The Eclipse Theia Community Release 2026-08

We are happy to announce the fifteenth Eclipse Theia community release, “2026-08,” incorporating the latest advances from Theia releases 1.72, 1.73, and 1.74.

New to Eclipse Theia? It is the next-generation platform for building IDEs and tools for the web or desktop, based on modern web technologies. With Theia AI, you can build AI-powered tools and IDEs with ease. Visit the Theia website for more info.

The Theia project also delivers a product: the Theia IDE, a modern, AI-powered, and open IDE for cloud and desktop environments. The Theia IDE is based on Theia and includes advanced AI-powered features. See the Theia IDE website or download the latest release for a hands-on experience!

Community releases are provided quarterly and serve as stable anchors for downstream adopters. Each combines features, fixes, and hardening from monthly releases. Learn more about the advantages of the community release, or visit Theia’s releases page.

Two related projects saw notable news during this release cycle. Eclipse Enclave is now available as open source at the Eclipse Foundation: a sandbox that runs Claude Code, Codex, and other AI coding agents in isolated containers behind a network allowlist, so you can let agents run at full autonomy without exposing your machine. ReviewGuard MCP takes a similar angle for AI-assisted code review: the MCP server holds the GitHub token and fixes the allowed operations at server start, so the agent cannot talk its way past them. On the modeling side, Eclipse GLSP editors now expose their diagrams via the Model Context Protocol, which lets agents inspect, query, modify, and validate models through a protocol that Theia AI and other MCP clients already speak.

Now let’s explore the key highlights from this fifteenth Theia community release!

Highlights of the Theia 2026-08 Community Release

This community release ships cumulative improvements from releases 1.72, 1.73, and 1.74, totalling 195 merged pull requests. For a comprehensive list of changes, please see the individual announcements. Here are some top highlights:

Installing MCP Servers and Skills from the AI Registry

The headline feature of this community release is the integration of the Eclipse Foundation AI Registry in Theia 1.73. Finding and installing MCP servers or skills used to mean tracking down a server, copying its configuration into your settings by hand, and hoping it came from a source you trust. With the AI Registry integration, MCP servers and skills appear as first-class entries in the Extensions view, right next to Open VSX extensions, with a type filter to narrow results. Search, click install, done.

The registry is a vendor-neutral trust registry hosted by the Eclipse Foundation alongside Open VSX. It records which organizations have reviewed and endorsed an artifact, and ships ready-to-use install configurations per tool. Installed entries stay connected to the registry: newer versions offer an update, locally modified configurations can be restored, and entries that disappear from the registry are flagged. Installs can also start from the registry website through a theia://install-mcp?id=<serverId> link. For adopters, the integration ships as the new @theia/ai-registry package and is opt-in per product.

Provider-Native Server Tools

Tools that an agent calls normally run locally in the Theia client. Theia 1.73 adds provider-native server tools that the LLM provider executes on its own infrastructure, starting with web_fetch and web_search for Anthropic and url_context and google_search for Gemini. Theia shows the call and its result in the chat, but nothing executes on your machine. For common needs like web search, this replaces a dedicated MCP server: there is nothing to install, configure, or keep running. Server tools are enabled in the capabilities panel under a “Server Tools” group per provider, are auto-approved since they never touch your system, and the selection is remembered per provider.

Theia 1.74 extends this to OpenAI web search via the Responses API, and makes the deferred-tool search visible as a server tool call, so you can see it run and how many tools it loaded.

Perspectives and the AI First Layout

Theia 1.74 introduces perspectives, a declarative way to describe which views belong to a workflow and where they should sit. Two commands drive it while the feature is experimental: Switch Perspective (Experimental) activates one, and Reset Perspective (Experimental) puts views back where the perspective wants them after you have dragged things around. Switching back to Default restores your previous layout.

The release ships one concrete perspective, AI First, which puts the AI Chat in the main area, the new AI Sessions view on the left, Explorer and Source Control on the right, and collapses the bottom area. The AI Sessions view itself is new: the session list was previously only reachable from the chat welcome message, and is now a standalone widget you can keep visible next to a running conversation. For adopters, perspectives are a clean way to ship their own view arrangements for specific workflows.

Longer Sessions: Compaction and Per-Session Model Selection

Long agent sessions eventually hit the model’s context limit, which used to mean a failed request and a fresh start. Theia 1.74 adds server-side compaction: the provider summarizes older turns on its side so the session keeps working. Activation is layered from a global setting (ai-features.chat.serverSideCompaction, on by default) through a per-provider override to a per-session override. Whether it actually happens depends on a model capability, currently honored by Anthropic and by OpenAI through the Responses API. Compactions are shown inline in the chat, counted in the token usage tooltip, and persisted with the session.

Complementing this, a model selector in the chat input lets you run the current session with a different model than the agent’s configured default. The override is deliberately session-scoped, with no “save as default”, and responses carry a badge showing which model produced them. It is useful for a cheaper model on a throwaway question or a stronger one for a single hard task, without editing global settings and forgetting to change them back.

Keeping Track of Multiple Agent Sessions

As agents run longer and delegate to each other, knowing which session needs you becomes a real problem. Theia 1.73 marks sessions that pause for a tool confirmation or a structured question with a waiting-for-input indicator in the chat overview: a bell icon, an attention-coloured dot, and a bold title, tied into the agent notifications so you can jump straight to the session that is waiting.

Theia 1.74 builds a delegation hierarchy on top of that. Sub-sessions created by delegation are grouped under their parent as a collapsible, multi-level tree that survives a restart, and a waiting descendant auto-expands and marks its ancestors, so you notice it even on a collapsed parent. Session status is also more precise: waiting for a tool approval now reads “Requires your approval” instead of the generic waiting message. For adopters, all of this is backed by an observable, aggregated status on ChatModel (idle, running, awaitingApproval, awaitingToolCall, awaitingInput, failed) that session views and notifications read instead of re-deriving state from request flags.

AI Agents in Multi-Root Workspaces

Before Theia 1.72, file access, task lookup, launch configurations, and command execution for AI agents were effectively scoped to the first workspace root. With multi-root workspace support for AI agents, agents see and address all roots equally using a <rootName>/<relativePath> convention. A new variable contribution injects the root information into every agent system prompt, tasks and launch configurations disambiguate by an explicit workspaceRoot parameter when names collide, the shell execution tool requires a cwd, and .gitignore is handled per root. Adopters that built on the previous single-root assumptions should check the release notes, as a few APIs were removed in the process.

Safer Defaults Around Tool Execution

Theia 1.72 shifts the default for tool confirmation toward caution. The * magic key in ai-features.chat.toolConfirmation is replaced by a regular ai-features.chat.defaultToolConfirmation preference visible in the Settings UI, and its default changes from always_allow to confirm. New MCP servers and newly added skills now ask before they run unattended. To keep that from becoming tedious, the confirmation dialog shows the tool description and a collapsible preview of the arguments, so you can see what createPullRequest is about to post at the moment you decide. Keyboard shortcuts approve or deny the latest pending call, and the Tools tab gains an action that bulk-approves the built-in tools while leaving MCP servers untouched.

In the other direction, a new ai-features.workspaceFunctions.allowedExternalPaths preference opts specific paths outside the workspace into read-only agent access, for instance a shared documentation folder. The allow-list is empty by default, applies only to read-only tools, uses path-component matching rather than string prefixes, and rejects .. traversal.

Mermaid Diagrams in the Chat

Fenced mermaid code blocks in agent responses now render as actual diagrams, themed to match the workbench, with collapse, source toggle, zoom, and pan. The rendered SVG is sanitized with DOMPurify and Mermaid runs with securityLevel: 'strict'. The architect, coder, and PR review prompts now suggest small diagrams for architecture and flows, which makes explanations a lot easier to follow than a wall of prose.

esbuild as an Alternative Bundler

For adopters, the most consequential change in this cycle is likely esbuild support in Theia 1.72 for browser, browser-only, and Electron applications. In our measurements, a full build of the Theia IDE browser application drops from roughly 20 seconds to about 2 seconds, with watch-mode rebuilds benefiting accordingly. Source maps are rewritten to absolute file:// URLs, so Chrome and Node debuggers bind breakpoints to TypeScript sources without custom configuration.

Webpack remains available and stays the default for existing applications, so the new path is opt-in for now. Two limitations are worth knowing before you switch: esbuild does not currently emit compressed artifacts and does not support code splitting.

For Agent Builders

Three changes make building and iterating on custom agents easier. Theia 1.73 moves custom agents from a shared customAgents.yml into per-agent folders under .agents/agents/<id>/, with metadata in YAML frontmatter on top of the prompt body, matching the convention skills already use. Existing setups are migrated only when you accept the migration notification, and the original file is kept as a backup.

Theia 1.74 adds the AI Tool Sketchpad, which closes a gap in that iteration loop. Prompts, models, skills, and MCP servers could already be changed at runtime, but trying a different tool shape meant writing code and rebuilding. A sketched tool declares only name, description, parameters, and return behavior, and either returns a static value or asks you for the result at runtime, so you can simulate any tool response while the agent is running. Definitions live in sketchedTools.yml and are re-registered on every change. The Sketchpad ships as the @theia/ai-tool-sketchpad package for Theia-based applications and is not part of the Theia IDE.

Finally, Theia 1.73 implements the VS Code lm.registerTool and lm.invokeTool APIs. Tools registered by a VS Code extension are fully integrated with Theia AI: they can be referenced in prompts via ~{toolId}, are auto-completed in the chat input, and can be invoked by the LLM like any other tool.

Security Hardening

Browser-based deployments should upgrade for this one. Theia 1.73 gates HTTP API and WebSocket access behind same-origin checks, engine.io transport-level validation, and a SameSite cookie-based connection token. Theia 1.74 closes the remaining gaps on the backend endpoints: the filesystem endpoints require a valid connection token, enforced by a reusable HttpConnectionValidator that adopters can apply to their own routes, plugin resource requests are again scoped to the requesting plugin’s directory, and a broad dependency update resolves a set of advisories.

One change is visible in daily use: rendered agent markdown in the AI chat no longer loads external resources automatically. Images, media, iframes, and CSS resource URLs are replaced by a placeholder showing the blocked URL, which you can allow explicitly.

VS Code Compatibility and Base Technology

VS Code extensions delivered as ESM modules used to fail to activate with errors like Cannot find package 'vscode'. Theia 1.72 adds proper ESM support by loading such plugins via dynamic import() and mapping import 'vscode' to a per-importer synthetic module. As a direct consequence, the built-in vscode.github and vscode.github-authentication extensions are no longer excluded by default. The same release also honors when clauses on view containers and auto-derives defaults for plugin configuration properties, which fixes duplicate sidebars with Anthropic’s Claude Code extension and a configuration error with the Vue extension respectively.

On the foundation side, VS Code API compatibility moves from 1.116.0 to 1.130.0 across the three releases, Electron is updated from 39.8.7 to 42.3.0 with newer Node.js, Chromium, and V8 versions, development defaults to Node 24, and @theia/core now accepts React 18 or 19 as a peer dependency so adopters can move ahead of Theia’s own switch.

Additional Highlights

Beyond the headlines, this community release brings a long list of smaller improvements. On the AI side: a new default/fast model alias backing the Explore, GitHub, Project Info, Command, and Chat Session Naming agents with cheaper, faster models; a #terminalCommand variable that pulls a specific command and its output from the terminal into a chat request; a token usage indicator that scales to the actual context window of the active model instead of a hardcoded 200k; OAuth 2.1 for remote MCP servers with PKCE and dynamic client registration; workspace file tools that run on the backend and return full-tree queries in tens of milliseconds instead of minutes; clean rendering of provider errors instead of escaped JSON blobs; and a PR review agent that walks the reviewer through a change rather than listing issues.

On the platform side: file watchers now honor files.watcherExclude end to end, which stops Linux setups from exhausting fs.inotify.max_user_watches; terminal paste works on macOS Electron again; custom editors support preview tabs and report a correct viewColumn; recently opened workspaces appear in the Windows jump list; session-scoped preferences can be injected via a --session-preference CLI flag without ever being written to disk; the Problems view gains “Copy as Text” and JSON export; the Keyboard Shortcuts dialog captures real keystrokes; and a new @theia/telemetry package adds a vendor-neutral telemetry framework that reports nothing unless an adopter contributes a sink. Two packages are on their way out: @theia/ai-vercel-ai is deprecated in favor of the dedicated OpenAI and Anthropic providers, and @theia/preview is superseded by the built-in Markdown extension.

For a comprehensive breakdown, please read the in-depth release news:

• Eclipse Theia 1.72 announcement
• Eclipse Theia 1.73 announcement
• Eclipse Theia 1.74 announcement

Build or Extend Your Own IDE or Tool with Eclipse Theia

If you are interested in building custom tools or IDEs based on Eclipse Theia, EclipseSource provides consulting and implementation services for Eclipse Theia, for AI-powered tools and IDEs, for Tools or extensions based on VS Code as well as for web-based tools in general. We also provide a Eclipse Theia online training that you can take anytime for a running start with Theia extension development. If you want to extend Theia with features or bug fixes, EclipseSource provides sponsored development for Theia, too. Finally, for cloud deployments we offer consulting and support for web-based tools.

👉 Contact us to find the perfect (AI-driven) solution for your projects!

👉 Subscribe to our YouTube channel.