XERJ MCP Server: Setup for Claude, Cursor and Other Apps

• Searches your indexed code and docs

  Keyword search over an index of your files. Plain words rank the file that defines a symbol above files that only mention it, and results give the file and line.

  Try: “Where do we rotate the WAL segment?”

  Read · xerj_search

• Finds where a symbol is defined

  Looks up a symbol in your locally indexed data and returns the declaration with its file and line.

  Try: “Where is euler_to_rotationmatrix defined?”

  Read · xerj_def

• Runs up to 50 searches at once

  Searches 1 to 50 identifiers or questions in one call and returns results in input order, with citations. It retrieves evidence and does not write answers.

  Try: “Look up these 20 function names and cite the files.”

  Read · xerj_batch_search

• Shows how other projects solved a problem

  Searches a corpus of other codebases and returns the matching definition, the file and line to cite, and a licence warning for each hit.

  Try: “How did other projects implement a retry policy?”

  Read · xerj_code_search

• Finds text by meaning

  Searches a semantic field by meaning. The server embeds your query itself, with no outside API key. The default embedder is lexical, not neural.

  Try: “Find notes about slow database queries.”

  Read · xerj_semantic_search

• Finds near matches to a vector

  Runs a nearest-neighbour search on a vector field using a vector you supply.

  Try: “Find the 5 nearest items to this embedding.”

  Read · xerj_vector_search

• Combines keyword and meaning search

  Fuses several sub-queries, such as keywords plus a vector, into one ranked list.

  Try: “Search for payment errors using keywords and meaning together.”

  Read · xerj_hybrid_search

• Turns a plain request into a query

  Turns a plain-words filter into a validated search query you can send to the search tool. It is rule-based, not an AI model.

  Try: “Build a query for events deeper than 100 km.”

  Read · xerj_plan

• Lists the fields of each indexed folder

  Reads the local index catalog and lists each index with its record count, time range and field names, so the assistant stops guessing field names.

  Try: “What fields can I query in my indexes?”

  Read · xerj_map

• Checks that the XERJ node is up

  Reports whether the node is running, its health, the number of indexes and the first 20 index names.

  Try: “Is XERJ running, and what is indexed?”

  Read · xerj_status

• Estimates an index before building it

  Counts the files and bytes in a folder by type and compares them with free disk space, before anything is indexed.

  Try: “How big would indexing ~/projects be?”

  Files · xerj_index_budget

• Deletes one indexed folder

  Deletes the indexes and catalog rows for a folder you indexed. Without confirm set to true it returns the plan and deletes nothing.

  Try: “Remove the index of my old-notes folder.”

  Delete · xerj_autoindex_remove

• Saves a memory for later

  Stores a durable memory in a namespace, searchable by words and by meaning. It can skip near-duplicates.

  Try: “Remember that we chose Postgres for the billing service.”

  Write · xerj_memory_store

• Recalls saved memories

  Finds the most relevant memories in a namespace by text, meaning, or both.

  Try: “What did we decide about the billing database?”

  Read · xerj_memory_recall

• Summarizes a knowledge graph

  Reports whether a second-brain graph exists, how many links are live or retired, its most linked nodes and a timeline.

  Try: “Give me an overview of my project brain.”

  Read · xerj_brain_overview

• Shows one node and its links

  Returns one node's neighborhood in the graph, with each link's direction, type, confidence and supporting quote.

  Try: “What is linked to the auth service node?”

  Read · xerj_brain_ego

• Adds a link between two nodes

  Records a link between two nodes in the graph along with the quote that justifies it. Repeating the same link changes nothing.

  Try: “Link the payments service to the ledger with this quote.”

  Write · xerj_brain_link

• Retires a link in the graph

  Marks a link as no longer valid. It stays in the history and can still be seen at earlier dates.

  Try: “Retire the link between the old API and the billing service.”

  Write · xerj_brain_unlink

• Sends a report about XERJ

  Files a field report about using XERJ. It opens a GitHub pull request on the XERJ repo when it can, and otherwise posts to xerj.org. It sends by default.

  Try: “Write a field report on how XERJ worked for this task.”

  Write · xerj_field_report

• Rates a hub corpus

  Sends a structured rating of a corpus on hub.xerj.org after you use it. Query text is reduced to its shape first. It sends by default.

  Try: “Rate the corpus you just used.”

  Write · xerj_corpus_feedback

• Finds hub corpora that fit a question

  Lists hub corpora that could answer a question, with their upstream repo, licence and whether each is already on your machine.

  Try: “Which hub corpora cover the Rust tokio runtime?”

  Network · xerj_hub_find

• Looks up a definition in a hub corpus

  Looks up one definition, with its signature, file and line, in a named hub corpus. It never downloads a corpus. It also sends an automatic review to hub.xerj.org by default.

  Try: “In the redis corpus, where is dictAdd defined?”

  Network · xerj_hub_def

• Asks a hub corpus a question

  Answers a question from hub corpora, using packs on your machine first and hosted ones second. Without a corpus name it picks one.

  Try: “How does the redis corpus handle key expiry?”

  Network · xerj_hub_ask

Watch out for

• It can delete indexed folders

  One tool deletes the indexes for a folder you indexed. It deletes nothing unless confirm is true, so read the plan it shows before the assistant confirms. github.com

• One tool sends reports by default

  The field report tool posts to xerj.org or opens a GitHub pull request, and its description says it sends without asking. Set XERJ_DISABLE_FEEDBACK=true to turn this off, as the README mentions. github.com

• The dev node has no password

  The README starts the node with --insecure, which turns off encryption and auth, so every caller can do everything. Keep the node on localhost or turn auth on. github.com

• Hosted hub answers leave your machine

  Hub questions use local packs first and hosted ones second. Do not ask about private topics. github.com

Security advisories