mcp-injector

AI coding assistants often fail because they retrieve the wrong context. On a large codebase, blindly dumping raw files into the prompt leads to hallucinations, slow responses, and high API costs.

Foldwork fixes this. It is a deterministic repository understanding engine that pre-indexes your entire codebase into a local SQLite catalog. It acts as the Context Layer for your IDE, serving exactly the functions the AI needs—no more, no less—maximizing the first-try success rate and reducing token usage by 41-89%.

By combining AST body folding (which strips out function bodies while preserving signatures) with canonical determinism (which guarantees byte-identical outputs to maximize Anthropic's KV cache hits), Foldwork transforms massive enterprise monorepos into lightweight, cache-friendly payloads. This drastically reduces token consumption, cuts API costs by up to 90%, and eliminates context window overflow.

No cloud. No telemetry. Runs entirely on your machine.--

Real-World Codebase Context Benchmarks

Estimate the impact of AST code compression on large open-source repositories (calculated at $2.00 / million input tokens for Claude Sonnet 5):

Numbers are reproducible. Run the open-source benchmark tool on any public repository:
mcp-benchmark repository

What It Looks Like

Run mcp-benchmark on your own project to see your exact savings before installing anything:

mcp-benchmark ./your-project

════════════════════════════════════════════════════════════════════════════════
  mcp-injector Benchmark — context
  Tier 3 compression  |  $2.00/1M tokens  |  2026-07-15T12:00:00Z
════════════════════════════════════════════════════════════════════════════════

FILE                                          RAW TOKENS    COMPRESSED     SAVED   COST SAVED*
──────────────────────────────────────────────────────────────────────────────────────────
cmd/license-gen/main.go                            3,633           214       94%       $0.0072
main.go                                           17,555         1,917       89%       $0.0347
website/api/webhook.go                             2,682           295       89%       $0.0053
main_test.go                                       1,576           353       78%       $0.0031
──────────────────────────────────────────────────────────────────────────────────────────
TOTAL (4 files)                                   25,446         2,779     89.1%       $0.0503

  * Based on $2.00 / 1M input tokens

  💡 Running this codebase through Claude 10×/day costs $0.51/day raw.
     With mcp-injector:  $0.01/day.  You save $0.50/day ($15/month).

Tools

get_project_map

Returns a compressed structural overview of the workspace. Function bodies are folded and replaced with placeholders to reduce token usage.

• tier (integer, optional): Compression tier to apply (default: 2).
• unfolded_files (array of strings, optional): Workspace-relative paths or glob patterns for files to serve at full resolution (uncompressed).
• path_prefixes (array of strings, optional): Scope the project map to specific microservices or packages, drastically reducing payload bloat.
• git_context: always includes current branch, changed files, and recent commits in the response.
• secrets_redacted: count of credentials automatically redacted before sending to Claude.

Example call:

{
  "tool": "get_project_map",
  "arguments": {
    "tier": 3,
    "unfolded_files": ["src/auth/handler.go", "**/*_test.go"],
    "path_prefixes": ["src/auth/"]
  }
}

injector_retrieve

Retrieves the full uncompressed source of a file from the local cache.

• path (string, required): The workspace-relative path of the file to retrieve.
• retrievalKey (string, optional): The SHA-256 retrieval key returned in a prior compressed payload.
• start_line (integer, optional): 1-indexed start line for range retrieval.
• end_line (integer, optional): 1-indexed end line for range retrieval.
• expand_graph (boolean, optional): Resolves and appends cross-file dependencies (limited to 50 1st-degree dependencies).

injector_search

BM25-ranked full-text symbol search over the local SQLite catalog. Supports FTS5 boolean logic (e.g., user AND (auth OR login)).

• query (string, required): FTS5 query string (bare terms, "phrase", prefix*).
• limit (integer, optional): Maximum results (default: 20).
• search_paths (array of strings, optional): Scope search to specific isolated directories.

injector_diagram

Generates a Mermaid sequence diagram for a given symbol by traversing its outbound dependencies (halts after 500 nodes).

• symbol (string, required): The exact symbol name.
• max_depth (integer, optional): Maximum traversal depth (default: 3).
• include_primitives (boolean, optional): Include basic types (String, boolean) and framework boundaries.

injector_regex_search

Fallback for exact literal or regex searches against file contents. Bypasses FTS5 tokenization.

• query (string, required): The string or regex pattern to search for.
• is_regex (boolean, optional): Treats query as extended regex (-E).

injector_write_file

Write a full file to disk. CRITICAL: Prevents data loss by intercepting and rejecting payloads containing compressed fold markers.

injector_blast_radius

Analyzes the architectural impact of changing a symbol by traversing the dependency graph. Supports inbound and outbound directional traversal.

injector_git_context

Integrates with local Git history to surface commit context, authorship, and code evolution directly into the LLM context.

injector_inspect_table

Enables direct database introspection capabilities. Currently supports PostgreSQL and MySQL. CRITICAL: You must start the daemon with the FOLDWORK_DB_DSN environment variable set to your database connection string (e.g. postgres://user:pass@localhost:5432/dbname) to activate this tool.

injector_clear_cache

Wipes the SQLite index cache and triggers a clean cold-start full re-index.

injector_stats

Returns index status, current compression ratio, total files indexed, and cache hit rate.

injector_sync (Deprecated)

Read tools automatically wait for pending indexing implicitly. You never need to manually call this tool.

Quick Install

Install the daemon locally and configure your IDEs:

curl -fsSL https://foldwork.dev/install | sh

Automatically configures Claude Desktop, Cursor IDE, VS Code, Devin Desktop, and Antigravity.

Getting Started

Step 1: Check if your project qualifies for the free tier

Run the benchmark CLI on your project to see your token savings and line count:

mcp-benchmark ./your-project

If your project is under 50,000 lines, mcp-injector is completely free. The benchmark output shows your exact line count.

Step 2: Install the daemon

curl -fsSL https://foldwork.dev/install | sh

The installer auto-detects Claude Desktop, Cursor, VS Code, Devin Desktop, and Antigravity and writes the MCP config automatically. You should see output like:

* mcp-injector v0.2.0 installed to /usr/local/bin/mcp-injector
* Claude Desktop configured
* Cursor configured
Restart your IDE and mcp-injector will be active.

Step 3: Restart your IDE

The MCP server starts automatically when your IDE launches. No separate daemon process to manage.

Step 4: Verify it is working

In Claude Code or Cursor, ask Claude:

"Use get_project_map to show me the structure of this project"

Claude will call the mcp-injector tool and return a compressed map of your entire codebase. If you see module names, entry points, and dependency information - it is working.

Step 5: Get the full source when needed

When Claude needs to see the complete implementation of a compressed function, it automatically calls injector_retrieve. You can also trigger this explicitly:

"Show me the full implementation of UserService.java"

Claude will fetch the uncompressed source from the local cache.

Editing Code: You MUST use the injector_write_file tool to edit code. If Claude tries to write back folded placeholders into your source code, the daemon will hard-reject the payload to protect you from data loss.

Step 6: Check your savings

injector_stats

Or ask Claude directly: "Call injector_stats and tell me my current token savings."

Agent Use Cases & Advanced Usage

Now that your AI has deterministic tools to search, traverse, and retrieve code, you can ask it high-level architectural questions that usually fail on raw codebases: