Birdview makes coding agents map a project's architecture before they edit it

Birdview is a new open-source tool that changes the order coding agents work in: map the project's architecture first, list the modules a change will touch, then edit. It reached 159 stars on GitHub in its first 36 hours. The licence is MIT and version 0.1.0 was tagged on 2026-09-12.

What it asks the agent to do

The workflow has two stages. First, the agent inspects the project and writes an architecture.json file that lists modules, ownership, evidence for each claim, and how the modules relate. Second, for every code-changing task, the agent writes an entry to activity.jsonl declaring the planned scope, the target modules, the files it will edit, the lifecycle phase, and the check results, all pinned to a specific revision of the architecture map.

Both files are validated against JSON Schemas and against cross-record rules: stable map identity, contiguous sequences, valid targets, correct file ownership, and consistent check results. The output is a standalone HTML view that runs with no server and no network. The project's own words: "See AI changes before they happen."

Where it fits

The project defaults to an "auto" mode, written into a project's AGENTS.md file, which asks the agent to inspect and update the map, render it, and declare affected modules before every edit. An "on-demand" mode requires an explicit request. Both are guidance rather than a write interceptor: the agent still holds the pen, but it now has to state where it is drawing before it draws.

That framing is honest about the current limit. The README says out loud that "activity is declared by an agent; Birdview does not automatically observe coding operations." A completed event does not prove the checks passed, only recorded check results do. There is no live transport and no automatic refresh: updates require regenerating the HTML.

What is missing from v0.1

The README lists the constraints plainly. The package is marked private and is not published to npm. Live view is manual. Bilingual authoring is supported for Chinese and English; other languages are user content. There are no benchmarks in the repository, and the project does not claim any measured improvement in code quality, only that "forced context check exposes affected modules early, reduces blind changes, and keeps implementation aligned with the system around it."

For a team using Claude Code, Codex, or Cursor on a repository large enough that the agent's edits sometimes touch modules outside the intended blast radius, that trade may still be worth trying: the cost is a JSON file per project and a JSONL append per task, and the return is a document the human can read before the change lands. The pattern is closer to a pre-flight checklist than to an observability tool. It only works if the agent is honest in its declarations, which is a real dependency, and it depends on the map staying in step with the code, which will be work.

Read the traction honestly. 159 stars in 36 hours is early attention, not proof of use, and there is no independent report of a team running it against a real codebase yet. The cost of trying it is low.

Source

Primary source: Birdview on GitHub.