Seshat MCP

What seshat-mcp MCP server does

The seshat-mcp MCP server turns a repository into a compiled representation of its code structure. The graph covers functions, classes, routes, tables, dependency edges, data flow, constraints, and selected history signals. Agents can use that representation to understand how a symbol is used, what a change may affect, and which files are most relevant before making an edit.

The server is intended for repository investigation rather than general file search. It can identify public entry points, locate entities by name or organizational attributes, trace callers and callees, and report symbols that appear unreachable. Its answers include coverage information, and its history analysis describes commit-resolution correlation without presenting correlation as causation.

How it works

A repository is synchronized into the service and compiled into a typed symbol graph. The agent then queries that graph through MCP tools. Public repositories are cloned from GitHub; private repositories require a connected GitHub account. Source is processed in Papyrus Labs cloud infrastructure, where the graph is cached for subsequent queries.

The workflow can start with project and module discovery, continue with symbol or entry-point queries, and then move into dependency, data-flow, or blast-radius analysis. For JavaScript, TypeScript, and Python call sites, the server can distinguish states such as unawaited asynchronous calls, promises used as conditions, failures without surrounding error handling, and dropped results.

History data comes from the repository’s commit record and is backfilled during the first synchronization. Lineage describes changes to a function’s body, calls, signature, or data, along with CI outcomes, reverts, related changes, and the last recorded change that forced an update. Hotspot and co-change queries provide additional historical context.

Setup and configuration

Installation uses the published npm package and a setup command:

npx -y @papyruslabsai/seshat-mcp setup <your-key>

The setup process writes MCP configuration and stores the API key. A free key is available from the Seshat service, and the first extraction is free. Manual configuration can run the package with npx and provide SESHAT_API_KEY in the server environment:

{
  "mcpServers": {
    "seshat": {
      "command": "npx",
      "args": ["-y", "@papyruslabsai/seshat-mcp"],
      "env": { "SESHAT_API_KEY": "your-key" }
    }
  }
}

The README names Claude Code and Cursor as manual-configuration examples, while also stating that any MCP client can be configured this way.

Tools and capabilities

The seshat-mcp MCP server exposes tools for several investigation stages:

• list_projects and list_modules show synchronized projects and code organization.
• query_entities and find_entry_points locate symbols, routes, exports, and public API surfaces.
• get_entity, get_dependencies, get_data_flow, and get_blast_radius explain relationships, behavior, and transitive change impact.
• get_optimal_context ranks the smallest useful set of files to read before editing.
• find_by_constraint searches for functions associated with a table or trait, while find_dead_code identifies unreachable symbols.
• query_call_sites checks per-call states in JavaScript, TypeScript, and Python.
• get_lineage, get_hotspots, and get_co_change_clusters use commit history to show evolution, unstable areas, and files that commonly change together.

Limitations and notes

Analysis runs in the Papyrus Labs cloud rather than locally. Public repositories are cloned from GitHub, private repositories need GitHub account access, and processed source is cached to serve queries. The README says cross-cutting audit tools for test coverage, topology, and semantic clones are not yet part of the available tool list.

Usage beyond the free tier is priced per query, with a stated typical investigation range of five to fifteen queries. The repository describes the server as MIT licensed.