Last updated: 2026-08-22 (v0.24.2). This file is refreshed at each release.
mcp-debugger gives AI agents step-through debugging over the Model Context Protocol: 28 tools across eight language adapters (Python, JavaScript/TypeScript, Ruby, Rust, Go, Java, .NET, C/C++), plus a mock adapter for testing. This roadmap answers two questions we hear from people evaluating the project: is the tool surface stable enough to build on? and what's left before 1.0?
1.0 is defined by criteria, not dates. We cut 1.0 when all of the following hold:
- Tool-schema freeze. The 28-tool surface and its response shapes are declared stable under SemVer: breaking changes to tool names, parameters, or response shapes require a major version. Prerequisite work: complete full parameter/response documentation in the tool reference for the four tools currently summarized without schemas (
list_supported_languages,attach_to_process,detach_from_process,list_threads). - Per-platform CodeLLDB packages published and verified. The five
@debugmcp/codelldb-*platform packages (which let npm installs pull only the native debug engine for the current platform) are live on npm and verified across install channels (npx, global npm, Docker). - A full release cycle with no breaking behavioral changes. One complete minor release soaks with no regressions that change documented tool behavior.
- Supply-chain milestones. OpenSSF Best Practices Silver, and build-provenance attestations covering every distributed artifact — npm tarballs (done), the Docker image, and the PyPI launcher (#422).
- Environment self-check — ✅ shipped: the
mcp-debugger doctorcommand checks every adapter's runtime prerequisites in one pass, and the diagnostics guide consolidates prerequisites and failure signatures (#423). - Turnkey Kubernetes debugging — ✅ shipped: the Kubernetes debugging recipe with registry-free example manifests and per-language attach presets, covering ephemeral sidecar containers and port-forwarded attach (#424).
- Published-artifact canary — ✅ shipped: the Canary workflow runs a weekly (and on-demand, as the release gate) install-and-debug matrix over what users actually install (npx, global npm, Docker) across x64/arm64 Linux, arm64 macOS, and Windows, catching packaging regressions before users do (#425).
To set expectations, mcp-debugger is deliberately not heading toward:
- An IDE-replacement UI. This is infrastructure for agents (and the humans supervising them), not a visual debugger. The
expose_sessionDAP mirror exists so your IDE can watch a live session; that's the extent of the UI ambition. - Languages without a maintained DAP backend. Adapters wrap existing, actively maintained Debug Adapter Protocol implementations (debugpy, js-debug, CodeLLDB, Delve, …). We don't write or maintain debug engines.
- Editing or refactoring tools. This is a debugger. Code modification belongs to other tools in the agent's kit.
If a 1.0 criterion above seems wrong or missing, open an issue — the criteria are meant to be argued with.