Project

Roadmap

Brevi's direction stays the same: AI-first, with speed and safety as the reasons an agent should run its Python through Brevi. Token savings are added step by step and each one is measured in the release test before it counts. This page lists what is planned, not promised dates.

Now#

Public preview on GitHub. The source, the releases and these docs, under Apache-2.0.
Sandbox on every path (v0.24.1). The allow check runs before every build path, Python libraries need allow native, and a sandboxed program never falls back to CPython. From v0.24.6 only the caller grants capabilities under --sandbox (--allow, --allow-py, the MCP allow and allow_py arguments); the program's own allow lines grant nothing.
Library programs off the typed path (v0.24.6). A program the typed build rejects that imports a library Brevi lacks (requests, pydantic, yaml) takes the dynamic build, and only those imports run in CPython.
The standard library, natively. Standard-library calls no longer need CPython: what the typed build lacks runs in the native dynamic build (v0.24.3), and the common ones move onto the typed fast path module by module (math and time first). In v0.24.4: enum (members, auto(), @unique, str and int mixins), json.JSONDecoder().raw_decode, io.StringIO, path text from os and pathlib outside the sandbox, and bin, hex, oct, fractions.Fraction, re.S/re.M and csv.writer(sys.stdout) on the typed path; all 48 final programs of the v0.24.3 release tests now build natively. In v0.24.6: random outside the sandbox, textwrap and argparse (so command-line programs build natively), bytes values, hashlib and base64. Next: csv.writer on other files, datetime.strptime on the typed path, IntEnum and Flag, then json indent and dataclasses.asdict on the typed path. The table shows where each function runs.
More programs on the fast path. On a corpus of 813 saved agent programs, 565 build typed in v0.24.5 (550 in v0.24.4, 496 in v0.24.3), and v0.24.5 fixes the typed-build wrong answers the v0.24.4 release test found (trains, poker). Next: multi-file programs where one module sends the whole program to CPython.
Weekly releases. One release a week, on Mondays, after that week's release test passes; an important fix may ship sooner. Changes land on main under "Unreleased" in the changelog until then.
Cheaper heavy tasks. Tell agents their Python is compiled so they write the plain solution, and make plain idioms (generators, string indexing) build typed and fast.

Next#

Package managers. pip install brevi, npx brevi mcp, brew install, cargo install brevi and the MCP registry. The packaging is ready.
brevi.run. The compiler as a hosted API for agents: check, run and test over HTTPS, sandboxed, with cached builds. Then a playground in the browser.
More of Python natively. match statements, async, and more standard modules in dynamic builds.

Later#

Windows parity. Library calls and --standalone on Windows; --standalone on macOS.
Faster cold builds, for agents that build many one-off programs.
Larger and fresher benchmarks on new model versions, with results published here (the release test already runs weekly).

Always#

Every program that once passed must keep passing: the full set of saved solutions is rerun on every change and on every new Rust and Python release.
Brevi matches CPython or refuses. A silent difference is treated as the worst kind of bug.

Changes ship in releases listed in the changelog.