Our verdict
Beacon, from New York start-up Asymptote Labs, captures the full history of your coding-agent sessions (prompts, responses, tool calls, commands, edits and token use) across Claude Code, Cursor, Codex, OpenCode, Cline and, it says, 20+ other harnesses, in one normalised format. On top of that it runs a memory loop: you pick traces, an evaluator or you draft a lesson, you approve it, and future agents can retrieve it over MCP or as an installed Agent Skill. The docs say this loop is deliberately review-gated and nothing is written to memory without your approval. The core is MIT-licensed on GitHub and can run fully locally, with nothing leaving your machine. The catch is the default: the interactive setup preselects Beacon Cloud (labelled 'recommended'), and on Cloud your sessions are used for research and AI training unless you opt out. Pick 'Local only' unless you have read the privacy policy and are allowed to share the code and data your agents touch.
You switch between several coding agents and keep re-explaining the same fixes and repo conventions, and you are happy to run it locally (or have read and accepted Beacon Cloud's terms).
Your sessions contain client or employer code you are not allowed to send to a third party and you might click through the installer defaults, or you only use one agent that already has good built-in memory.
Pricing: As of 21 Oct 2026, per beacon.sh/pricing: Open Source (self-hosted, MIT) $0, no account and 'nothing leaves your machine'. Beacon Cloud $0 (hosted timeline, 50 findings, session search over MCP). Beacon Cloud Pro $20/month billed annually (unlimited findings, agent memories and skills from your sessions over MCP). Enterprise custom. Check site.