jCodeMunch indexes the symbols in a large codebase, or in several repos at once, so your team and its agents can ask where something lives, why it's written that way and what depends on it. A local folder indexes up to 2,000 files by default; for a bigger repo, raise max_folder_files in ~/.code-index/config.jsonc.
search_symbols, get_repo_outline).get_symbol_provenance).get_cross_repo_map, get_group_contracts). When you ask for it, who imports a file and what a change breaks reach across those repos too (find_importers, get_blast_radius with cross_repo).get_blast_radius, find_references).get_pr_risk_profile).On our benchmark, a lookup returned 96.2% fewer tokens than grep-and-read (15 lookups on express, fastapi and gin at pinned commits, cl100k tokenizer). The method and raw data
When to adopt it. When the codebase grows past what grep and glob handle comfortably: the agent reads several whole files to answer one question, sub-agents open the same files again, or the code spans more than one repo.
Why it can be the one primary. It runs locally over stdio, starts no language servers, and keeps its index in ~/.code-index/, not in your repo. A new install shows the agent a six-tool front door. It's read-only: it plans renames and checks deletes, and your agent makes the edits. A Studio license (up to 5 developers) or a Platform license (the whole organization), paid once, covers the team. See the terms and the security answers.
Run your own A/B before you standardize. Take 10 to 20 real tasks from your backlog and run each with and without jCodeMunch, on the same model and settings. Compare what you care about: tasks finished, time, tool calls and tokens. Make sure the agent knows the index is there (jcodemunch-mcp init --policy writes a short policy into CLAUDE.md), because a server the agent never calls can't change the result. The 30-day guarantee covers the trial.
In one outside test on a real Vue+Firebase production codebase, the agent completed 80% of runs with jCodeMunch vs 72% without, and timed out in 32% vs 40% (one codebase, 25 runs per arm, Sonnet 4.6, the March 2026 version). Most developers don't personally pay the bill. They care that the agent finishes the ticket, and that a senior engineer doesn't have to check every refactor by hand.
jcodemunch-mcp init on every dev machine. Auto-detects Claude Code / Cursor / Windsurf, writes config, installs hooks, indexes the repo. One command per machine. Or run one shared copy over HTTP on a host you control (jcodemunch-mcp serve --transport streamable-http), behind your own HTTPS reverse proxy, with an optional bearer token everyone shares.
PreCompact and session-memory hooks preserve context across long tasks.
get_blast_radius, get_pr_risk_profile, check_rename_safe, get_layer_violations. Senior-engineer judgment, codified into tool calls.
audit_agent_config inspects what your team is already paying for: tool schemas, MCP server bloat, redundant context. Surface the leak before you spend on training.
All three munchers conform to the published Apache-2.0 jMRI spec. No vendor lock-in. Your team's investment in retrieval discipline outlives any one tool.
Name collisions, refactor plans and dead code, read-only (check_rename_safe, plan_refactoring, find_dead_code).
One structural pattern across 70+ languages (search_ast).
Sub-agents can look up a symbol, its callers and its impact.
Not a hobby project: 17 generations of index format.
"Roughly 5× more efficient context retrieval."
"Preserves your context budget for actual reasoning."
"Structural questions you simply can't ask Grep or Glob."
"The whole game is what you choose not to put in the prompt."