
Self-improving memory for coding agents
Beacon is an open-source, self-improving memory layer for AI coding agents. It captures agent sessions across Claude Code, Cursor, Codex, OpenCode and 20+ other harnesses, unifies them into one searchable timeline, and uses Jev to identify which runs are worth learning from. It turns high-signal workflows, corrections, and debugging patterns into reusable skills, flags security risks, and surfaces costly behavior patterns — so every agent starts smarter than the last.
Captures prompts, commands, diffs and token usage from Claude Code, Cursor, Codex, OpenCode and 20+ other tools, putting every agent and machine into a single timeline.
Beacon reviews sessions and surfaces three kinds of findings: security risks (e.g. secrets pasted into prompts, destructive commands, permission bypasses, unknown MCP server calls), reusable skill drafts from repeated workflows, and behavior patterns like retry loops, abandoned sessions, and expensi
Turn findings into downloadable reusable skills that any agent can use, and fix prompts in one click.
Agents can search your past sessions over MCP; Pro and higher plans add full MCP access to sessions, findings and feedback.
See where your token spend goes across agents and machines.
On Pro and above, agents can save and recall memories over MCP and pull skills straight from your sessions.
Share any session with a link.
Daily digest on the free plan, real-time email alerts on Pro and above (plus custom alerts on Enterprise), with a full evidence trail on every finding.
Beacon captures every agent trace across your environment via the Beacon CLI (available for macOS, Windows, and Linux; installable via Homebrew) with zero infrastructure to manage on the cloud plans.
Beacon uses Jev to cheaply evaluate and classify agent runs at scale, identifying which sessions are actually worth learning from.
High-signal workflows, corrections, and debugging patterns are distilled into reusable Agent Skills, memories agents can recall over MCP, and findings with concrete fixes — creating a compounding feedback loop across all your agent tools.
Free tier forever with 50 findings; Pro at $20/month (billed annually); Team at $30/seat/month billed annually plus a $10/month base fee for up to 20 seats ($40/seat/month shown in the comparison table); Enterprise with custom pricing via demo.
$0 forever
20+ agents supported, every agent and machine in one timeline, token spend visibility, 50 findings (risks, skills, habits), one-click skills and prompt fixes, MCP session search for your agents, and session sharing by link.
$20/month (billed annually, save 25%)
Everything in Cloud plus unlimited findings, full MCP access (sessions, findings and feedback), agent memories over MCP, skills pulled from your sessions, full evidence trail on every finding, real-time alerts, and priority support.
$30/seat/month billed annually + $10/month base fee, up to 2
Pro for every member in one shared workspace: shared trace namespace, privacy settings for what your team shares, org API keys and agents for CI and bots, members and admin roles, and one invoice for every seat.
Custom (book a demo)
Everything in Team plus block or require approval for risky actions, org-wide findings, agents on endpoints/CI/cloud, fleet rollout through Jamf, Intune or Fleet, SSO/RBAC and audit logs, your own bucket and region, and dedicated SLA and su
Beacon captures agent sessions across Claude Code, Cursor, Codex, OpenCode and 20+ other harnesses into one timeline, then turns that history into a compounding knowledge layer: security findings, reusable skills, agent memories, and behavior insights so every agent starts smarter than the last.
Yes. Beacon is open source and free to get started, with no credit card required. It's described as running locally via the Beacon CLI, with optional cloud plans on top.
Beacon works with 20+ agent harnesses, explicitly including Claude Code, Cursor, Codex, and OpenCode.
Beacon Cloud is $0 forever with 50 findings included. Cloud Pro is $20/month billed annually. Cloud Team is $30/seat/month billed annually plus a $10/month base fee for up to 20 seats. Enterprise pricing is custom and requires booking a demo.
Findings are insights Beacon surfaces from your agent sessions in three categories: Security (risky actions like pasted secrets, destructive commands, or permission bypasses), Skill (drafted reusable skills from repeated workflows), and Behavior (costly habits like retry loops, abandoned sessions, or expensive models on simple tasks — each with a suggestion to work faster or cheaper).
Install the Beacon CLI for macOS, Windows, or Linux — for example via Homebrew: `brew tap asymptote-labs/tap && brew install beacon && beacon endpoint install`. A setup prompt can also be copied for configuration, and you can sign up free for the cloud.
Yes. All plans let your agents search your past sessions over MCP. Pro and above add full MCP access to sessions, findings and feedback, plus agent memories and skills pulled straight from your sessions.
Yes. The Team plan adds a shared namespace, privacy settings, org API keys, and admin roles. Enterprise adds governance features like blocking or approving risky actions, org-wide findings, fleet rollout through Jamf/Intune/Fleet, SSO/RBAC and audit logs, your own bucket and region, and a dedicated SLA.