MCP & CLI

raph exposes the same capabilities two ways. Prefer MCP when it is connected; fall back to the CLI otherwise.

MCP server

Start the server (stdio):

raph start

Install project config for supported agents automatically:

raph agents mcp setup --path .

This writes the correct config for OpenCode (opencode.json), Claude Code (.mcp.json), Codex (.codex/config.toml), Cursor (.cursor/mcp.json), and Pi (.pi/mcp.json).

Tools

• search — ripgrep-style search (auto/literal/regex/vector, type & scope filters)
• search_codebase, index_codebase
• store_memory, update_memory, deprecate_memory, get_memory_history
• search_project_knowledge, search_shared_knowledge, search_global_preferences
• store_rule, list_rules
• add_document, list_documents, read_document, link_nodes
• crawl_url, crawl_website
• graph_neighbors, graph_neighbors_cross_corpus, best_vector_match

CLI for agents

Every command takes --format json|text. raph defaults to JSON when it detects it is being called by an agent (via the RAPH_FORMAT/RAPH_JSON env or a non-terminal stdout), and to text in an interactive terminal. So an agent can just run:

raph search "auth middleware" --type func
raph mem search "deployment" --scope project
raph rules list --all
raph doc list --type handoff

and parse the JSON, with no flags required. Humans get readable output automatically.