Local codebase search for AI assistants via MCP. Requires Hawkeye and Hawkeye AI Bridge installed.
Hawkeye AI Bridge MCP Server (io.github.Hawkeye-Zaragsoft/hawkeye-aibridge)
This MCP server enables local codebase search for AI coding assistants by exposing Hawkeye’s local project index over Model Context Protocol (MCP). It is designed to work with tools such as Claude Code, VS Code, Cursor, and AI agents. Hawkeye and Hawkeye AI Bridge must be installed, and search relates to indexed project content.
🛠️ Key Features
Local-first codebase search via MCP
Queries Hawkeye’s local project index
Supports AI assistants including Claude Code, VS Code, Cursor, and AI agents
Token/runs context notes in the provided excerpt (token budget “runway”)
🚀 Use Cases
Searching local repositories for AI-assisted coding and refactoring workflows
Using an AI agent to query an existing local index for relevant code
⚡ Developer Benefits
Integrates Hawkeye indexing with MCP clients
Enables developers to route code search through Hawkeye’s indexed results
⚠️ Limitations
Requires Hawkeye and Hawkeye AI Bridge installed
The excerpt includes a partial token-budget discussion and does not describe additional constraints
Runway at 10M tokens/month - Days until a hypothetical 10M/month context budget is exhausted, if every one of 50 daily search steps looked like this one. Longer bar, more runway.
Broad grep - 1.6 days
Compact grep - 7.0 days
Hawkeye indexed - 29.0 days
image
Local codebase search for Claude Code, VS Code, Cursor, and AI agents.
Hawkeye AI Bridge lets AI coding assistants query Hawkeye’s local project index through MCP, so they can find exact references across large codebases, assets, scripts, localization, config files, and project content without uploading the full repository.
Instead of giving an AI agent huge raw grep output, Hawkeye can return compact file/line results that help the agent decide which files are actually worth reading next.
Search large projects locally
Reduce token usage from noisy raw search output
Find exact references across code and non-code project files
Useful for C++, C#, Unreal Engine, Unity, legacy projects, and proprietary codebases
Hawkeye runs locally/on-premises and does not use AI for indexing or searching
This repository is the official public setup and discovery repository for using Hawkeye AI Bridge as an MCP server.
It does not contain the proprietary Hawkeye AI Bridge source code. Hawkeye and Hawkeye AI Bridge are proprietary software owned by Zaragsoft.
Why use Hawkeye AI Bridge?
Hawkeye helps developers quickly search and understand large projects. With MCP support, AI coding assistants can ask Hawkeye for fast local search results, saving tokens on each call.
Typical use cases:
Find all references to a class, function, symbol, asset, localization key, or file.
Search large codebases without relying only on editor search.
Help AI coding assistants understand more of the project before making changes.
Navigate large game/software projects with less reliance on tribal knowledge.
Keep search local/on-premises.
~80% token reduction vs. blind file reading.
If you do not specify any groups then all groups will be used.
Recommended install path: Place the AIBridge folder at C:\Program Files\Hawkeye\AIBridge\. Use this path in all examples below. If you install elsewhere, substitute your actual path throughout.
Client Setup
Claude Desktop
You don't need this zip for Claude Desktop. Download Hawkeye.mcpb from the release page and install it directly:
If .mcpb is a recognised file type on your system: double-click the file.
Otherwise: go to Settings → Extensions → Install Extension and select the file.
Move the AIBridge folder to a permanent location — recommended: C:\Program Files\Hawkeye\AIBridge\. Don't run it from Downloads; the registered path must keep working.
Register the MCP server with Claude Code:
powershell
claude mcp add hawkeye --scope user -- "C:\Program Files\Hawkeye\AIBridge\HawkeyeAIBridge.exe"
Install the Skill so Claude knows when to use Hawkeye. hawkeye-search.skill is a zip archive (it just contains SKILL.md), so it needs to be extracted, not renamed:
Why cmd /c? Cursor passes the command path to the shell without quoting it, which breaks paths containing spaces (like C:\Program Files\). Wrapping it with cmd /c works around this.
Adjust the path if Hawkeye AI Bridge is installed somewhere else.
OpenCode
Add the following to your opencode.json (create it in your project root or home directory if it doesn't exist):
command is an array of strings, not a command + args split
"type": "local" is required
Installing the Skill for OpenCode
The Skill is required for OpenCode. Without it, OpenCode will try to call hawkeye.exe directly on the command line rather than routing tool calls through the MCP server.
OpenCode expects skills to be extracted from the .skill zip archive:
Once both are in place, restart OpenCode and try: "Find where PlayerController is defined." You should see it call hawkeye_search_minimal.
Installing the Skill for Claude
The Skill teaches Claude when and how to use the Hawkeye tools. Without it, Claude has to guess from tool descriptions alone — results will be worse.
The shipped file is called hawkeye-search.skill, but it's actually a zip archive containing a single SKILL.md — it needs to be extracted, not renamed. (Renaming the zip to SKILL.md produces a file with that name but zip contents inside, which Claude can't read as a skill.)
Open File Explorer and navigate to %USERPROFILE%\.claude\skills\ (create the skills folder if it doesn't exist).
Create a new folder named hawkeye-search.
Right-click hawkeye-search.skill and choose Extract All..., then extract into the hawkeye-search folder you just created. (Windows recognizes the .skill file as a zip archive even though the extension isn't .zip — if "Extract All" isn't offered, copy the file, rename the copy's extension to .zip, then extract that instead.)
Confirm the result is hawkeye-search\SKILL.md — a plain text file, not the original archive.
To restrict the Skill to a single project, place SKILL.md at <your-project>\.claude\skills\hawkeye-search\SKILL.md instead.
Available Tools
Tool
Description
hawkeye_search_minimal
Token-efficient search — returns only file paths and line numbers. 91% fewer tokens than grep. Best for chat-based lookups.
hawkeye_search
Full search with metadata — returns complete result objects with context. Best for detailed analysis.
hawkeye_get_groups
List all searchable code groups configured in Hawkeye.
hawkeye_get_editors
List configured editors for opening files.
hawkeye_execute_editor
Open a file in a configured editor at a specific line number.
hawkeye_health_check
Verify Hawkeye service is running and accessible.
hawkeye_reload_index
Trigger a headless Hawkeye index reload (no window, no stolen focus) on request.
validate_hawkeye_path
Check that the configured Hawkeye path exists and is valid.
get_settings
Retrieve current Hawkeye configuration (path, max results, case sensitivity).
save_settings
Save Hawkeye configuration.
copy_to_clipboard
Copy text to the system clipboard.
Token Savings
Approach
Tokens used
Blind file reading (no Hawkeye)
~100,000 per session
With hawkeye_search_minimal
~20,000 per session
Savings
~80% reduction
Example Prompts
Ask your AI coding assistant:
List groups inside Hawkeye to find all the groups.
Use Hawkeye to find all references to this class.
Search the workspace for where this asset name is used.
Find likely files related to the inventory system in group 4,7.
Search for this localization key across the project.
Find where this Blueprint or Unreal asset is referenced.
Show me files related to feature loadout before editing code.
Find all the sounds for explosion inside groups 1,5,7
Use Hawkeye to find code and content references before changing this file.
If you do not specify any groups then all groups will be used.
Privacy and Security
Hawkeye runs locally/on-premises.
Hawkeye does not upload your source code.
Hawkeye itself does not use AI for indexing or searching.
The MCP server only exposes Hawkeye functionality to tools you configure locally.
You control which MCP clients can connect to it.
Always review AI-generated code changes before applying them.
You may have registered it in a different scope. Re-register with --scope user:
powershell
claude mcp remove hawkeye
claude mcp add hawkeye --scope user -- "C:\Program Files\Hawkeye\AIBridge\HawkeyeAIBridge.exe"
claude mcp list shows hawkeye but with 0 tools
The exe was registered but failed to start. Common causes:
The AIBridge folder was moved or deleted after registration.
Hawkeye itself can't be found — run hawkeye --version in a fresh terminal to check.
Fix: claude mcp remove hawkeye then re-register.
Hawkeye not found
code
Error: Failed to start Hawkeye process
Open a new terminal and run hawkeye --version.
If that fails, set HAWKEYE_PATH (see Prerequisites above) and close/reopen your terminal.
Environment variables only apply to terminals opened after you set them.
Skill isn't being used by Claude
Confirm the file exists at %USERPROFILE%\.claude\skills\hawkeye-search\SKILL.mdand that it's a plain text file, not the .skill archive under a different name. Right-click it → Properties; if the size matches the original .skill file (~14 KB) and it won't open in a text editor, it was renamed instead of extracted — delete it and re-run the Expand-Archive step above.
Restart your Claude Code session — Skills are loaded at session start.
Cursor: connection failed with path error
If Cursor logs show 'C:\Program' is not recognized as an internal or external command, the path with spaces is being passed unquoted. Use the cmd /c workaround shown in the Cursor section above.
No groups returned
code
No groups available / Empty group list
Run hawkeye --getgroups manually to test.
Verify Hawkeye configuration is correct.
Check working directory permissions.
Timeout errors
code
Error: Operation timed out waiting for Hawkeye
Check if the Hawkeye process is hanging.
Check system resources.
Restart Hawkeye and try again.
License
Dual License:
SKILL.md and skill components: MIT License — freely modifiable and redistributable.
HawkeyeAIBridge executable: Proprietary (Zaragsoft) — free to use, cannot be modified, reverse-engineered, or redistributed.