fastedits 0.2.1

AST-aware code editing — 45% fewer output tokens than unified diffs

FastEdit

AST-aware code editing powered by a fine-tuned 1.7B model. Diffs, SEARCH/REPLACE, and apply_patch all force the agent to repeat back old code to say where the edit goes. FastEdit uses tree-sitter to find the target by name — the agent writes only the change plus a line or two of context.

Agent output token savings

Model	Edit tool tokens	FastEdit tokens	Saved	Reduction
GPT-5.4	3,404	1,557	1,847	54.3%
Opus 4.6	4,286	2,291	1,995	46.5%
Opus 4.7	4,771	2,645	2,126	44.6%
Grok 4.20	2,946	1,661	1,285	43.6%

The problem

Every AI code editor today makes the model output old code to locate edits. Whether it's unified diffs, SEARCH/REPLACE blocks, or apply_patch — the model has to repeat back the lines it wants to change:

# Claude Code / Codex — model outputs old AND new code
@@ -1,4 +1,6 @@
 def process(data):
-    result = transform(data)
-    return result
+    try:
+        result = transform(data)
+        return result
+    except Error as e:
+        return {"error": str(e)}

You're paying double: the model writes every old line (to say "find this") plus every new line (to say "put this"). On a 50-line function where you change 3 lines, that's 50 lines of output just for location, plus 3 lines of actual edit. ~94% of output tokens are wasted on telling the model where to put the code.

How FastEdit works

FastEdit eliminates location tokens entirely. Instead of making the model repeat old code, it uses two things:

1. AST awareness — tree-sitter parses the file and finds the target function/class by name. No need to output old lines for location.
2. A fine-tuned 1.7B SLM — when the edit is complex, a small merge model takes the original chunk (~35 lines) + edit snippet and produces the merged result.

# FastEdit — model writes ONLY the change
fastedit edit api.py --replace process --snippet '
def process(data):
    try:
        # ... existing code ...
    except Error as e:
        return {"error": str(e)}
'

--replace process uses tree-sitter to find the function. # ... existing code ... tells the system to preserve untouched lines. The model never outputs old code — zero tokens spent on location.

The three edit modes

Mode	What happens	Tokens	Speed
--after symbol	Text insertion after the named symbol	0	Instant
--replace symbol (deterministic)	Context anchors splice new lines in	0	Instant
--replace symbol (model)	1.7B SLM merges snippet into ~35-line chunk	~40	<1s

The system tries deterministic text-matching first. It classifies each snippet line as "context" (matches the original) or "new" (the edit), then splices new lines between the matched anchors. This handles 74% of real edits with zero model calls.

When deterministic matching can't resolve the edit (indent structure changes, full rewrites, <2 matching lines), the 1.7B model takes over. It only ever sees a ~35-line function — never the whole file — so it's fast and accurate.

Install

Prerequisite: tldr must be on PATH (used for AST analysis).

pip install fastedits
fastedit pull          # downloads the 1.7B model (~3GB, one-time)

For Apple Silicon (recommended for local use):

pip install fastedits[mlx]
fastedit pull

For GPU servers (vLLM, TGI) or local servers (LM Studio, llama.cpp, Ollama):

# Optional: install vLLM for GPU serving
pip install fastedits[vllm]

# Or point at any OpenAI-compatible server:
FASTEDIT_BACKEND=llm FASTEDIT_LLM_API_BASE=http://localhost:1234/v1 fastedit edit ...

CLI

# View file structure (functions, classes, line ranges)
fastedit read src/app.py

# Edit a function (AST-scoped merge)
fastedit edit src/app.py --replace handle_request --snippet '
def handle_request(data):
    validate(data)
    # ... existing code ...
    logger.info("done")
'

# Insert new code after a symbol (0 tokens)
fastedit edit src/app.py --after handle_request --snippet '
def health_check():
    return {"status": "ok"}
'

# Batch edits to one file
fastedit batch-edit src/app.py --edits '[
  {"snippet": "import redis", "after": "import json"},
  {"snippet": "def cache_get(key): ...", "after": "connect"}
]'

# Delete, move, rename (all instant, no model)
fastedit delete src/app.py deprecated_handler
fastedit move src/app.py helper_func --after main
fastedit rename src/app.py old_name new_name

# Undo last edit / show diff
fastedit undo src/app.py
fastedit diff src/app.py

MCP server

FastEdit runs as an MCP server for AI agents (Claude Code, Cursor, etc.):

pip install fastedits[mcp]
python -m fastedit.mcp_server

Add to Claude Code config:

{
  "mcpServers": {
    "fastedit": {
      "command": "python",
      "args": ["-m", "fastedit.mcp_server"]
    }
  }
}

10 tools: fast_edit, fast_batch_edit, fast_multi_edit, fast_read, fast_search, fast_diff, fast_delete, fast_move, fast_rename, fast_undo

Auto-redirect Edit → fast_edit (optional)

A PreToolUse hook intercepts Claude's built-in Edit tool and redirects to fast_edit. Zero tokens wasted — Edit never executes. Works on Mac, Linux, and Windows (PowerShell too).

Add to .claude/settings.json or your project .claude.json:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Edit",
        "hooks": [{"type": "command", "command": "fastedit-hook"}]
      }
    ]
  }
}

fastedit-hook is installed automatically with pip install fastedits — no paths, no python3 vs python issues.

The model

FastEdit includes a fine-tuned 1.7B parameter model (Qwen2.5-Coder-1.5B architecture) trained specifically for code merging. It takes an original code chunk + edit snippet and produces the merged result.

Most edits never reach the model:

• --after is pure text insertion (0 tokens, instant)
• --replace tries deterministic text matching first (0 tokens, instant)
• Only when the snippet has complex structural changes does the 1.7B model activate

The model is scoped to ~35-line chunks via AST, so it runs in <1s on Apple Silicon (MLX) or GPU (vLLM).

Accuracy

Tested across 22 structurally distinct edit patterns (73 cases):

Path	Accuracy	Tokens	Latency
Deterministic (74% of edits)	100%	0	<1ms
Model (26% of edits)	92%	~40	~500ms
Combined (production)	~98%	~10 avg	~130ms avg

The deterministic path handles the easy majority perfectly and for free. The model handles the complex minority. The AST scoping prevents the failure modes that plague whole-file approaches (ordering errors, content loss).

Per-language model accuracy (156-example benchmark):

Language	Accuracy
Python, Java, Kotlin, C, PHP	92%
JavaScript, TypeScript, Rust, Swift	85%
Go, C++, Ruby	77%

How it compares

	FastEdit	Claude Code / Codex	Aider SEARCH/REPLACE
How it locates the edit	AST — names the symbol	Model outputs old lines	Model outputs SEARCH block
Tokens for location	0	~50% of output	~50% of output
What the model sees	~35-line chunk	Entire file context	Entire file context
Failure mode	Symbol not found (immediate, clear error)	Can't find old lines (silent misapply)	Can't find SEARCH block
Languages	13	Any	Any

Supported languages

Python, JavaScript, TypeScript, Rust, Go, Java, C, C++, Ruby, Swift, Kotlin, C#, PHP

Environment variables

Variable	Default	Description
FASTEDIT_MODEL_PATH	~/.cache/fastedit/models/...	Path to model
FASTEDIT_BACKEND	mlx	Backend: mlx or llm
FASTEDIT_LLM_API_BASE	http://127.0.0.1:8000/v1	LLM server URL (any OpenAI-compatible)
FASTEDIT_LLM_MODEL	fastedit	Model name to send in API requests
FASTEDIT_LLM_API_KEY	not-needed	API key (if server requires one)

License

MIT