Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

Beacon

Open-source recorder for 20+ coding agents that turns reviewed sessions into project memory and skills, with a paid cloud on top.

Reviewed by ReviewRamble editors · Updated

View on GitHub Open source

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

3.8/5

Beacon keeps one clean record of every coding agent you use; the memory it builds is only as good as the lessons you stop to approve.

What is Beacon?

Beacon, from Asymptote Labs, records what your coding agents do across Claude Code, Cursor, Codex, OpenCode, Cline and more than 20 other harnesses, plus Claude.ai and ChatGPT through a browser extension, CI jobs and some cloud agents. It normalises prompts, responses, tool calls, commands, file edits, approvals, MCP activity and token use into one OpenTelemetry-based event format and writes them to a local JSONL file. A terminal browser and a local web dashboard read that file, and it can be forwarded to Splunk, Datadog, Elastic, Microsoft Sentinel, S3 and other destinations.

The memory layer sits on top. You pick traces, optionally score them with an external evaluator, review candidate lessons and approve them into a per-project memory store, which agents read back through a read-only local MCP server or as installed Agent Skills. The same loop runs over MCP in the hosted Beacon Cloud on its Pro plan.

The Go CLI is MIT-licensed. The repository was created in May 2026, had about 3,567 GitHub stars on 18 November 2026, and ships releases every few days; v1.3.63 came out on 17 November 2026. Asymptote Labs sells Beacon Cloud plans on top of the open-source build.

Is Beacon worth it?

Beacon is a strong session recorder with a memory feature attached, and the memory works only as well as your review habit. Capture is the better half. One install hooks into more than 20 harnesses, the collectors listen only on 127.0.0.1 ports 4317 and 4318, and the raw history lands in ~/.beacon/endpoint/logs/runtime.jsonl, where any tool can read it. Moving between Claude Code and Codex no longer leaves two separate histories.

Memory is deliberately manual, and nothing runs from hooks. You run beacon memory evaluations run on chosen traces, which by default sends a bounded, redacted projection of each trace to TypeSafe's hosted Jev evaluator. A trace becomes a candidate only if its task-success probability is at least 0.50 and the mean of three probabilities is at least 0.60. The evaluator returns scores, not text, so you or your agent still write the lesson before approving it. Approved memory is scoped to one repository and kept in a local memory.db; cross-project memory is not automatic. Teams that forbid external calls can skip scoring or point it at an internal evaluator.

The default setup leans towards the cloud. Interactive install signs you in through beacon.sh with Beacon Cloud preselected, and confirming that choice starts forwarding at once, including the last 30 days of sessions on the device. Standard mode uploads prompts, responses and tool output after local redaction; Metadata-only strips that content on the device first; Local keeps everything on the machine. The free cloud allows 50 findings over an account's lifetime, and memory over cloud MCP and skills drawn from sessions need Pro at $20 a month, or $15 billed annually.

Two more defaults deserve a look. In the default shared mode, any local account can read and rewrite the system log, so multi-user machines need the restricted log access setting. The optional browser extension keeps full Claude.ai and ChatGPT prompt and response text by default, which the docs call the most sensitive default in the product. Compared with claude-mem, which has a model summarise every session automatically, Beacon costs nothing to run until you ask it to score traces, but it also remembers nothing until someone approves a lesson.

Where it earns its place

  • A solo developer who moves between Claude Code and Codex when usage caps run out

    Stop re-explaining the project every time the agent changes

    Install in Local mode, let Beacon capture both harnesses, and approve lessons from good sessions into project memory that either agent recalls through beacon mcp serve. Local mode is free; scoring traces with Jev needs an evaluator key.

  • A platform team of 15 engineers on Cursor, Claude Code and Copilot CLI

    Turn a fix one engineer finds into a skill the whole team loads

    Beacon Cloud Team, at $30 a seat a month plus a $10 base fee, gives a shared namespace for traces; an approved memory renders as a SKILL.md the team commits to the repository.

  • A security team at a 300-person software company

    See what coding agents actually ran on developer laptops

    Deploy in system mode with restricted log access, forward runtime.jsonl to the existing SIEM such as Splunk or Sentinel, and run the offline threat rules in beacon scan. Fleet rollout through Jamf or Intune and blocking risky actions are Enterprise features at custom pricing.

Before you commit

Setup
brew install beacon and beacon endpoint install on macOS; a .deb or .rpm through a checksum- and signature-verified script on Linux; an MSI on Windows. Choose Local in the wizard to keep data on the machine. It also installs Beacon Skills into ~/.agents/skills and ~/.claude/skills unless you turn that off.
Integrations
Claude Code, Cursor, Codex, OpenCode, Cline, Gemini CLI, GitHub Copilot CLI, OpenClaw, Hermes Agent and 20+ more; Claude.ai and ChatGPT through a browser extension; LiteLLM, Envoy and Portkey gateways; Anthropic, OpenAI, Claude Agent SDK and Vercel AI SDK; Splunk, Datadog, Elastic, Sentinel, Sumo Logic, Wazuh, S3 and GCS.
Data and privacy
Local mode keeps traces in runtime.jsonl and memory in memory.db on the machine, with redaction and size limits applied before writing. Beacon Cloud receives Standard (redacted content) or Metadata-only data over HTTPS with a revocable per-device key. Memory scoring sends redacted trace projections to the evaluator you choose.
Cost at scale
Open source and Beacon Cloud are free; Pro is $20 a month ($15 annually); Team is $30 per seat a month plus a $10 base fee ($5 annually) for up to 20 seats; Enterprise pricing is custom.
Limits
Coverage varies by harness: Cursor reports no token counts and Codex no file events. Memory is per project and needs manual review. The free cloud stops analysing new activity after 50 findings. About 40 issues and pull requests are open, and releases land every few days.

What’s good

  • Captures 20+ harnesses into one local JSONL format
  • Collectors bind to localhost; Local mode sends nothing
  • Review-gated memory, readable over MCP or as Agent Skills
  • Forwards to Splunk, Datadog, Elastic, Sentinel and S3

What’s not

  • Installer preselects Beacon Cloud forwarding
  • Memory scoring calls an external evaluator by default
  • Cloud memory and session-drawn skills need Pro at $20 a month
  • In shared mode any local account can read the system log

Best for

Developers and platform teams who switch between several coding agents and want one auditable record of what they did, with memory they curate themselves.

Beacon alternatives

Beacon: common questions

Is Beacon worth it?
Beacon keeps one clean record of every coding agent you use; the memory it builds is only as good as the lessons you stop to approve. We rate it 3.8 out of 5.
How much does Beacon cost?
Free and open source (MIT). Beacon Cloud is free with 50 findings; Pro $20/month ($15/month billed annually); Team $30/seat/month plus a $10/month base fee ($5 billed annually), up to 20 seats; Enterprise custom.
What are the downsides of Beacon?
Installer preselects Beacon Cloud forwarding; memory scoring calls an external evaluator by default; cloud memory and session-drawn skills need Pro at $20 a month; in shared mode any local account can read the system log.
Who is Beacon best for?
Developers and platform teams who switch between several coding agents and want one auditable record of what they did, with memory they curate themselves.
How much does Beacon cost for a team or in production?
Open source and Beacon Cloud are free; Pro is $20 a month ($15 annually); Team is $30 per seat a month plus a $10 base fee ($5 annually) for up to 20 seats; Enterprise pricing is custom.
Is Beacon safe for company data?
Local mode keeps traces in runtime.jsonl and memory in memory.db on the machine, with redaction and size limits applied before writing. Beacon Cloud receives Standard (redacted content) or Metadata-only data over HTTPS with a revocable per-device key. Memory scoring sends redacted trace projections to the evaluator you choose.
What are the best Beacon alternatives?
The closest alternatives we have reviewed are claude-mem, Memorilla and Mem0.

Answers checked Nov 18, 2026.