Agent Toolkit for Coding CLIs
A drop-in toolkit for your agentic coding CLI
Bureau is an open-source toolkit that supercharges Claude Code, Gemini CLI, Codex, and OpenCode. It's not part of the Reverb platform; it's a standalone tool you hand to your agent. Bureau provides MCP servers, 66 specialized agent roles, and structured workflow skills, all working across every supported CLI.
Agentic coding CLIs are fragmented. Each has unique strengths but incompatible tooling. When you switch between them, you lose your setup: context, tools, custom workflows, all gone. Bureau solves this by providing a consistent layer across all CLIs, so you can pick the best model for each task without starting over.
Bureau comes with 66 specialized roles, including architect, debugger, code-reviewer, security-compliance, and more. Every role works as both an isolated subagent (with separate context) and an interactive main agent (in your conversation). Use Claude for architecture, Gemini for broad search, GPT for quick edits.
Bureau ships with structured workflow skills that agents activate automatically. The flagship is Assess Mode, a two-phase code review protocol: first it builds a complete mental model of your changes (with 4 comprehension styles to choose from, including hunk-by-hunk inline review), then it audits every file against your configurable quality standards. It runs as an interactive guided tour when you're in conversation, or produces a structured markdown report when delegated to a subagent. Other built-in skills include Micro Mode (step-gated editing with DAG-based planning for maximum control over each atomic edit), Scrimmage Mode (systematic self-attack testing across 5 vulnerability categories after every change), and Blast Radius Mode (impact analysis before edits that enumerates callers, dependents, tests, and contracts affected, then classifies changes as safe, needs review, breaking, or blocked).
Bureau integrates GitHub's open-source spec-kit CLI so agents write comprehensive specs before coding. The agent writes a spec, asks questions interactively, turns the spec into an implementation plan and concrete tasklist, then implements based on those docs. On-the-fly updates cascade through specs, plans, and tasks automatically. This significantly reduces agent mistakes, bugs, and unintended implementation omissions.
Bureau installs and configures a cohesive set of MCP servers: Sourcegraph and Serena for code search, Brave and Tavily for web research, Context7 for API docs, Qdrant and Memory MCP for persistence, Semgrep for security, and Playwright for browser automation.
Bureau injects context files at agent startup so your agent automatically knows how to use every tool and when to delegate to subagents, with no explicit instructions needed. This includes a handoff guide (which model to delegate to for which task), a tools guide (progressive disclosure of each MCP server's capabilities), and skill triggers that activate protocols like Assess Mode when they match your prompt.
Bureau makes it trivial to plug Reverb into your agent setup. Since Bureau already manages your MCP servers, adding Reverb Vinyl for graph memory is just another server in the set. Bureau is the easiest on-ramp to the Reverb platform.
Join the waitlist for early access to Reverb Cloud and updates on the platform.