molmcp
Allows indexing and querying capabilities from GitHub repositories, enabling LLM agents to discover code structure, symbols, and relationships via the discovery engine.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@molmcp find a function that downloads a file from URL"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
MolMCP
Multi-plane MCP for the MolCrafts ecosystem.
Protocol: MCP 2026-07-28 via FastMCP 4.0.0b5 (+ MCP Python SDK v2). Handshake-era clients still work — FastMCP 4 negotiates per connection.
Optional science packages (molvis, molq, molexp, …): if not installed,
that plane is omitted from catalogs and client configs (silent). Explicit
molmcp serve <plane> still errors with an install hint. This is runtime
behavior — not a test skip.
molmcp serve (no plane) starts the molcrafts core and FastMCP-mounts
enabled providers into that one process (molvis_open, molq_list_jobs, …).
molmcp init <host> writes that one MCP entry and the usage skill.
--disable molcrafts errors; --disable molq omits that mount.
Command | Role |
| Composed core + provider mounts |
| Debug: vis-only process, bare |
| User-level skill + MCP JSON |
Science APIs are never MCP tools. Discover them on molcrafts (packages →
open), then call them from agent Python or molvis_exec.
Client config (default: everything)
One standard mcpServers JSON, which every host reads — Claude Code and
Cursor natively, Grok alongside its own config.toml.
molmcp init grok # skill + composed serve
molmcp init grok --disable molq --disable molexp
molmcp init grok --disable molq --enable molq # re-enable after a disable
molmcp init claudeHost is required (grok, claude, cursor, codex). JSON is one
molcrafts entry running molmcp serve, with --disable flags for omitted
mounts. Tool ids look like molcrafts__molvis_open.
In Grok,
~/.grok/config.tomloutranks the JSON sources. If an old molmcp entry lives there it still wins —grok inspectshows each server's origin.
Related MCP server: OpenCodeHub MCP Server
Configuration
Settings live in ~/.molmcp/settings.json, edited through the CLI. There are
no environment variables.
molmcp config list # resolved settings + layers
molmcp config set sources.molpy pkg:molpy # index a package
molmcp config set indexWorkspace true --project # index this repo too
molmcp config add excludes vendor
molmcp config remove sources.molpyA project may carry .molmcp/settings.json (checked in) and
.molmcp/settings.local.json (untracked); both layer over the user file.
Writes go to the user file unless --project / --local is passed, because a
plane server inherits its working directory from whichever client launched it.
What gets indexed. Auto-discovery finds installed MolCrafts distributions.
The working directory is not a source unless indexWorkspace says so — it
used to be, which meant an unconfigured install indexed whatever it happened
to be started next to.
Key | Meaning |
| Extra sources to index, |
| Index the working directory as well (default |
| Narrow which indexed sources the knowledge tools surface |
| Extra ignore globs for the file walk |
| Where the index lives and how big it may get |
| Environment to auto-discover from (a venv root, python, or site-packages) |
| Force a distribution in or out of auto-discovery |
| Provider-specific paths |
molcrafts.json is no longer picked up from the working directory; pass
--config PATH if you keep one.
CLI
uv run molmcp planes # list planes
uv run molmcp init grok # skill + MCP config
uv run molmcp config list # resolved settings
uv run molmcp route "draw dopamine"
uv run molmcp serve # composed core + mounts
uv run molmcp serve molvis # debug one plane
uv run molmcp search "Conformer" # offline index search
uv run molmcp index
uv run molmcp cache # index size; --prune / --gc / --vacuum to reclaimInstall
uv sync --extra dev
uv run pytest -vDesign rules
FastMCP composition —
molmcp serveis molcrafts + namespaced mounts.Bare register, namespaced mount — a provider registers
open; the stack exposesmolvis_open. Debugmolmcp serve molvisstill showsmolvis__open.No science tool mirror — no
show_smiles/draw_dopamine; discovery + Python.Providers register via
molmcp.providersentry points.No environment switches — configuration is settings and CLI flags, so
molmcp config listis the whole truth.
Documentation
Full manual: docs.molcrafts.org/molmcp
(sources in docs/):
Local sources: docs/concepts/architecture.md, docs/guides/molvis-workbench.md.
Available Tools
9 toolscomposeComposeBRead-onlyIdempotent
Bind packages + suggest + explore/open pages into one budgeted pack.
sources is intersected with knowledgeScope when set.
Inputs: refs, task, sources, budget_chars. Returns: object
infoInfoARead-onlyIdempotent
Ops/health view of sources and registry (not the main discovery path).
Inputs: workspace. Returns: object
list_planesList PlanesARead-onlyIdempotent
List the core connection and optional provider planes.
Each row has id, serve_command, when_to_connect,
tools_hint, and disableable. molcrafts is always on;
only provider planes can be dropped from a client config.
Inputs: no named parameters. Returns: object
openOpenARead-onlyIdempotent
L2 symbol page: signature, doc, examples, tests (inject before coding).
Miss → ok=false / SYMBOL_NOT_FOUND. Empty examples are honest zeros.
Inputs: ref (required), include_source. Returns: object
outlineOutlineBRead-onlyIdempotent
L1 module directory for one source (optional path prefix).
Inputs: source (required), path, top_symbols_limit. Returns: object
packagesPackagesARead-onlyIdempotent
L0 directory page: every package + summary for context injection.
Read the markdown (or data.packages[].summary) and choose sources yourself — this is a catalog, not a ranking.
With knowledgeScope set, only those packages appear.
Inputs: no named parameters. Returns: object
routeRouteARead-onlyIdempotent
Which optional provider plane(s) to connect for task.
Routing only — no science. molcrafts is already this connection. Do not invent domain MCP tools for chemistry APIs.
Inputs: task (required). Returns: object
searchSearchCRead-onlyIdempotent
Index helper: find refs (prefer after packages/outline, with source=).
Source symbols are evidence only. executable=true only for Molexp bind.
sources is intersected with knowledgeScope when set.
Inputs: query (required), mode, path, kinds, +3 more. Returns: object
suggestSuggestARead-onlyIdempotent
Optional shortcut: which package pages to read for task.
Inputs: task (required). Returns: object
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 9 tools
The main exploration tools are explicitly layered (packages/outline/open), and info/list_planes/route are separated by role. Search, suggest, and compose have closely related purposes, but their descriptions distinguish ref-finding, package-page recommendation, and bundling, so an agent can usually select the right one.
Tool names mix bare nouns (info, packages, outline, open) with verbs (compose, search, suggest, route) and one snake_case verb_phrase (list_planes). There is no consistent verb_noun convention or common prefix, making the surface feel ad hoc despite the underlying hierarchy.
Nine tools is well within the ideal scope for a server of this complexity. Each tool has a distinct function in the discovery, search, composition, health, and routing workflow, with no obvious redundancy.
The browse (packages/outline/open), search, suggest, compose, health, and plane-routing flows form a complete read-only discovery workflow. The only minor gap is the lack of a raw package/registry detail view beyond info and open, but agents can still accomplish the core task.
Maintenance
Related MCP Connectors
Hosted code graph over MCP: exact callers, dependencies, and cross-repo blast radius for AI agents.
Codebase knowledge graph — architectural awareness for AI coding agents via MCP
Code intelligence for coding agents: semantic, AST, graph, and full-text search. 279+ languages.
Shared product-knowledge graph for AI agent fleets — vault tools + agent self-signup over MCP.
Related MCP Servers
- AlicenseNot gradedqualityAmaintenanceSemantic code indexer with GraphRAG knowledge graph. Index your codebase, search in natural language, and expose everything via MCP so AI agents understand architecture — not just files.523Apache 2.0
- AlicenseNot gradedqualityCmaintenanceProvides code intelligence for AI coding agents by indexing repositories into a hybrid knowledge graph, enabling agents to query dependencies, impact, and context through 28 MCP tools.4Apache 2.0
- FlicenseNot gradedqualityCmaintenanceEnables coding agents to query a local-first code intelligence graph of Python repositories—covering functions, classes, modules, and their relationships—via MCP, supporting subgraph retrieval, caller lookup, and impact analysis without re-reading the codebase.-
- AlicenseNot gradedqualityCmaintenanceEnables AI coding agents to query a local, continuously updated symbol graph of a codebase, providing ranked search, caller/callee exploration, dependency paths, and git-diff impact analysis through MCP tools.10 npmApache 2.0