Asymptote-Labs/agent-beacon

The problem

If you use more than one coding agent — Claude Code for one repo, Cursor for another, Codex for a quick fix — the knowledge produced in each session dies with that session, and mostly inside that harness. The correction you gave Cursor about how migrations run is not visible to Claude Code tomorrow. Beacon addresses this by recording agent sessions locally and distilling workflows, corrections and debugging patterns into a knowledge store that later agents can retrieve through MCP or Agent Skills.

What's different

The usual approaches here are harness-native memory (Claude Code's own CLAUDE.md/memory files, Cursor rules) and hand-maintained convention docs. Both are tied to one tool and depend on you writing things down. Beacon's angle is capture-first: it watches sessions across 20+ harnesses, then asks you to review what was extracted before it becomes shared memory. The two capabilities that actually matter are cross-harness capture into one store and exact session replay (prompts, tool calls, edits, approvals, token usage) from a durable local JSONL log. The trade-off is a background daemon and a review step; it is heavier than a rules file, and the quality of the extracted knowledge is not something the README proves.

Who it's for

Developers running two or more agent harnesses on the same codebase, or teams that want conventions to survive individual sessions and be shared across agents. It is a reasonable fit for people who already tolerate a local telemetry-style agent. Solo developers using exactly one harness with a well-tuned rules file will get little beyond replay. Anyone who cannot run a background collector on their machine should look elsewhere.

Cost of adoption

Go binary distributed via Homebrew, .deb/.rpm, and Windows MSI; macOS, Linux and Windows are covered. The interactive setup signs in through beacon.sh and preselects Beacon Cloud, so read the confirm screen: choosing Managed forwards history to Asymptote Labs, choosing Local keeps it on disk. Setup also writes skills into ~/.agents/skills and ~/.claude/skills unless you opt out, and cloud mode registers an MCP server in your harnesses. It is MIT licensed, with a CI badge and an active Discord, but the project is only about five months old and the README is long on workflow narrative and thin on extraction quality or evaluation details.

Verdict

Beacon is a young project riding a sharp spike: 2751 stars in roughly five months, and #1 on the weekly Go board with +856 this week. That momentum is real but recent, and the board history before late October is unknown, so treat it as an early-stage bet rather than an established tool. The core idea — session capture as the substrate for cross-harness memory — is sound and the local-first JSONL design is honest about where data goes. Decide mainly on whether you want a collector daemon and a cloud destination in your workflow; if you do, the replay and local mode are useful on their own.