Patchloom

One binary. Every platform. Structured file edits for AI agents.

Patchloom is a single-binary CLI that gives AI coding agents safe, structured file editing on any operating system. It edits JSON, YAML, and TOML by selector (not regex), preserves comments, understands code structure across 20 languages, batches multiple file edits into one tool call, and works identically on Linux, macOS, and Windows.

Not a generic filesystem MCP. Default MCP filesystem servers read/write files as text. Patchloom adds dry-run previews, parser-backed config and markdown edits, AST ops, multi-file batch/tx with undo, and stable error_kind peels for hosts. Full coding agents (Claude Code, Codex, Cursor) own the loop; Patchloom is the tool layer they (or a Rust embedder) call.

# Edit a YAML value by selector without breaking comments or formatting
patchloom doc set config.yaml database.port 5432 --apply

# Batch 6 file edits into a single tool call
patchloom batch --apply <<'EOF'
doc.set package.json version "2.0.0"
doc.set config.yaml app.version "2.0.0"
doc.set config.toml project.version "2.0.0"
replace README.md "1.0.0" "2.0.0"
replace CHANGELOG.md "1.0.0" "2.0.0"
file.create VERSION "2.0.0"
EOF

Why Patchloom? | Install | Quick start | Commands | Comparison | When to use what | Architecture | Status

Why Patchloom?

The problem

AI agents edit files through tool calls. Each call is a round-trip back to the LLM. When a task touches config files, that process has three failure modes:

1. Syntax corruption. The agent uses text replacement on JSON, YAML, or TOML and produces invalid output (mismatched braces, broken indentation, lost comments).

2. Round-trip tax. Editing 6 files means 6 separate tool calls. Each one waits for the LLM to generate, execute, read the result, and plan the next call.

3. Platform fragmentation. On Linux the agent uses sed, jq, grep. On Windows, none of those exist. The agent falls back to verbose PowerShell or makes errors with unfamiliar syntax.

How patchloom solves each one

What changes with patchloom

Without patchloom (6 tool calls)

Agent: edit file 1  ─── tool call ───▶  15s
Agent: edit file 2  ─── tool call ───▶  15s
Agent: edit file 3  ─── tool call ───▶  15s
Agent: edit file 4  ─── tool call ───▶  15s
Agent: edit file 5  ─── tool call ───▶  15s
Agent: edit file 6  ─── tool call ───▶  15s
                                    Total: ~90s

With patchloom batch (1 tool call)

Agent: batch with
  all 6 edits     ─── tool call ───▶  25s



                  5 round-trips saved
                                    Total: ~25s

Key capabilities

When to use patchloom vs native tools

Patchloom is not a replacement for all file operations. Its instructions tell agents exactly when to use it and when native tools are faster:

Correctness over speed

Patchloom is not faster than native tools for simple, single-file edits. Use native tools for those. But native text replacement cannot safely edit structured files: a sed on YAML can corrupt indentation, strip comments, or produce invalid syntax. doc set parses the file, changes the value by selector, and writes valid output. That guarantee is the point.

Where patchloom is faster is multi-file batching. Six file edits via native tools means six round-trips to the LLM. One batch call does the same work in a single round-trip.

Task                    PL-CLI    MCP    Native
──────────────────────  ──────  ──────  ──────
search                   18.5s   12.7s   13.9s  ◀ ~same
replace                  36.1s   26.6s   26.1s  ◀ ~same
doc_set                  30.9s   16.9s   13.7s  ◀ native fastest
md_table                 15.5s   13.5s   15.3s  ◀ MCP fastest
tx_multi_file            41.4s   28.5s   22.9s  ◀ native fastest
batch_6_files            50.6s   46.6s   30.3s  ◀ native fastest
batch_mixed_ops          24.7s   13.6s   20.9s  ◀ MCP fastest
yaml_comment_preserve    18.1s   11.6s   16.1s  ◀ MCP fastest
md_insert                15.0s   11.7s   15.7s  ◀ MCP fastest
file_ops                 26.0s   16.6s   17.2s  ◀ ~same
tidy                     45.0s   30.3s   41.7s  ◀ MCP fastest
──────────────────────  ──────  ──────  ──────
TOTAL                   321.9s  228.5s  233.8s

MCP mode wins overall (228.5s vs 233.8s native) because structured tool calls skip shell syntax construction entirely. MCP wins 5/11 tasks; native wins 3/11; 3 are ties. CLI mode is always slowest due to shell construction overhead.

Related MCP server: mcp-json-yaml-toml

Install

# Homebrew (macOS/Linux)
brew install patchloom/tap/patchloom

# crates.io (requires Rust 1.95+, includes MCP server)
cargo install patchloom

# Scoop (Windows)
scoop bucket add patchloom https://github.com/patchloom/scoop-bucket
scoop install patchloom/patchloom

# winget (Windows; PackageIdentifier Patchloom.Patchloom)
winget install Patchloom.Patchloom

# Chocolatey (Windows; community feed; newer versions may lag moderation)
choco install patchloom

# npm / npx (downloads the platform binary from GitHub Releases)
npx patchloom --version
# or: npm install -g patchloom

Pre-built binaries for Linux, macOS, and Windows are on the Releases page. See Installation for shell installer scripts, source builds, and shell completion setup.

• MCP Registry name: mcp-name: io.github.patchloom/patchloom

Editor extension

Install the companion extension for VS Code, Cursor, Windsurf, or VSCodium:

• VS Code Marketplace

• Open VSX Registry

The extension auto-discovers the CLI (or installs it for you), generates AGENTS.md, configures MCP servers, and adds Quick Actions to the command palette. See the Editor Extension guide for details.

Quick start

1. Set up your project

patchloom init

This creates AGENTS.md in a new project or appends the rules to an existing agent instructions file, offers shell completions, and detects MCP configuration opportunities. Pass -y to skip confirmation prompts.

If you only want the rules text:

patchloom agent-rules >> AGENTS.md

# Or tailor the output:
patchloom agent-rules --mode mcp >> AGENTS.md            # MCP-only (no CLI examples)
patchloom agent-rules --platform windows >> AGENTS.md    # Windows-only syntax

If .vscode/ or .cursor/ exists, init also prints ready-to-copy .vscode/mcp.json or .cursor/mcp.json snippets.

Your AI agent reads AGENTS.md and learns when to use patchloom vs native tools.

2. Edit a config file safely

# Parser-backed: changes the value, preserves comments and formatting
patchloom doc set config.yaml database.port 5432 --apply

3. Batch multiple edits into one call

patchloom batch --apply <<'EOF'
doc.set config.json version '"2.0.0"'
md.upsert_bullet AGENTS.md "Rules" "- Always test"
replace src/main.rs "v1" "v2"
EOF

Values are JSON-first: unquoted 2.0 becomes a number. Force a string with nested quotes as above (Unix shells).

Or use a JSON plan with format and validate lifecycle:

{
  "version": 1,
  "operations": [
    { "op": "doc.set", "path": "config.json", "selector": "version", "value": "2.0.0" },
    { "op": "md.upsert_bullet", "path": "AGENTS.md", "heading": "Rules", "bullet": "- Always test" },
    { "op": "replace", "path": "src/main.rs", "old": "v1", "new": "v2" }
  ],
  "format": [{ "cmd": "cargo fmt --all" }],
  "validate": [{ "cmd": "cargo test", "required": true }]
}

patchloom tx plan.json --apply

tx plans are trusted input. format and validate run their cmd fields through the host shell (sh -c on Unix, cmd /C on Windows), so only run plans you trust.

4. Or use MCP for structured tool calls (no shell syntax)

After installing with MCP support, start the server:

patchloom mcp-server

MCP-capable agents call patchloom tools directly as structured JSON, with no shell quoting or command construction. The agent sends {"path": "config.json", "selector": "version", "value": "2.0"} instead of building patchloom doc set config.json version '"2.0"' --apply.

Coding agents: set PATCHLOOM_MCP_SURFACE=core for an 11-tool pack (list_files, search/read/replace, doc/md, execute_plan, server_info) so schemas stay small. Product default remains full inventory when the env is unset. Prefer Patchloom MCP alone for list+edit (no second filesystem MCP). See the MCP setup guide for Cursor / Claude / Codex paste configs and the full security model.

Using VS Code, Cursor, or Windsurf? The Patchloom extension handles setup automatically: it installs the binary, runs init, and configures your editor's MCP settings.

As a Rust library

Host teams embedding Patchloom instead of a private edit stack: see the embedder host case study (for_agent, peels, fuzzy refuse, apply_fragment, path-only ops).

Add patchloom as a dependency (omit CLI/MCP/AST with default-features = false):

[dependencies]
patchloom = { version = "0.22", default-features = false }

use patchloom::api::{self, ApplyMode, ReplaceOptions, edit_error_kind, EditErrorKind};
use std::path::Path;

// Replace text (preview only, no disk write)
let result = api::replace_text(
    Path::new("src/config.rs"),
    "old_value", "new_value",
    &ReplaceOptions::default(),
    ApplyMode::Preview,
    None,
)?;
println!("{}", result.diff);

// Agent hosts: shared primary+fallback policy (unique, require_change, fuzzy @ 0.90)
let opts = ReplaceOptions::for_agent();
// Fail closed: zero matches become EditErrorKind::NoMatch
match api::replace_in_content("body", "missing", "x", &opts) {
    Ok(r) => println!("changed={}", r.changed),
    Err(e) => assert_eq!(edit_error_kind(&e), Some(EditErrorKind::NoMatch)),
}
// for_agent auto-refuses over-wide fuzzy; custom options use api::fuzzy_span_suspicious
// Buffer multi-op + host write: api::refuse_batch_if_suspicious_fuzzy after apply_content_edits
// Invalid options and bad regex peel InvalidInput (CLI/tx typed errors included)
match api::replace_in_content("body", "", "x", &ReplaceOptions::default()) {
    Err(e) => assert_eq!(edit_error_kind(&e), Some(EditErrorKind::InvalidInput)),
    Ok(_) => panic!("empty pattern must error"),
}

// Set a value in a JSON file
api::doc_set(
    Path::new("config.json"),
    "version",
    serde_json::json!("2.0"),
    ApplyMode::Apply,
    None,
)?;

// Multi-doc YAML: merge into document 0 (selector None = root only)
api::doc_merge(
    Path::new("stream.yaml"),
    serde_json::json!({"env": "prod"}),
    ApplyMode::Apply,
    None,
    Some("0"),
)?;

// Sole-path text load: binary → EditErrorKind::Binary; invalid UTF-8 → InvalidEncoding
let _text = api::load_text(Path::new("notes.md"))?;

All API types are Send + Sync. Beyond the api module, utility modules are also public: containment (workspace path guarding), exec (shell command execution), files (file-walking, load_text_strict, binary detection), backup (restore_path_from_latest_backup for post-Apply validate/revert), and write (atomic file writes with policy transformations). Library users needing temp dirs (e.g. agents) can use PathGuard::builder(cwd).allow_temp_directory() (handles /tmp on macOS); see the containment and api module rustdocs. Multi-doc bare keys and wrong-root merges peel to EditErrorKind::TypeError via edit_error_kind. Create/rename dest-exists peels to EditErrorKind::AlreadyExists (or api::is_already_exists / api::error_kind_str for CLI-stable "already_exists" strings). Fine-grained kinds also have bool peels (is_not_found, is_conflicts, is_changes_detected, is_type_error, is_format_failed, is_guard_rejected, is_invalid_input, is_no_match, is_ambiguous) matching edit_error_kind.

Replace fail-closed / shell-token options: CLI replace --require-change and --command-position (also plan/MCP fields and ReplaceOptions on the library). Agent hosts: ReplaceOptions::for_agent() on primary and fallback replace paths (auto span refuse); custom options still call api::fuzzy_span_suspicious / FuzzySpanPolicy after fuzzy Apply; buffer multi-op hosts call api::refuse_batch_if_suspicious_fuzzy after apply_content_edits (#2064). Library-only AST mutators: ast_rename / ast_replace_in_symbol / ast_rename_batch (feature ast + files), and FunctionSigEdit::parse_rust. Ordered host onboarding: Embedder host checklist (#2009). Full surface: docs.rs/patchloom.

Getting started

Commands

Agent-optimized (these are faster or safer than native tools)

General-purpose (also useful in scripts and CI)

How patchloom compares

The key difference: patchloom is designed for AI agent workflows. One batch or tx call replaces N sequential tool calls, cutting round-trips and eliminating partial-failure states.

vs agent-native editing tools

The table above compares patchloom to human CLI tools. But agents already have built-in editing: Claude Code's edit_file, Cursor's apply, Grok Build's search_replace, Aider's /code blocks. Why add patchloom on top?

Agent-native tools use text matching. They find a block of text and replace it. This works for source code but fails on structured config files:

Agent uses search_replace on YAML

database:
  # Production settings
  host: db.prod.internal
  port: 5432  # PostgreSQL default
  pool_size: 10

The agent replaces port: 5432 with port: 5433. Result depends on implementation. Many agents lose the inline comment, break indentation, or fail to match because of surrounding context changes.

Agent uses patchloom doc set

patchloom doc set config.yaml \
  database.port 5433 --apply

The YAML parser changes the value at the selector path. Comments, indentation, key ordering, and all other formatting are preserved. The output is always valid YAML.

When to keep using native tools: Single-file reads, simple text search, single-file text replacement where comments don't matter. Patchloom's agent-rules tell agents exactly when to use each approach.

When to use what

Longer write-ups: Comparisons · Embedder host checklist · MCP setup

Library hosts: Embedder host checklist: dual-path ReplaceOptions::for_agent() (primary + fallback, includes refuse_suspicious_fuzzy), peels via edit_error_kind / is_* / is_fuzzy_span_suspicious with #[non_exhaustive] _ arm, multi-op ContentEditsResult.op_honesty + refuse_batch_if_suspicious_fuzzy after buffer multi-op (#2064), plan/tx widest matched_text, optional apply_content_edits_to_file_with_span_policy. Custom options still use api::fuzzy_span_suspicious / FuzzySpanPolicy after fuzzy Apply.

Context budget: line-range read, search --count / --files-with-matches, one batch/tx, --jsonl for large streams.

How it works with your AI agent

Two integration modes, same capabilities:

flowchart LR
    subgraph CLI["CLI mode (any agent)"]
        direction TB
        A["patchloom agent-rules >> AGENTS.md"] --> B["Agent reads AGENTS.md"]
        B --> C{"What kind of edit?"}
        C -->|Simple edit| D["Native tool (faster)"]
        C -->|Config edit| E["patchloom doc (safer)"]
        C -->|Markdown edit| F["patchloom md (smarter)"]
        C -->|Multi-file edit| G["patchloom batch (batched)"]
    end

    subgraph MCP["MCP mode (MCP-capable agents)"]
        direction TB
        H["patchloom mcp-server"] --> I["Agent discovers tools via MCP"]
        I --> J["Structured JSON tool calls"]
        J --> K["No shell syntax needed"]
    end

Status

4000+ tests across 24 commands. Tested with Grok 4.3, GPT-5.4, and Claude Opus 4.6.

Full command reference

Every command, flag, transaction operation, and exit code is documented in the Command Reference (also available at docs/reference/README.md).

License

Licensed under either of:

• MIT license (LICENSE)

• Apache License, Version 2.0 (LICENSE-APACHE)

at your option.

Contributing

See CONTRIBUTING.md.

For local verification before opening a pull request, run make check. It matches the main Linux CI gate: formatting, clippy, unit tests (including feature-matrix jobs), integration tests, PTY tests, release-notes structure, test hygiene, and generated-doc freshness (check-patchloom-md, check-readme). While iterating locally, make check-fast is the same except it skips only check-patchloom-md (it still runs check-readme so a drifted test-count badge fails before CI).

All commits must be signed off with git commit -s.

Agent integration tests

make agent-test runs 19 pytest scenarios that verify AI agents correctly use patchloom when given instructions. make bench-agent runs 3-way benchmarks (CLI vs MCP vs native) across 11 tasks. Use MODEL=X to switch models and RUNS=N for variance reduction. Requires an LLM API key. Not part of make check. See tests/agent/README.md for details.

Security

For current security reporting guidance, see SECURITY.md.