跳过主要内容
B

bifrost

brokkai/bifrost

DeepSeek Harness (dsh) bundle for Bifrost: multi-language code intelligence over MCP.

安装

dsh plugin --profile web add github:brokkai/bifrost

README

Bifrost

Bifrost — code intelligence for AI, with structural queries and multi-language analysis designed for repository-scale use

Latest release crates.io PyPI npm Apache-2.0

Quickstart · Documentation · Ten-minute evaluation · Discord

Why Bifrost?

bifrost is Brokk's Rust-based static analysis toolbox for AI coding harnesses, editors, and large repositories.

Bifrost gives every supported language a shared intermediate representation, so the same structural query and navigation workflows work across a mixed-language repository instead of stopping at language boundaries.

  • One multi-language IR. Parse unbuilt or partially broken workspaces and normalize their source structure for cross-language analysis.
  • A real query language. Use JSON CodeQuery or the Rune Query Language (RQL) to find language-neutral code shapes and traverse indexed declarations, references, calls, imports, and type relationships.
  • Built for agents and editors. Expose structured MCP tools to coding agents, LSP features to editors, and the same analyzer through the CLI, Python, and Rust.
  • Designed for active repositories. Snapshot isolation, incremental updates, content-based caching, and git/worktree awareness keep analysis responsive as a repository changes.

See Choose Bifrost for the right interface for your workflow, and the Language and Analysis Capabilities matrix for language-by-language support, precision tiers, and current analysis boundaries.

Run Your First Query

Install the released CLI, clone the small verified evaluation fixture, and run its saved RQL query:

curl -fsSL https://raw.githubusercontent.com/BrokkAi/bifrost/refs/heads/master/install.sh | bash
git clone --depth 1 https://github.com/BrokkAi/bifrost.git
cd bifrost/docs/fixtures/ten-minute-evaluation
bifrost --root . --query-file queries/find-audit.rql

Users with uv or pipx can instead run uv tool install brokk-bifrost or pipx install brokk-bifrost; both install the same native bifrost command from a platform wheel in an isolated environment. uvx brokk-bifrost --version runs it without a persistent install.

Node users can run npm install -g @brokkai/bifrost or use npx -y @brokkai/bifrost --version. The npm package installs the same released native command for the current platform.

The installer downloads the checksum-verified release binary into ~/.local/bin on macOS (Apple Silicon and Intel), Linux (x86-64 glibc or musl, ARM64 glibc), WSL, and Android under Termux. Homebrew users on macOS and Linux (x86-64 and ARM64 glibc) can run brew install brokkai/tap/bifrost instead. On Windows, and on ARM64 musl distributions such as Alpine, build from source with cargo install brokk-bifrost --locked. See Install Bifrost for the full platform table.

The result identifies the normalized Python call and its exact source location:

{
  "isError": false,
  "structuredContent": {
    "results": [
      {
        "enclosing_symbol": "src.app.handle",
        "end_line": 5,
        "kind": "call",
        "language": "python",
        "path": "src/app.py",
        "result_type": "structural_match",
        "start_line": 5,
        "text": "audit(value)"
      }
    ],
    "truncated": false
  }
}

Continue with the ten-minute evaluation to run the same query through the CLI, an MCP-connected coding agent, and VS Code.

See Bifrost in Action

Turn Source into a Query in VS Code

Selecting Python source in VS Code, inspecting its normalized Rune IR, using the generated starter RQL, and browsing the exact query result

Select a source construct and run Bifrost: Show Rune IR to inspect its language-neutral .rune form and get a conservative starter RQL query. Run that query from the editor, browse typed results grouped by file, and jump to the exact source range. The extension also provides definitions, references, hover, rename, symbols, hierarchy, diagnostics, completion, and other LSP features.

Use the Same Analyzer from the CLI

Running a saved RQL query with the Bifrost CLI and receiving an exact structured match

Run saved RQL or JSON queries directly, or call the same named tools exposed over MCP for shell scripts and reproducible analysis workflows.

Language Coverage

Bifrost includes analyzers for C, C++, C#, Go, Java, JavaScript, Kotlin, PHP, Python, Ruby, Rust, Scala, and TypeScript. See the capability matrix for the supported analysis and precision boundaries in each language.

Documentation

The public documentation site lives in docs/ and is published at bifrost.brokk.ai.

Useful starting points:

Run the docs site locally with:

cd docs
npm install
npm run dev

GitHub Pages publication is handled by .github/workflows/docs.yml. Release tag builds publish both the latest docs site and a versioned snapshot under versions/<tag>/.

Repository Layout

The repository is organized around the public product surfaces:

  • Core analysis: crates/ contains the Rust workspace crates; src/ provides the top-level facade and CLI; schemas/ holds shared data contracts.
  • Integrations and distribution: editors/ contains editor integrations; plugins/ contains coding-agent plugins; and npm/, packaging/, and bifrost_searchtools/ expose or package the analyzer.
  • Documentation and examples: docs/ is the documentation site; examples/ contains runnable examples.
  • Analysis data and project defaults: semantic-packs/ contains shipped semantic knowledge; .bifrost/ contains project-level analysis defaults.
  • Development support: benchmark/ covers reproducible measurements; scripts/ and tools/ hold maintenance and release helpers. Root manifests plus .github/, .cargo/, and .config/ define builds, checks, and automation.

License and Commercial Use

Bifrost's public open core is licensed under Apache-2.0 and may be used in research, internal systems, hosted services, and commercial products. Read the License page for the project terms and Third-Party Notices for the additional terms that apply to official binaries and semantic packs.

Contributing

Project contributors   @jbellis @DavidBakerEffendi @foundev @sontek @alexswearingen @joshuadavidthomas

The public repository is projected from Bifrost's open-core source, so its GitHub contributor graph does not reflect everyone who helped build the project.

For local development, test commands, repository-local Python workflow, and release tagging, see CONTRIBUTING.md.

相关插件