parecode

An MCP server that gives coding agents context-window-aware search and safe, atomic multi-file edits — built to cut token usage on large codebases without giving up correctness.

Requirements

• Node.js 20 or newer (ESM, native test runner, stable fetch-free runtime).

• ripgrep on PATH (rg on Linux/macOS, rg.exe on Windows). Install via your package manager:

  • macOS: brew install ripgrep

  • Debian/Ubuntu: apt install ripgrep

  • Windows: winget install BurntSushi.ripgrep.MSVC or choco install ripgrep

• A supported MCP client (Claude Code is the reference target).

Parecode does not bundle ripgrep — it shells out to the system binary so you stay on a single, audited version.

Related MCP server: LocalNest MCP

Install

npm install -g parecode

Pure JavaScript — no native dependencies, no C/C++ toolchain required.

Quick start

Register the server with Claude Code:

parecode init                       # user scope; installs MCP + SessionStart hook + parecode-explore plugin (defaults)
parecode init --scope project       # commit MCP config + hook to the repo
parecode init --no-hook             # register MCP only; skip the SessionStart hook
parecode init --no-plugin           # skip the parecode-explore Claude Code plugin
parecode init --print               # print the equivalent command without running it
parecode init --remove-hook         # remove the SessionStart hook (MCP stays registered)
parecode init --remove-plugin       # uninstall the parecode-explore Claude Code plugin

The SessionStart hook injects a short directive at the start of each session telling Claude to prefer ParecodeSearch / ParecodeEdit over the equivalent native tools. Without it, Claude's first-party Grep / Read / Edit tools typically win by default and Parecode's token savings never land. The hook payload is a static string; parecode hook session-start prints it. Pass --no-hook if you would rather opt in explicitly per session via your own tooling.

The bundled parecode-explore Claude Code plugin adds a read-only subagent (pinned to Haiku, given only ParecodeSearch) and a matching skill, so exploration-style questions ("where is X?", "how does Y work?", "find all usages of Z") get answered in a cheap, isolated context window instead of burning tokens in your main session. init registers a local marketplace pointing at the npm-installed copy and runs claude plugin install parecode-explore@parecode. If your Claude Code build doesn't support the plugin subcommand the step soft-fails with a warning and the rest of init still succeeds; pass --no-plugin to skip it entirely, or --with-plugin to make any plugin-step failure hard-fail.

Then in any session, the ParecodeSearch, ParecodeExpand, and ParecodeEdit tools become available. Run parecode doctor to confirm registration, hook status, and .codegraph/ pairing if present.

What it does

• ParecodeSearch — ripgrep-backed search that returns matches with surrounding context windows in a single call, with per-file byte chunking so large result sets do not blow up your context.

  • pattern accepts a single string or an array of strings; arrays dispatch parallel ripgrep runs sharing the same paths / contextLines, and each match carries a patterns: string[] field listing which input patterns contributed. One call replaces N back-to-back greps for related-keyword flow tracing.

  • Overlapping or adjacent windows within the same file are merged automatically (gap ≤ contextLines), with bridging lines loaded from disk.

  • Per-match and response-level estimatedTokens are returned so the agent can self-budget before consuming results.

  • Opt-in relatedSymbols: true surfaces likely event-flow neighbours (Handle<X>, On<X>, <X>Handler/Listener/Closed/Completed/Started) discovered in each match, capped at 10.

  • Omitted line ranges are reported so the agent can widen with ParecodeExpand without re-reading the whole file.

• ParecodeExpand — widen a known (file, startLine, endLine) range with optional contextBefore / contextAfter padding. Designed as the natural follow-up to a ParecodeSearch match. Returns the same estimatedTokens shape so the same self-budgeting heuristic applies. Prefer this over a full-file Read after locating a line.

• ParecodeEdit — batched multi-file edits with whitespace-tolerant fuzzy matching (and an opt-in Unicode-lookalike mode), pre/post stat conflict detection, and atomic same-directory rename writes. Cross-file edits run in parallel.

• parecode stats — local JSONL session log with token-saved estimates. Zero network. Zero telemetry.

Measured savings

On search-and-edit tasks — finding call sites, multi-file refactors, "do X to every Y" — Parecode cut cost ~40% and assistant turns ~75–83% in matched A/B tests:

Method: the identical task run with Parecode on vs off, a fresh session per run, n=3 per arm with alternated order, Sonnet 4.6. Savings come from collapsing many Grep / Read / Edit round-trips into single ParecodeSearch / ParecodeEdit calls — so the win scales with how much searching and multi-file fan-out a task involves, and shrinks toward zero on single-file or reasoning-heavy tasks. These are measured per-session token and cost numbers, not the estimates in the scan below.

Retroactive Savings Scan

Curious how much Parecode would have saved you if you had installed it earlier? You can scan your past Claude Code sessions:

parecode stats --retroactive --since 30d

Sample output:

Parecode — last 30d (retroactive scan)
─────────────────────
Sessions:                   42
Tool calls:                156
Calls batched (est):        89
Tokens saved (est):  1,200,000

* Note: Retroactive savings are estimated, not measured.

Privacy disclaimer: This scan runs entirely locally against Claude Code's session transcripts (~/.claude/projects/**). By default, it parses only structured fields (tool names, paths, patterns, and token counts). It does not send any data over the network. The --include-content flag (which allows reading tool input/output) is strictly opt-in and loudly flagged if used.

Privacy

Parecode performs no network calls at runtime. Session logs are written to your OS data directory (resolved via env-paths) with 0600 permissions on Unix. Logs are self-maintaining: each server start prunes session data older than 30 days and caps the envelope log at 5 MB. Prune earlier with parecode prune <days> or wipe the data dir. Update to the latest release (and refresh hooks/plugin) with parecode update.

License

MIT