The Pulse
JetBrains MPS 2026.2 adds tools for coding agents
JetBrains released MPS 2026.2 with expanded tools for coding agents and support for Linux ARM64. The company says changes to its agent toolkit reduced token usage by more than 50% in its measurements with Claude Code.
JetBrains has released MPS 2026.2 with expanded coding-agent tools for its language workbench, according to the company's announcement. The toolkit gives agents access to MPS projects through a Model Context Protocol server and a catalog of MPS-specific skills.
JetBrains says its measurements using Claude Code with Opus 5 and 5.5 showed token savings exceeding 50%. JetBrains says the savings apply to common tasks such as domain-specific language modeling and implementing language definitions. The release also adds Linux ARM64 support, alongside changes to packaging, the interface and editor tests.
More operations for agents
The toolkit's type-system tool lets agents compare types, determine the expected type for a role and trace type calculations. It can also help them debug type-system rules and identify the rule responsible for a reported type-system error.
Other MCP tools let agents invoke intention actions and quick-fixes and keep their skill databases current with the instance they connect to. Agents can also read and filter MPS logs and open, close or create projects. Another tool automatically corrects SModel expressions in parsed Java statements, supporting work on concept functions and behavior methods.
The toolkit's bundled plugin is enabled by default and requires an open project and an MCP-aware coding agent. JetBrains had already shipped substantial toolkit improvements in MPS 2026.1.2 and the 2026.2 RC2 release candidate. The company said those versions received the improvements planned for the 2026.2 release.
Changes to agent requests
The revised MPS Model Context Protocol API uses more consistent parameter names, handling of null and default values, and error reporting. It also provides more detailed tool descriptions and uses bulk operations with compact summaries where they are most useful.
Short read results now appear directly in responses, while larger dumps are written to files when needed. The API also adds validation of full models in a single operation, lightweight metadata views for discovering concepts, and checks that reject invalid search scopes and mutation inputs.
Skills include helper scripts, and project targeting has been clarified for situations in which several projects are open. The toolkit documentation notes that its APIs, tool names, prompts and bundled skills may change between releases.
Platforms, packaging and tests
The release supports standalone MPS-based applications on Linux ARM64, and the Build Script wizard includes that platform in generated build scripts. Build Language can import archive contents directly into distribution layouts. Plain TAR files and archives compressed with gzip or bzip2 are supported.
When files are imported into a TAR layout, symbolic links and permissions are preserved. MacOS archive distributions now use.tar.gz instead of.zip, and the wizard generates the same format for standalone macOS applications.
In the interface, Find Usages results can be grouped by virtual package, with folders reflecting the package of each root node. Grouping by virtual package is enabled by default in Find Usages; other tool windows retain their usual defaults. The Import dialog now respects the models and languages selected when code is pasted between models.
Editor tests can use custom node-matching rules to specify which differences to compare or ignore, or apply a marker to an individual node. Failure messages now identify the relevant node and cell and include each mentioned node's presentation, concept name and node ID.