EclipseSource

If your Eclipse RCP application or Eclipse Tools is expected to remain important for the next five or ten years, now is a good time to revisit your migration strategy. AI-assisted development is changing the economics of modernization: codebase analysis, characterization testing, business-logic extraction, UI reimplementation, and repetitive migration work can all be accelerated significantly. But migration is still a re-architectur, not a mechanical conversion. The key decisions around what to keep, redesign, reuse, or rebuild still require deep architectural and domain expertise. Our latest article looks at what has changed in 2026, how to approach reuse vs. rewrite, how to manage the dual-tool phase, and why migrating while you still control the timeline is often the strongest position. Read the full article: https://lnkd.in/dE277xwN #EclipseRCP #SoftwareModernization #AI #eclipsetheia #theiaide #theiai #eclipseclouddevtools

How do you build a custom tool or IDE in 2026? AI is no longer just another feature on the roadmap. It changes how tools are designed, how workflows work, who interacts with them, and how quickly teams can build and validate new ideas. Our updated guide looks at what this means in practice: designing for both humans and AI agents, creating review-centric UX, exposing capabilities through MCP or CLI, keeping AI integrations flexible, choosing the right platform and foundation, and using AI to build better PoCs faster. If you are planning a new tool, modernizing an existing IDE, or thinking about how AI should fit into your product architecture, this might be a good place to start. Read the full article: https://lnkd.in/dj4Bx5ee #eclipsetheia #theiaide #theiai #eclipseclouddevtools #AIEngineering #DeveloperTools #IDE #MCP #SoftwareArchitecture

Eclipse Theia 1.75 is now available! Highlights include support for Agent Plugins from the AI Registry, a new Memory capability for agents, MCP Apps that can render interactive UI directly in AI chat, and a unified AI configuration experience. The release also includes major improvements to the Source Control Graph, walkthrough support on the Welcome page, startup performance optimizations, updated base technologies, and many additional fixes and enhancements. Explore what’s new in Eclipse Theia 1.75: https://lnkd.in/da-e8W3z #eclipsetheia #theiaide #theiai #eclipseclouddevtools #opensource #AI #MCP #developerexperience #clouddevelopment

Eclipse Enclave is a sandbox that runs AI coding agents in isolated containers behind a network allowlist you control and is now **open source at the Eclipse Foundation**. We originally built it in late 2025 for our own team, internally called YOLOArena, because we wanted to run Claude Code and other coding agents in YOLO mode, without confirming every command and without worrying about our laptops, which hold SSH keys, source code and other secrets. Every agent session runs in its own container (an "Enclave"), with its own view of the filesystem and its own network gateway. The agent sees the project directory you started it in and nothing else, unless you share it. All outbound traffic passes through a DNS allowlist and proxy that you control, so you can leave the permission prompts switched off and still know what the agent can and cannot reach. One command in a project directory: $ enclave And Enclave does the rest: builds the container with the agent pre-installed, sets up the mounts, caches the agent's authentication so you log in once instead of on every run, configures the network, and drops you into a session with the permission prompts already off (for Claude Code, that means --dangerously-skip-permissions). Claude Code, Codex CLI, OpenCode, Theia AI, Pi and Mistral Vibe work out of the box, and other agents can be added with Enclave extensions. It wraps the devcontainer you already have instead of replacing it, and you can start as many sessions as you like in parallel and re-attach when you are ready to steer and review. MIT licensed. Linux and macOS natively, Windows through WSL2. The intro video with demos and the full write-up: https://lnkd.in/dBdtMXhG https://lnkd.in/dYmEYATU Project page at the Eclipse Foundation: https://lnkd.in/dz9CjXW6 The code has been in daily use here for months, but as an open source project Enclave is only a few weeks old. Issues, PRs and stars are what tell us where to take it next: https://lnkd.in/d98FEJ4H #AIAgents #AgenticAI #AICoding #EclipseFoundation #DevSecOps

🚀 Eclipse Theia 1.74 is released! This month’s release brings 60 merged improvements across Eclipse Theia and Theia AI. Among the highlights are a new experimental AI First perspective that puts AI Chat at the center of the workbench, a dedicated AI Sessions model selection, and server-side compaction to keep long-running AI sessions going as they approach model context limits. The release also adds improvements for multi-agent sessions, provider-native server tools, security hardening, a new telemetry framework, updated base technologies, and many smaller workbench and platform enhancements. Explore all the news and noteworthy changes: https://lnkd.in/eT3BZKD5 #eclipsetheia #theiaide #theiai #eclipseclouddevtools #opensource #AI #AIAgents #DeveloperTools #IDE

Claude Code still doesn't read AGENTS.md. I honestly don't get why, the feature request has been open since August 2025 with thousands of reactions, and AGENTS.md is now a Linux Foundation standard used across tens of thousands of repos. However, the good news: you don't need to wait, and you don't need to maintain two instruction files. Create a CLAUDE.md with exactly this: → @AGENTS.md → ## Claude Code → <your Claude-specific rules here> The @-import pulls your canonical AGENTS.md into every session; the section below holds tool-specific instructions. If you have nothing Claude-specific to add you can also simple symlink: ln -s AGENTS.md CLAUDE.md (just note that symlinks need extra setup on "Some OS :-), so the import is the safer default for mixed teams). This is also the pattern I recommend in our AI coding trainings for teams running various tools such as Claude Code, Codex, Cursor, Kiro or Theia Coder. One tool-agnostic source of truth, optional thin tool-specific overlays on top. #AIcoding #ClaudeCode #TheiaIDE #SoftwareEngineering #SystematicAICoding

We built a small MCP server for a problem we kept running into with AI-supported code reviews: How do you let an agent help with a real GitHub review without giving it too much control? ReviewGuard MCP keeps the GitHub token away from the agent and defines the permitted operations when the server starts. In the default setup, the agent can read a PR and prepare a draft review, but it cannot submit anything publicly. It also supports more autonomous review setups, with explicit controls such as limiting the agent to one PR, restricting review actions, and preventing it from resolving other people’s threads. It is something useful we wanted to share because others are probably dealing with the same issue. The server works over stdio or HTTP and can be used with Claude Code, Codex, Theia AI, or other MCP clients. Article and links: https://lnkd.in/dttMqDUZ How are you currently controlling what coding agents are allowed to do in GitHub? #MCP #AICoding #CodeReview #GitHub #DeveloperTools #systematicaicoding

In our Systematic AI Coding trainings, we have always advocated for context lifecycle management as a fixed, central part of the workflow - deliberately scoping sessions, ending them, and carrying distilled state forward. The anti-pattern: one endless conversation about mixed topics, kept alive forever, "so the AI doesn't forget things." The main arguments are generation quality (long, polluted contexts dilute attention and drag stale decisions into new tasks) and structure for developers (a session with a clear scope forces you to think in well-defined work packages). However, there is a third argument: cost. Prompt caching makes continuing a conversation look cheap - cache reads cost only a fraction of regular input tokens. But two effects work against the "just keep it running" strategy: 1. Costs compound with context length. Every turn re-reads the entire history. A bloated context multiplies the cost of every single message - even at discounted cache rates, a long mixed session quietly exceeds what several focused sessions would cost. 2. The cache expires - and long sessions invite pauses. This is the part I measured: if you step away and the cache expires, your next message re-writes the entire accumulated context at premium rates. The runs after a pause cost a multiple of the warm-cache runs. And it's precisely the long-running "keep everything" sessions that you tend to pause - for a meeting, a coffee, overnight. If you want to see the effect yourself: 1. Warm the cache of any long-running conversation with a message Snapshot the session cost (A) 2. Immediately send another message and snapshot again (B) → Diff 1 = B − A 3. Wait until the cache has expired (5 minutes by default, up to 1 hour depending on configuration), send the same message again, and snapshot (C) → Diff 2 = C − B Diff 1 and Diff 2 are the same message - warm vs. cold cache. The ratio between them is the price of that pause. So the endless-conversation strategy loses twice: while it runs, you pay linearly growing costs for degrading quality. When you pause it, you pay the full re-write on top. The alternative: work in focused bursts with a good plan and while the context is warm, then deliberately close the session and hand distilled state to the next one - a summary, an updated project context file, a handoff document. Explicit compaction over accumulation. Larger context windows have shifted where the boundary sits - sessions can reasonably be coarser-grained today than in the 200k era. But they don't remove the boundary: the principle stays the same, and ironically, the bigger the context, the more expensive the cold re-write after a pause. Quality and structure were always reason enough. In times of token scarcity, the economics now point in exactly the same direction. How do you handle session boundaries in your AI coding workflow? #systematicaicoding #dibecoding #videcoding #aicoding #visecoding #ClaudeCode

💡 Tip of the day: /model opusplan in Claude Code With "opusplan", Claude Code uses Opus while you're in plan mode, where the deep reasoning actually matters, and then automatically switches to Sonnet for execution. The result: you get Opus-level architecture and planning quality, but the (usually much longer) implementation phase runs on the faster and cheaper model. No manual switching, it just happens when you exit plan mode. You can set it permanently: claude --model opusplan or via "model": "opusplan" in your settings. I picked this one up during a "Systematic AI Coding" onsite training I gave last week - a participant brought it up in a discussion. That's why I enjoy live trainings so much: you always walk away having learned something as a trainer, too! #systematicaicoding #dibecoding #videcoding #aicoding #visecoding #ClaudeCode