Architecture is files + a live Q&A engine, not plugins. Portable across any IDE, any LLM, zero vendor lock-in.

Two engine fixes landed during the benchmark, both committed in this branch:

1. _ask_with_agent_md now surfaces project-level docs (conventions.md, module_registry.md, map.md, structure.md) into its answer prompts. Removes the "module knowledge does not include project-wide conventions" refusal pattern.
2. The structured-facts answer agents now have search_code, read_file, list_directory, read_file_metadata, search_by_type bound at runtime, so the LLM can grep and read actual source instead of paraphrasing the KG.

Full report (data, methodology, per-cell tables, caveats): artifacts/benchmark-2026-05-09/REPORT.md.

rb-setup — first-time configuration

Run this once per project, right after installing the plugin. Interactive picker for the LLM provider (OpenAI / DeepSeek / Groq / 阿里灵积 / NVIDIA NIM / Ollama local / any OpenAI-compatible endpoint), then writes .env to the project root with OPENAI_BASE_URL, OPENAI_API_KEY, OPENAI_MODEL, RB_ASK_TIMEOUT_SECONDS. It can also write an explicit local Codex host-runner config for experimental no-API-key rb-ask. Also ensures .env is in .gitignore. Skip it if you already have a working .env.

rb-refresh — build / refresh the knowledge base

Deploys the multi-agent cluster to read your code: each module gets its own Agent that produces a knowledge doc under .repobrain/agents/*.md, plus a map.md routing index. Run after install, after significant code changes, or when rb-ask returns stale answers. The first refresh auto-creates .repobrain/ — no separate init step needed. Pass quick for an incremental update, failed-only to rerun only previously failed modules.

Time: a few minutes for small repos, longer for large ones. Requires rb-setup to have completed. Full LLM refresh requires an API-key/OpenAI-compatible provider; local host-runner mode can use RB_REFRESH_SCAN_ONLY=1 rb-refresh --workspace . for scan artifacts.

rb-ask — routed Q&A on the codebase

The main reason this plugin exists. Routes your question to the right ModuleAgent (and GitAgent when applicable), then returns an answer grounded in actual source with file paths and line numbers. Use it before manually grepping or reading files — it's faster and more accurate. Good question shapes: "where is X defined/handled?", "why was Y done this way?", "how does the auth flow work?", "what depends on module Z?".

Requires a knowledge base — if you see "no index" or empty answers, run rb-refresh first.

rb-init — scaffold a new multi-agent repo

Creates a new project from the RepoBrain template. Two modes: quick (fast scaffold, clean copy) and full (adds runtime profile, .env, mission file, sandbox config, optional git init). This is for starting a new repo — you do not need it before rb-refresh on an existing project.

The plugin also bundles the agent-repo-init skill (the same backend that rb-init invokes — Codex / Claude can also match it by description) and the optional rb-mcp MCP server (ask_project + refresh_project) for tool-style integration.

rb-refresh — Multi-agent self-learning (8-step pipeline)

rb-refresh --workspace my-project

1. Scan codebase (languages, frameworks, structure)
2. Multi-agent pipeline generates conventions.md
3. Generate structure.md — language-agnostic file tree with line counts
4. Build knowledge graph (knowledge_graph.json + mermaid)
5. Write document/data/media indexes
6. LLM full-context analysis — group files by import graph + directory + prefix, pre-load into context (~30K tokens per sub-agent), filter out build artifacts. Each sub-agent reads the full source code and outputs a comprehensive Markdown knowledge document (agents/*.md). Large modules get multiple agent docs (one per group, no merging). Global API concurrency control prevents rate-limiting. Fully language-agnostic — works with any programming language.
7. RefreshGitAgent analyzes git history, generates _git_insights.md
8. Map Agent reads all agent docs → generates map.md (module routing index with descriptions and key topics)

rb-ask — Router-based Q&A

rb-ask "How does auth work in this project?"

Router reads map.md → selects modules → reads agents/*.md → LLM answers with code references. Multiple agent docs are read in parallel, then a Synthesizer combines answers.

Falls back to the legacy Router → ModuleAgent/GitAgent swarm when agent docs are not yet generated.

Key design choices

• LLM as analyzer: No AST parsing or regex — source code is fed directly to LLMs. Works with any programming language out of the box.
• Smart grouping: Files grouped by import relationships, directory co-location, filename prefixes. Build artifacts filtered. Hard character limit (800K) prevents context overflow.
• No information loss: Large modules produce multiple agent.md files — no merging or compression. Parallel reads + Synthesizer recombines at answer time.
• Global API concurrency control: RB_API_CONCURRENCY limits total simultaneous LLM calls.
• Language-agnostic module detection: Pure directory structure — no __init__.py or any language-specific marker required.

Instead of reading hundreds of documentation files, Claude Code can call ask_project as a live tool — backed by a dynamic multi-agent cluster: Router routes questions to the right ModuleAgent, returning grounded answers with file paths and line numbers.

Setup:

# Install engine
pip install "git+https://github.com/study8677/repobrain.git#subdirectory=engine"

# Refresh knowledge base first (ModuleAgents self-learn each module)
rb-refresh --workspace /path/to/project

# Register as MCP server in Claude Code
claude mcp add repobrain rb-mcp -- --workspace /path/to/project

Tools exposed to Claude Code:

MCPClientManager lets your agents connect to external MCP servers (GitHub, databases, etc.), auto-discovering and registering tools.

// mcp_servers.json
{
  "servers": [
    {
      "name": "github",
      "transport": "stdio",
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "enabled": true
    }
  ]
}

Set MCP_ENABLED=true in .env to make configured servers available, and set RB_ALLOW_MCP=true only when you want rb-ask to auto-connect those external servers. Stdio MCP servers inherit process environment plus configured env values, so treat enabled servers as local-permission code.

The default sandbox is for trusted local workspaces, not untrusted code isolation. Retrieval graph files redact common secrets before writing to disk, but full mode can still preserve source snippets. See Sandbox docs.

rb ask / rb refresh are available when both cli/ and engine/ are installed. rb-ask / rb-refresh are the engine-only entrypoints.

repobrain/
├── cli/                     # rb CLI — lightweight, pip-installable
│   └── templates/           # .cursorrules, CLAUDE.md, .repobrain/, ...
└── engine/                  # Multi-agent engine + Knowledge Hub
    └── repobrain_engine/
        ├── _cli_entry.py    # rb-ask / rb-refresh / rb-mcp + python -m dispatch
        ├── config.py        # Pydantic configuration
        ├── hub/             # ★ Core: multi-agent cluster
        │   ├── agents.py    #   Router + ModuleAgent + GitAgent
        │   ├── contracts.py #   Pydantic models: claims, evidence, refresh status
        │   ├── ask_pipeline.py    # agent.md + graph-enriched ask
        │   ├── refresh_pipeline.py # LLM-driven refresh → agents/*.md + map.md
        │   ├── ask_tools.py
        │   ├── scanner.py   #   multi-language project scanning
        │   ├── module_grouping.py # smart functional file grouping
        │   ├── structure.py
        │   ├── knowledge_graph.py
        │   ├── retrieval_graph.py
        │   └── mcp_server.py
        ├── mcp_client.py    # MCP consumer (connects external tools)
        ├── memory.py        # Persistent interaction memory
        ├── tools/           # MCP query tools + extensions
        ├── skills/          # Skill loader
        └── sandbox/         # Code execution (local / microsandbox)

CLI (pip install .../cli) — Zero LLM deps. Injects templates, logs reports & decisions offline.

Engine (pip install .../engine) — Repository knowledge runtime. Powers rb-ask, rb-refresh, rb-mcp. Uses the OpenAI-compatible endpoint written by rb-setup (OpenAI, DeepSeek, Groq, DashScope, NVIDIA NIM, Ollama, or custom). Experimental local mode can set RB_HOST_RUNNER=codex so rb-ask runs through the user's local codex login instead of an API key; this is for personal/local use and is not a hosted product backend.

Skill packaging:

• engine/repobrain_engine/skills/graph-retrieval/ — graph-oriented retrieval tools for structure and call-path reasoning.
• engine/repobrain_engine/skills/knowledge-layer/ — project knowledge-layer tools for semantic context consolidation.

For local work on this repository itself:

python3 -m venv venv
source venv/bin/activate
pip install -e ./cli -e './engine[dev]'
pytest engine/tests cli/tests