Plugins
Install the Baz plugin in Claude Code, OpenAI Codex, or Cursor.
The Baz plugin brings Baz into your coding agent: indexed search across every repository in your organization, the codebase-exploration skill that tells the agent when to use it, and two commands built on top, one for planning a change and one for reviewing it.
Pick your agent:
/plugin
OpenAI CodexInstall with codex plugin install
CursorAdd the Baz rule and MCP server manually
What the plugin adds
Every platform gets the same three indexed search tools, served by the Baz MCP server:
| Tool | What it does |
|---|---|
repo_search(keywords, domains?) |
Semantic search across indexed architecture summaries, spanning every repository in your organization and the per-domain summaries of a single large monorepo. |
remote_grep(repository, pattern, path) |
Indexed regex grep inside one repository, scoped to a path, with roughly two lines of context per match. |
remote_file_search(repository, pattern) |
Glob file-name search inside one repository, for when you have a naming hunch but not the exact path. |
Searches run against the Baz index, so they work on repositories you have never cloned. No checkout, no local tree walk, no API rate limits.
Alongside the tools, the plugin ships:
- The
baz-codebase-explorationskill. Loads automatically when relevant. It routes searching through Baz, keeps file reads on your own tooling, and enforces a search budget so the agent stops searching and starts reading once it has a hit. - Two commands.
/baz:plan-with-bazto plan a change,/baz:reviewto review one. Both spend the search budget on the questions only an org-wide index can answer. - Session hooks. They pass your session id and current repository to Baz so tool calls and the final plan land on your Baz timeline.
Commands
Plan a change
/baz:plan-with-baz is manually invoked. It explores the relevant repositories, including ones you have never cloned, then writes a structured implementation plan for you to approve before any code is written. See Planner for what happens to the plan once it reaches Baz.
Review a change
/baz:review reviews your changes the way a reviewer with the whole organization’s code in front of them would. It resolves a diff, reads the changed files for context, then uses the Baz tools to check the change against repositories you do not have checked out: the caller in another service that still passes the old signature, the consumer that reads a field you just renamed, the registration site your new enum value is missing from.
/baz:review # everything not yet on the base branch
/baz:review committed # only commits on this branch
/baz:review uncommitted # only staged + unstaged edits
/baz:review --include-untracked # also review new, untracked files
/baz:review --base develop # compare against a different base
/baz:review --pr 42 # review an open PR (needs gh or glab)
/baz:review --fix # review, then apply the fixes it finds
Unlike the planning command, the review skill also triggers on plain requests such as “review my changes” or “check this branch for security issues”.
Findings come back grouped by severity (blocking, should fix, consider), each with a file:line, the concrete conditions that make it fail, and a fix. Cross-repo findings name the repository they live in. Style nits and pre-existing issues the diff did not touch are deliberately left out.
Four cross-repo checks run in priority order, until the search budget is spent:
| Check | What it catches |
|---|---|
| Broken consumers | A call site in another repository that still passes the old signature, shape, or field name. |
| Both sides of a contract | One end of a request, response, event payload, or schema changed while the other end was not updated. |
| Diverging from a peer | A new enum value, handler, provider, or route that is missing from a registration site an existing peer is wired through. |
| Reinventing something | A retry, validator, parser, or client the organization already has as a drop-in helper. |
The cross-repo checks need Baz connected. If it is not, the review does not silently shrink to a local one. For a change with an outward-facing surface (a changed signature, schema, event, or public export) it stops, names the symbols whose consumers it could not check, and withholds the merge verdict rather than implying coverage it does not have. Purely local changes, such as formatting or a private helper, still review normally because there is nothing cross-repo to miss. Every report ends with a ## Coverage line stating which repositories were searched, so a partial run never reads as a complete one.
With --fix, or when you ask afterwards, the agent turns the findings into a task list, shows it, then works through it one finding at a time, running the project’s tests and linter if it has any. It only edits the repository you have checked out. Findings in other repositories are reported, never edited.
Server URL
All three plugins connect to the same hosted MCP server:
https://baz.co/mcp
Authentication runs through OAuth. The first time the agent calls a Baz tool, your browser opens and you sign in with your Baz account. You are in full control of when Baz is called, and nothing happens behind the scenes.
Other MCP clients
If your editor is not one of the three above, you can still wire the Baz MCP server by hand. You get the tools, but not the skill, the planning command, or the session hooks.
VS Code
{
"mcp": {
"servers": {
"baz": {
"type": "http",
"url": "https://baz.co/mcp"
}
}
}
}
Claude Desktop
{
"mcpServers": {
"baz": {
"command": "npx",
"args": ["-y", "mcp-remote", "https://baz.co/mcp"]
}
}
}
Windsurf
{
"mcpServers": {
"baz": {
"command": "npx",
"args": ["-y", "mcp-remote", "https://baz.co/mcp"]
}
}
}
Troubleshooting
Clear your ~/.mcp-auth directory
Clients that connect through mcp-remote store their credentials in ~/.mcp-auth (or wherever MCP_REMOTE_CONFIG_DIR points). If you hit connection or auth errors, clear it and restart the client:
rm -rf ~/.mcp-auth
Reset cached tools in VS Code
If Baz updates a tool signature (parameters, description), the MCP cache may need a reset:
- Open the Command Palette (
Cmd+Shift+Pon Mac). - Run
MCP: Reset cached tools.
The agent searches locally instead of calling Baz
That usually means the skill or rule is not loaded. On Cursor, confirm the rule file sits in .cursor/rules/. On any platform, you can force the routing by asking the agent to use the Baz tools explicitly, or by running /baz:plan-with-baz.
Managing configuration centrally
Review configurations, coding standards, and conventions are managed in one place at https://baz.co, so the same rules apply in the editor and in cloud reviews. Sign up free for a 14 day trial at https://baz.co.
FAQ
Do I need to clone a repository to search it?
No. Baz search runs against an index of your organization’s repositories, so the agent can plan a change against code you have never checked out. Reads still use your own tooling, so once a file path is known you open it however you normally would.
What is the Model Context Protocol (MCP)?
MCP is an open protocol that lets AI-enabled editors and CLIs talk to external servers. Baz ships a hosted MCP server at https://baz.co/mcp, and each plugin wires it for you.
What should I do if the connection fails?
Clear ~/.mcp-auth, reset cached tools if you are on VS Code, and confirm the configured transport or URL is correct. Then restart the client so the OAuth flow runs again.