maintainability / Serviceability: ☀️ Fully vendor-managed SaaS (GitHub/Microsoft); no in-house serving or model upkeep.
1.5.3 Option 2: Claude Code (Anthropic)¶
Add Anthropic Claude Code as the next approved AI coding assistant. Claude Code is Anthropic's agentic coding tool, designed as a terminal-first CLI with a VS Code extension and a JetBrains plugin. Claude Code is already available to company engineers under the existing Claude whitelist approval (managed under ADR-0000065) and can be used in the terminal, tmux sessions, or within the VS Code extension. It reads/writes files, runs commands, and performs multi-file edits autonomously.
Strengths: * Terminal-first agentic workflow - Claude Code's primary interface is the terminal. It excels at autonomous multi-file refactoring, running tests, and iterating on code until it passes; plan mode, subagents and background tasks make it arguably best-in-class for agentic editing. * Native VS Code panel - a graphical in-IDE experience (dockable chat panel, side-by-side inline diffs with accept/reject, plan mode, @-file mentions, checkpoints), not a terminal wrapper. GA. * MCP support - managed MCP allow/deny lists, OpenTelemetry usage/cost metrics, audit/compliance API, and data-residency routing via Amazon Bedrock or Google Vertex AI.
Weaknesses: * No inline code completions - Claude Code does not provide ghost-text completions while typing. This is the most frequently used AI assist feature and a significant gap for day-to-day coding productivity. * No Next Edit Suggestions - cannot predict the next edit location. * JetBrains integration is Beta - the JetBrains plugin runs the CLI in the integrated terminal and shows diffs in the IDE viewer, but it is still labelled Beta and is not a first-party native integration; IntelliJ users otherwise rely on the terminal CLI alongside their IDE. * Claude models only - limited to Anthropic's own models (Opus 4.7, Sonnet 4.6, Haiku 4.5). No access to GPT, Gemini, or other models. * Tier sizing - the pilot fits Claude Team Standard (~$200/mo for 10, comparable to Copilot Business); Team Premium ($100/seat/mo) and Enterprise (seat + API) cost more. * IDE Integration: Beta - mixed: strong where it acts, missing as-you-type completion. * Model Flexibility: neutral Provides the already-approved Claude (Opus 4.7, Sonnet 4.6), routable via Bedrock / Vertex / Foundry, but no non-Claude models - comparable to Amazon Q's single-vendor path. * Efficiency: ☀️ Strong agentic capabilities for complex refactoring tasks; lack of inline completions is the main day-to-day gap. * Compliance: ☀️ Anthropic Enterprise DPA (eff. 1 Jan 2026). ZDR available for enterprise API incl. Claude Code. No DPIA required. * Cost: ☀️ Included on every Team seat; Team Standard ~$200/mo for 10 is comparable to Copilot Business (~$190/mo). Team Premium / Enterprise + API are optional and pricier. * Cloud Vendor Neutrality: ☀️ No cloud vendor dependency. * Security: ☀️ Terminal/IDE-only. No CI/CD integration. * Maintainability / Serviceability: ☀️ Vendor-managed (Anthropic); updates ship via the CLI / extension. No in-house serving to operate. * Evolvability: ☀️ Already approved at company; fast-evolving (skills, plugins).
1.5.4 Option 3: JetBrains AI Assistant Enterprise¶
Add JetBrains AI Assistant Enterprise as the next approved AI coding assistant. JetBrains AI is the next-closest alternative after Copilot and would be the stronger choice if company used exclusively JetBrains IDEs.
Strengths: * Native JetBrains integration - first-party, deeply integrated into IntelliJ's code analysis engine. Leverages IntelliJ's semantic understanding of the codebase. * Junie agent - autonomous coding agent for JetBrains IDEs (GA); Junie CLI (Beta) extends the agent to VS Code/Neovim via the terminal. * Air-gapped on-premises deployment - AI Enterprise can run fully on-premises via JetBrains IDE Services 2026.0, connecting to on-prem OpenAI-compatible servers (vLLM, llama.cpp, LM Studio) and shipping JetBrains Mellum for completion - the strongest data-sovereignty story among the commercial options. * BYOK (Bring Your Own Key) - now GA - connect to OpenAI/Anthropic or local models via Ollama and LM Studio; BYOK does not require a JetBrains AI subscription.
Weaknesses: * VS Code reach is CLI-only - there is no native VS Code plugin; VS Code users get Junie only through the Junie CLI. * Custom instructions limited - custom instructions remain limited compared with Copilot. * Cost model structure: Cost depends on tier: AI Pro (business) ~$200/month for 10 users, AI Ultimate ~$600/month, AI Enterprise is custom-priced. Its genuine cost differentiator is the separately-required JetBrains IDE licence on top of the AI subscription. * Risk (Output): neutral Same output risk as all other options. Mitigated by ScanCode + review. * IDE Integration: neutral Excellent in JetBrains; VS Code is CLI-only (Junie CLI). MCP now GA. Junie CLI adds terminal access. Custom instructions limited. * Model Flexibility: ☀️ Multiple models + BYOK (GA) + local model support (Ollama / LM Studio). * Efficiency: ☀️ Strong for JetBrains users. Junie agent (GA) for autonomous tasks. * Compliance: ☀️ Enterprise DPA. Air-gapped on-premises option for strict data sovereignty. * Cost: neutral AI Ultimate ~$600/month for 10 users (AI Pro ~$200; Enterprise custom), plus separate IDE licences. * Cloud Vendor Neutrality: ☀️ No cloud vendor dependency. On-premises option available. * Security: ☀️ IDE-only. On-premises option for maximum control. * Maintainability / Serviceability: ☀️ Vendor-managed SaaS by default (the air-gapped on-prem AI Enterprise variant shifts some serving in-house). * Evolvability: ☀️ Strong for JetBrains-centric teams. BYOK enables future model flexibility.
1.5.5 Option 4: Amazon Q Developer Pro¶
Add Amazon Q Developer Pro as the next approved AI coding assistant.
Strengths: * Cost-effective - $19/user/month (Pro) with a perpetual free tier (50 agentic requests/month). * Runs Anthropic Claude via Bedrock - now uses Claude (Sonnet 4.6) as its coding model, with model selection in IDE and CLI; AWS reports it tops the SWE-bench leaderboard. * AWS integration - strong for teams working with AWS services (Lambda, S3, DynamoDB, etc.). * Java upgrade automation - built-in tooling for Java version upgrades and .NET porting.
Weaknesses: * AWS ecosystem lock-in - although the model is now Claude, Q Developer's broader value is optimised for the AWS ecosystem (Bedrock, IAM, AWS services). * Single-vendor model path - Claude is delivered exclusively via AWS Bedrock; there is no choice of OpenAI/Gemini/local models the way Copilot or JetBrains BYOK allow. * IDE integration shallower than Copilot - extension-level in VS Code and JetBrains, plus a Q CLI and MCP support, but no org-level custom instructions. * Efficiency: ☀️ Good code completion. Java upgrade tooling useful but narrow. * Compliance: ☀️ AWS enterprise agreements. * Cost: ☀️ $190/month for 10 users. Cheapest option with generous free tier. * Cloud Vendor Neutrality: ️ Drives AWS lock-in. Conflicts with OVH Cloud migration strategy. * Security: ☀️ IDE-only. Standard enterprise controls. * Maintainability / Serviceability: ☀️ Fully AWS-managed; no in-house serving or model upkeep. * Evolvability: neutral Useful only if company moves to AWS.
1.5.6 Option 5: Google Gemini Code Assist¶
Add Google Gemini Code Assist Standard or Enterprise as the next approved AI coding assistant.
Strengths: * Cost-competitive - $19/user/month (Standard), $45/user/month (Enterprise). * Large context window - 1M token context window at Enterprise tier. * GCP integration - strong for teams working with Firebase, BigQuery, Apigee.
Weaknesses: * GCP vendor lock-in - designed to drive Google Cloud adoption; deeper features assume Firebase, BigQuery, or Vertex AI. * Gemini models only - cannot use Claude, GPT, or other models. * IDE Integration Shallower than Copilot / JetBrains AI - extension-level only, no org-level custom instructions, no admin-managed MCP. Agent Mode and MCP are now GA on the Gemini 3 family, but Google is sunsetting the individual free tier and the Gemini CLI on 18 June 2026 in favour of the separate Antigravity IDE (a VS Code fork), which adds real churn risk for a tool company would standardise on. * Irrelevant for company - company does not use GCP, making GCP-specific features unused. * Risk (Input): ☀️ Contractual no-training guarantee. Google Cloud DPA. * Risk (Output): neutral Same output risk. Mitigated by ScanCode + review. * IDE Integration: neutral Inline completion + Agent Mode + MCP (GA, Gemini 3) in both IDEs via extension, but no org-level custom instructions or admin-managed MCP. (Surface churn risk is captured under Evolvability.) * Model Flexibility: ️ Gemini models only. Cannot use Claude or other approved models. * Efficiency: ☀️ Good code completion. Large context window (Enterprise). * Compliance: ☀️ Google Cloud DPA.
1.5.7 Option 6: Self-hosted open-weight models¶
Run open-weight models that company controls, instead of a vendor SaaS. All variants share the same software layers and differ only in where inference runs, the cost model, and the maximum model size:
- Models (Apache-2.0 / MIT only, to keep the licence story clean): Qwen2.5-Coder (7B/14B/32B, strong FIM) or StarCoder2 for inline completion; Qwen3-Coder-Next (80B-A3B, SWE-bench Verified ~70.6), GLM-4.5-Air (106B-A12B) or Devstral Small 2 (24B) for chat / agent; IBM Granite Code where vendor support / indemnification matters. Frontier-scale open models (GLM-4.7 ~358B, DeepSeek-V3.2 685B, Devstral 2 123B) score higher (low-70s SWE-bench Verified) but need multi-GPU. Avoid the Qwen2.5-Coder 3B (Qwen-Research licence); Devstral 2 (123B) and Llama 4 need legal review of the licence.
- Inference stack: an OpenAI-compatible engine - vLLM (server) or Ollama / LM Studio / Foundry Local / AMD Lemonade (single machine) - exposing /v1/chat/completions so any IDE plugin connects.
| Dimension | 6a On-premises server | 6b Azure-hosted | 6c Developer laptop |
|---|---|---|---|
| Where inference runs | company data centre (GPU box) | company's own Azure tenant (pin EU region) | the developer's own machine |
| Representative target hardware / setup | 1x H100-80 GB box | Azure AI Foundry model deployment + Foundry Agent Service "Hosted Agents" (scale-to-zero) | Apple M4 Max 64-128 GB, AMD Ryzen AI Max+ 395 128 GB, or Copilot+ PC (40+ TOPS, 32 GB+) |
| Max practical model size | 80B-A3B (single GPU); frontier on multi-GPU | any (Foundry catalog or your container; elastic GPU) | ~7-32B at usable speed; up to ~70-120B Q4 on 128 GB unified memory |
| Cost model | capex ~$35-45K + power + ML-ops | Azure consumption (per-GPU-hour, scale-to-zero) - opex | laptop premium ~$2.5-4K one-time (developers need laptops anyway) |
| Ops burden | High - own GPUs, drivers, serving, upgrades | Medium - managed lifecycle / scaling / observability (App Insights) | Low-medium - per laptop, standardised via an IT image |
| Inline completion | Yes (Continue / Tabby) | Endpoint for IDE plugins | Yes - fully offline |
| Data location | company premises (strongest) | company's Azure tenant, EU region, no training, Azure DPA | the laptop memory (~7-32B; 70B Q4 on 128 GB). |
- 6a On-premises server - maximum sovereignty and control: a GPU box on company premises serving a shared completion + agent model. Highest capex and ML-ops burden.
- 6b Azure-hosted - trades capex and GPU ops for Azure consumption billing and managed lifecycle. Microsoft's run containerised agent frameworks (LangGraph, Microsoft Agent Framework, custom code) with scale-to-zero, OpenAI-compatible invocation and built-in OpenTelemetry observability, deployed via the Azure Developer CLI; the model itself is served from the Azure AI Foundry catalog or your own container. Data stays within company's Azure tenant (pin an EU region / EU Data Boundary, no-training, Azure DPA) - strong, but not "never leaves any cloud", and it introduces Microsoft Azure as a cloud dependency company is not otherwise standardising on. The same managed pattern can be run on OVH Cloud GPU instances - company's actual migration target and an EU-sovereign provider - which would be more strategically aligned than Azure.
Weaknesses: * Ops / maintenance - 6a carries a full ML-ops burden (GPUs, drivers, serving, the upgrade treadmill); 6b reduces it to managed-deployment management; 6c spreads a lighter burden across each laptop. None has a vendor coding-assistant SLA (except IBM Granite weights). * Capex / setup (6a) and cloud dependency (6b) - 6a needs ~$35-45K and lead time; 6b adds Microsoft Azure as a cloud vendor that conflicts with the OVH direction (mitigated by hosting on OVH GPU instead). * Quality / size ceiling - all variants trail frontier hosted models (Claude Opus 4.7, GPT-5.x, Gemini 3.x) on the hardest long-horizon agentic work; the laptop variant (6c) has the lowest ceiling, with model size capped by laptop memory and speed by thermals / battery. * Licence diligence required - stick to Apache-2.0 (Qwen-Coder, Devstral Small, Granite) and MIT (DeepSeek, GLM) weights; Devstral 2 (123B, revenue-gated modified MIT), Llama 4 (community licence) and StarCoder2 (OpenRAIL-M behavioural restrictions) need legal review. * Efficiency: ☀️ Copilot-class inline completion (Qwen2.5-Coder / StarCoder2) plus strong agentic edits for routine work deliver real productivity gains; the quality gap is confined to the hardest long-horizon agentic tasks (largest on 6c). * Compliance: ☀️ 6a / 6c have no external processor (no DPA/DPIA dependency); 6b runs under the Azure DPA in an EU region. Clean GDPR / EU AI Act posture. * Cost: neutral No per-seat licence fees. Varies by variant: 6c near-zero marginal (laptops developers need anyway), 6b metered opex with scale-to-zero, 6a ~$35-45K capex + ML-ops. Acceptable for a pilot via 6c. * Cloud Vendor Neutrality: ☀️ 6a / 6c have no cloud dependency (strongest); 6b adds a Microsoft Azure dependency - prefer OVH GPU for an aligned cloud variant. * Security: ☀️ IDE-only, no CI/CD or production access. 6a / 6c isolate inference on owned hardware; 6b adds an Azure tenant company must secure. * Maintainability / Serviceability: ️ The main weakness. 6a / 6b require in-house ML-ops (GPU / serving / quantization upkeep, model updates, no vendor SLA); 6c is lighter but still per-laptop upkeep. Needs a capable ML-ops engineer. * Evolvability: ☀️ Aligns with data-sovereignty and OVH migration strategy; model layer is freely upgradable as open models advance.
1.5.8 Option 7: Do not approve any additional tool¶
(Evaluated dynamically under section 1.5.1 AS-IS variant).
1.6 Links and References¶
- ADR-0000065: Decision of use of AI Coding Assistants (parent ADR establishing the tactical usage framework)
- The enterprise whitelist managed under ADR-0000065
1.6.1 Vendor pricing and plan pages (verified 2026-05-22)¶
- GitHub Copilot Business/Enterprise pricing, premium-request allowances
- GitHub Copilot "AI Credits" model effective 1 June 2026; org budget controls
- Anthropic Claude Team Standard/Premium and Enterprise seat + usage
- Claude Code included on every Team seat
- JetBrains AI tiers, AI-credit model, air-gapped deployment
- Amazon Q Free/Pro tiers; now runs Claude (Sonnet 4.6) via Bedrock
- Google Gemini individual free tier and Gemini CLI sunset on 18 June 2026
1.6.2 Self-hosted open-weight option (Option 6)¶
- Qwen3-Coder-480B model card
- Qwen2.5-Coder family (Apache-2.0 code models, FIM completion)
- DeepSeek-V3.2 (MIT)
- GLM-4.7 (MIT)
- Devstral 2 / Vibe CLI
- IBM Granite Code (Apache-2.0)
- StarCoder2 (BigCode purpose-built completion model, OpenRAIL-M licence)
- vLLM and Ollama (OpenAI-compatible self-hosted inference engines)
- Continue.dev and Tabby (open-source IDE front-ends for VS Code + JetBrains)
1.6.3 Operational Frameworks & Risk Docs¶
- GitHub Copilot Enterprise Privacy (GitHub's data handling and training exclusion commitments)
- Claude Code Documentation (Anthropic's agentic coding tool for terminal and VS Code)
- ScanCode Toolkit Documentation (mandatory OSS licence scanning tool)
- OWASP Top 10 for LLM Applications (risk framework referenced in ADR-0000065)
- EU AI Act (regulatory framework)
- ISO/IEC 42001 - AI Management System (management system standard for AI)