Changelog

কী ship হয়েছে।

Every release, measured. New capabilities, benchmark updates, and the security controls behind them.

Changelog

v0.9
Jun 2026Benchmarks

Public head-to-head benchmark

Our first measured comparison against the OpenAI Codex and Google Gemini CLIs — with the full dataset published.

  • 8 bug-fix scenarios on identical checkouts: single bugs, three-at-once, and repos the agents had never seen.
  • 5× fewer tokens and 1.7× faster than Cursor, with every fix verified by a passing test (87/87).
  • Per-scenario data downloadable from the benchmarks page, with full methodology.
v0.8
May 2026New

Native GitHub App

Turn issues and failing checks into verified fix pull requests without leaving GitHub.

  • Install on selected repositories with fine-grained, least-privilege permissions.
  • Every pull request ships the evidence bundle, the patch, and a passing test run.
  • A red CI check can open a fix pull request automatically.
v0.7
Apr 2026Security

Enterprise & deployment controls

Run Codna entirely inside your own infrastructure.

  • Self-hosting, per-tenant key isolation, and bring-your-own-key.
  • Fail-closed egress and secret redaction — your code is never used for training.
  • Managed, VPC, on-prem, and fully air-gapped deployment options.
v0.6
Mar 2026New

MCP server for Cursor and Claude

Give the agent you already use the same deterministic understanding, as a local tool.

  • Codna’s repository graph and evidence bundles exposed over the Model Context Protocol.
  • Drop-in for Cursor and Claude — no change to your workflow.
  • The model reasons over a tight bundle instead of re-reading the whole repo.
v0.5
Feb 2026Performance

Deterministic engine — 250+ languages

Map an entire repository for zero LLM tokens.

  • 130 repositories mapped in 9.2 seconds at 100% ecosystem accuracy.
  • Median 59ms per repo — the dependency and blast-radius graph is built without a model.
  • 162× context reduction: fixes run from ~600 tokens instead of 100k+.
v0.1
Dec 2025Milestone

Initial private beta

The first end-to-end loop: understand, fix, verify.

  • CLI with repository triage and agent-driven fixes.
  • Dogfooding caught and fixed two real bugs — an egress footgun and a harness idempotency issue — shipped with regression tests.
  • Foundations for the MCP server and the GitHub App.