Beacon
Cross-harness observability and self-improving memory for coding agents
About Beacon
Beacon records what your coding agents actually did. The MIT-licensed CLI captures sessions across Claude Code, Cursor, Codex, OpenCode and more than twenty other harnesses, and puts them into a single timeline that spans every agent and every machine you work on. What separates it from a plain memory store is what it does with that record: it mines sessions for three kinds of finding. Security findings flag risky things an agent did, such as a secret pasted into a prompt, a destructive command, an agent bypassing its permissions, or a call to an unknown MCP server. Skill findings notice when you have walked an agent through the same steps more than once and draft a reusable skill from it. Behaviour findings spot habits that cost time or money: retry loops, abandoned sessions, approvals you always grant, expensive models used for simple tasks.
Two things are worth weighing before adopting it. The first is maturity: the repository is MIT with roughly 3,400 stars and was created in May 2026, so this is an early project with a loud launch rather than an established one, and its homepage social proof is almost entirely posts on X, several from accounts whose main subject is crypto. The second is shape. Beacon is open core, and the money is upstream: both the Enterprise and Book Demo links leave beacon.sh for Asymptote Labs, whose product is organisation-wide agent governance. The pricing page also contradicts itself, with the Team card reading $30 per seat per month while the comparison table below it reads $40.
You install the CLI with Homebrew on macOS, or the equivalent on Windows and Linux, and run the endpoint install step. From then on it records sessions from whichever supported harness you use, and they all land in one timeline rather than in per-tool silos. Beacon Cloud reviews those sessions and raises findings against them, each one showing what happened, which session it was in and how to fix it, with a one-click path from a finding to a downloadable skill or a corrected prompt. Your agents can query the record back over MCP: session search is available on the free tier, and the paid tiers add findings, feedback, saved memories and skills pulled straight from past sessions. Sessions can be shared by link, and spend is broken down so you can see where tokens went. The enterprise tier moves from observing to intervening, blocking or requiring approval for risky actions and rolling the agent out across a fleet through Jamf, Intune or Fleet.
- •One timeline across 20+ harnesses - Claude Code, Cursor, Codex, OpenCode and others on every machine you use, in one record rather than per-tool silos
- •Security findings - Flags secrets pasted into prompts, destructive commands, permission bypasses and calls to unknown MCP servers, each with the session it came from
- •Skills drafted from repetition - Notices workflows you have walked an agent through more than once and turns them into a reusable skill
- •Behaviour and cost findings - Surfaces retry loops, abandoned sessions and expensive models on simple tasks, with a suggestion attached
- •MCP access to your own history - Agents search past sessions over MCP on every tier, with findings, memories and skills on the paid ones
- •Open source CLI - MIT licensed, so what is collected is auditable rather than taken on trust
Developers running several coding agents across more than one machine who have no idea what those agents have collectively done, and teams who have started to treat agent sessions as a security surface rather than a private scratchpad. The free tier is a real one, not a trial, and fifty findings is enough to see whether the findings are any good before paying. The security angle is the strongest reason to look at it, and the clearest reason a team lead rather than an individual would adopt it. It overlaps heavily with the single-harness memory tools, so if you only want one agent-memory tool, the question is whether cross-harness coverage matters more than maturity. If you work in one harness on one machine, a simpler memory engine will cover you with less to run.
Pricing
- Beacon Cloud$0/mo
- Beacon Cloud Pro$20/mo
- Beacon Cloud Team$30/mo
- EnterpriseContact sales
From the vendor pricing page, 2026-11-06
