Skip to main content
On cloud, every tool takes a repo argument naming one of your repos, and the full surface below is available. On a local stdio server (Run locally) repo is a path (or omit it and set LETTUCE_DB), and the default surface is just understand + read_snippet — set CODEWAZE_MCP_FULL=1 to expose everything else.
Every tool takes a repo argument naming one of your repos. Each call returns compact rows with id, kind, qualified_name, path, line_start / line_end, signature, and summary fields. The relationship tools (neighbors, callers, …) also tag rows with the edge_kind (calls, imports, contains, inherits, references) and direction that led there.

​
High-level tools (start here)

Most agents only need these three.
​
Returns candidate symbols, code, callers, callees, related docs, and a confidence score (0..1) for task in repo. Answers most questions in one call.
​
Returns ranked candidate symbols (with code) for task in repo, plus a confidence level (high / medium / low / none).
​
Returns the top candidates in repo for query. Each carries file:line, signature, code, callers, and callees.

​
Finding code

​
Search symbols in repo matching query. Matches across names, signatures, summaries, and paths — button also matches IconButton.
​
Find symbols by exact name in repo (or LIKE pattern when fuzzy=true).
​
Return the nodes with the highest incoming-edge count in repo.

​
Following relationships

​
Nodes that call this node, and nodes called by it.
​
Modules imported by this node, and nodes that import it.
​
1-hop neighbors of a node. Each row carries the edge_kind and direction that led there.
​
Shortest path between two nodes, or {"path": null}.
​
Nodes transitively reachable from a node along the given edge kinds. Each row carries its depth.
​
Bounded subgraph around a node.

​
Reading source

​
A single node row: file:line, signature, docstring first line.
​
Read source. Pass a path (optionally with line_start/line_end or around a line), a name, or a node_id from an earlier result. Lines are line-numbered; ranges longer than max_lines are truncated.

​
A typical flow

For a request like “add rate limiting to the /login endpoint”:
  1. understand(repo, "add rate limiting to the /login endpoint") → candidate symbols with code, callers, callees, and related docs. Often enough to answer.
  2. If you need to see more of the file: read_snippet(repo, node_id, …).
To drive the navigation by hand:
  1. search(repo, "login") → the endpoint, with file:line.
  2. callers(repo, id) → who routes and tests it.
  3. callees(repo, id) → what it already depends on.
  4. search(repo, "rate limit", fuzzy=true) → locate the helper.
  5. find_path(repo, ...) → confirm it isn’t already wired in.
  6. read_snippet(repo, node_id=id) → read just those lines, then edit.