AI for large codebases. We build the analysis, runtime, and agent layers that let coding agents work on repositories too big to fit in a context window.
Everything below is open source and independently usable. Start with whichever layer matches your problem.
| bifrost | Multi-language static analysis for agents, editors, and large repositories. One IR across languages, a real query language (RQL), and MCP/LSP/CLI/Python/Rust interfaces. Apache-2.0 · docs |
| code-semantic-model-interchange | Experimental, language-neutral specification for portable semantic models of code. Apache-2.0 · docs |
| csmi-demo | Reproducible interoperability demos comparing static-analysis results with and without the same CSMI semantic pack. Apache-2.0 |
| quicksilver | Impact-selected mutation testing for Cargo workspaces: each mutant runs only the tests that execute it. LGPL-3.0 |
| bifrost-policy-scan | GitHub Action for Bifrost policy scans: SARIF upload, diff-aware PR gating. |
| bifrost-extension-template | Reference template for building your own analysis extensions on Bifrost's stable APIs. |
| mjolnir | Terminal client for ACP coding agents, with model-first routing and a multi-agent coding council. GPL-3.0 · docs |
| anvil | Portable Agent Client Protocol (ACP) server: model routing, tools, permissions, sandboxing, MCP. LGPL-3.0 · docs |
| hel | ACP session manager and remote execution environment. |
| muse-acp | Use a Meta Muse Code subscription from Zed, IntelliJ IDEA, and other ACP clients. |
| codex-acp | ACP server for Codex CLI, maintained for smoother client and IDE integration. |
Small, composable services for putting ACP agents to work on GitHub repositories.
| brokk-town | Local browser and terminal control plane for coordinated issue, review, release, and discovery workflows. |
| issue-bot | Walks GitHub issues, prepares fixes with a configurable ACP agent, and opens pull requests for review. |
| review-bot | Reviews pull requests with an ACP agent and independently verifies findings before posting them. |
| release-bot | Drives repository releases and verifies publication before recording success. |
| bug-bot / feature-bot | Discover useful bugs and feature opportunities, deduplicate them, and file reviewed GitHub issues. |
We publish our benchmarks rather than only our numbers.
| powerrank | The Brokk Power Ranking of LLM coders. |
| usagebench | Analyzer-neutral benchmark for finding usages of a code unit. MIT · results |
| dataflowbench | Benchmark for value flow, taint tracking, typestate, and witness quality across languages and tools. MIT |
- You want code intelligence in your own agent or editor -> bifrost, then the ten-minute evaluation.
- You want portable library semantics across analyzers -> CSMI, then the interoperability demos.
- You want faster mutation testing for a Cargo workspace -> quicksilver.
- You want to run coding agents from a terminal -> mjolnir.
- You are embedding an agent runtime in a product -> anvil.
- You want to use Muse Code or Codex through an ACP client -> muse-acp or codex-acp.
- You want agents to maintain a GitHub repository -> start with Brokk Town, or run one of the focused repository bots directly.
- You are comparing models or analyzers -> powerrank, usagebench, dataflowbench.
Discord · brokk.ai · Issues and PRs welcome on any repository above.