A Model Context Protocol server providing conda eco system context
The server provides 8 well-defined tools with generally good descriptions and complete JSON Schema input definitions. Tool names follow verb-noun conventions appropriately. However, there are notable gaps: (1) output schemas are undocumented in the visible source code, while descriptions mention return structure (e.g., ImportMappingResult, InfoResult TypedDicts), the actual JSON Schema definitions are not visible in the provided code, making it impossible to verify LLM-parseable output structure; (2) parameter descriptions are good but some lack constraint details (e.g., no min/max for 'limit' parameters, no enum for 'file' options in package_insights); (3) error handling guidance is absent, tools do not document what exceptions might occur or how to recover; (4) no tool annotations are present (readOnlyHint, destructiveHint, idempotentHint). The descriptions are generally well-written (100-300 chars typical) and parameter counts are reasonable (2-6 per tool). All tools have clear, actionable descriptions that explain WHAT they do and WHEN to use them.
Run cache maintenance for all registered external and tool-level caches. Returns a short status message after cleanup has been triggered.
Provides the full help text for the given tool including all subcommands and options. To be used to answer advanced CLI questions beyond the knowledge cutoff of models, to e.g. help with new features that recently landed in the tool. Args: tool: str = "conda" limit: max number of lines returned (0 means all) offset: number of initial lines skipped grep: Regular expression pattern to filter help lines (case-insensitive). Empty string returns all lines (default). Example: "install|update|create" returns lines matching any of these. Reduces context by ~90% for targeted queries. Returns: A string with the help text
Find conda artifacts that contain an exact file path. Searches package path metadata for packages that ship the specified path. Requires the full path as it exists within the package (exact match only). Args: path: The exact file path to search for. Must include directory prefix. Examples: "lib/libssl.so.3", "bin/python", "include/openssl/ssl.h" Note: Partial matches and wildcards are NOT supported. channel (str): Package channel to search, e.g. "defaults", "conda-forge", "bioconda" limit: Maximum number of results to return (0 means all) offset: Number of results to skip before applying limit Returns: dict with: - query_path: the searched path - artifacts: paginated list of artifact names - count: number of artifacts in the returned page - total: total number of matching artifacts - limit: limit used in this query - offset: offset used in this query
Output schemas are not formalized in JSON Schema, tools describe return structures textually (TypedDict names, dict keys) but the actual JSON Schema definitions are not visible in source code. LLMs cannot reliably parse unstructured descriptions of return values, risking hallucination and incorrect downstream tool chaining.
Enum constraints are missing for parameters with known valid values. 'channel' appears in 7 tools but is never declared as enum (should be: defaults, conda-forge, bioconda, nvidia, etc.). 'platform' in package_search (should be: linux-64, linux-aarch64, osx-64, osx-arm64, win-64). 'file' in package_insights (should be: some, all, list-without-content). Free-form strings invite hallucinated values.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | C | 62 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 47 | - | v1 |
Map a (possibly dotted) Python import name to the most likely conda package and expose supporting context. What this does: - Normalizes the import to its top-level module (e.g. "numpy.linalg" -> "numpy") - Retrieves an approximate candidate set of conda packages that may provide it - Applies a heuristic to pick a single "best" package - Returns a structured result with the decision rationale Heuristic labels: - identity: No candidates known; fallback to normalized import - identity_present: Candidates exist AND the normalized import name is among them - ranked_selection: Best package chosen via ranked hubs authorities ordering - fallback: Best package not in candidates (unexpected edge case) Returns: dict with structure matching ImportMappingResult TypedDict: - query_import: original query string supplied by caller - normalized_import: top-level portion used for lookup - best_package: chosen conda package name (may equal normalized_import) - candidate_packages: sorted list of possible supplying packages (may be empty) - heuristic: one of the heuristic labels above Args: import_name: Import string, e.g. "yaml", "matplotlib.pyplot", "sklearn.model_selection". channel (str): e.g. "defaults", "conda-forge", "bioconda", "nvidia" get_keys: Comma-separated field names to include in results. Empty string returns all fields (default). Example: "best_package,heuristic" returns only key fields.
Display information about the MCP instance (versions and pixi environment path) Can be compared with local output of "conda info" to see if they match. Returns: dict[str, Any]: Version information for MCP instance and dependencies. Includes pixi environment path if running in a pixi environment. See InfoResult TypedDict for structure.
Provides insights into a package's info tarball That includes the rendered recipe (meta.yaml) that allows for easy inspection of the package's build process, e.g. see build, host and run time dependencies. Which have big influence on what packages are linked against. The run_exports which end up as run time dependencies for other packages linked against this package. And the about information which contain the remote_url and sha to the repo location where the package recipe is maintained. That helps to open PRs in the right location to fix issues with the recipe. Args: url: The full package URL, e.g. "https://conda.anaconda.org/conda-forge/linux-64/numpy-1.24.3-py311h7f8727e_0.tar.bz2" file: can be set to "some", "all", "list-without-content" or a specific filename limit: max number of lines returned per file (0 means all; ignored for list-without-content) offset: number of initial lines skipped per file (ignored for list-without-content) get_keys: Comma-separated keys to extract from parsed file content (YAML/JSON). Empty string returns full file content (default). Example: "channels,conda_build_version" extracts those fields from about.json. Requires exactly one file (use 'file' parameter). Significantly reduces context by returning only needed fields. Returns: A dictionary with key=filename, value=content or parsed object.
Search available conda packages matching the given package_ref_or_match_spec, channel, and platform. Features: - Results are deduplicated. - Ordered by newest (version, then build_number descending). - limit=1 reliably returns the single newest record. - Supports paging via (offset, limit). - Optional field filtering via get_keys parameter. Args: package_ref_or_match_spec (PackageRef or MatchSpec or str): e.g. "numpy", "numpy>=1.20", "numpy=1.20.3", "numpy=1.20.3=py38h550f1ac_0" channel (str): e.g. "defaults", "conda-forge", "bioconda", "nvidia" platform (str): e.g. "linux-64", "linux-aarch64", "osx-64", "osx-arm64", "win-64" limit (int): Maximum number of results to return (0 means all). offset (int): Number of results to skip before applying limit (for paging). get_keys (str): Comma-separated field names to include in results. Empty string returns all fields (default). Example: "version,build,url" reduces context by ~60-70%. Returns: dict with keys: - results: list of package records (filtered by get_keys if specified) - total: total number of matching packages - limit: limit used in this query - offset: offset used in this query
Map a (case-sensitive) PyPI distribution name to the most likely conda package name. Returns: dict with structure matching PyPiToCondaResult TypedDict: - pypi_name: original input (trimmed) - conda_name: mapped (lowercase) conda name (fallback: pypi_name.lower()) - changed: conda_name != pypi_name.lower() 'changed' is True only when the resolved conda name differs from simple lowercase. Args: pypi_name: PyPI distribution name. channel (str): e.g. "defaults", "conda-forge", "bioconda", "nvidia"
Numeric parameters (limit, offset) lack bounds specifications. No min/max declared for 'limit' or 'offset' across tools. Baseline rubric requires stating 'max number of results' but not the actual constraint (e.g., 'limit: max 1000' would guide LLM not to request 999,999 results).
No error handling guidance. Tools do not document potential failures (invalid URL, corrupt tarball, malformed import name, network timeout) or recovery steps. Descriptions lack statements like 'Returns ToolError if URL is unreachable' or 'Try alternative channel if no results found.'
Tool annotations are completely absent. No readOnlyHint, destructiveHint, or idempotentHint declarations. While all tools appear to be read-only (no mutations), explicit annotations would clarify this for agents and enable safer execution models. Spec requires tool annotations for current-generation MCP servers.
'info' and 'cache_maintenance' have generic names that could be more action-specific. 'info' → 'get_server_info'; 'cache_maintenance' → 'clear_caches' or 'run_cache_cleanup'. Specificity helps LLMs distinguish intent.