Beacon aims to make a correction learned in one coding agent useful to the next—not to lower the model’s token rate. Its creators introduced the project on Hacker News on October 21, describing the frustration of re-explaining work whenever they switched between Claude Code, Codex and Cursor.
What Beacon actually does
The official repository describes shared prompts, responses, tool calls and file activity across agent applications, without replacing their models.
Beacon normalizes captured activity, then makes reviewed lessons available through Model Context Protocol (MCP) tools or Agent Skills. The documentation lists more than 20 harnesses, but its coverage matrix has gaps: do not assume every integration captures approvals, token usage or identical session detail.
The memory workflow is explicit: select traces, create candidates, review them, approve useful lessons, then retrieve or promote them into project skills. Local memory is project-scoped, not automatically shared across unrelated repositories.
What changes—and what still costs money
| Cost or workflow | Separate agent histories | With Beacon |
|---|---|---|
| Repeated explanations | Re-enter conventions after switching tools | Retrieve approved project lessons |
| Agent access and inference | Existing subscription or usage terms | Those bills still apply |
| Memory processing | Manual notes or each tool’s features | Local review; optional external evaluation |
| Storage and operations | Existing tool arrangements | Local retention or chosen forwarding destination |
| Pricing data: How we verify → | ||
How we verify pricing and cost boundaries.
Beacon’s MIT license permits open-source use. That is not a promise of a cost-free end-to-end workflow. The plan documentation distinguishes local Open Source, Beacon Cloud, paid Beacon Cloud Pro and Enterprise. Hosted memory over MCP is a Pro/Enterprise feature; the local memory path is separate. The separate snapshot covers billing and quotas, not inference tokens. Annual equivalents require annual billing; checkout and tax terms may differ.
Optional local evaluation can call an external evaluator such as TypeSafe Jev. A dry run previews selected traces and estimated calls without a network request. The evaluator returns probabilities—not the lesson itself; a person or agent still drafts the lesson for review. Manually authored candidates provide another path.
Neither the HN post nor the reviewed documentation establishes a measured percentage reduction in token spending. Retrieval can save repeated explanation, but evaluation, drafting and retrieved context can also add usage.
Who benefits, and where the trade-off falls
Likely beneficiaries: developers who frequently switch tools, teams repeating the same repository-specific corrections, and maintainers who want inspectable project knowledge outside one harness.
Less compelling: short-lived projects, effective existing shared instructions, or traces whose review costs more than it saves.
Beacon could reduce the friction of switching harnesses. It does not transfer subscription allowances, reset rate limits, guarantee cheaper inference or turn a weaker model into a stronger one.
Check the data boundary before enabling capture
“Local-first” needs a setup decision. The README says interactive setup preselects Beacon Cloud and offers an explicit Local choice. Signing in alone does not forward data; confirming the managed option connects the endpoint and starts forwarding.
For a local-only pilot, choose Local and inspect the destination settings. External evaluation is another boundary: the docs describe bounded, redacted trace projections sent on explicit evaluation commands. Local memory recall does not itself require that evaluator call.
Do not treat automatic redaction as proof that every sensitive item is removed. Review what is captured and retained before using confidential repository sessions. Approved memory should contain reusable conventions, not credentials or customer records.
What to do now
- Pilot one repository. Capture a representative task and inspect the trace before creating memory.
- Approve one concrete lesson. Try a testing convention or a correction you repeatedly explain; check whether a second supported agent retrieves it correctly.
- Measure accepted-task cost. Include retries, evaluator calls, drafting, retrieved-context tokens and review time; compare the same tasks with and without memory.
- Keep a rollback path. Remove obsolete lessons when repository conventions change, and verify the harness integrations you actually use.
For the underlying model budget, use our OpenAI pricing page, Anthropic pricing page and token calculator. The AI API cost guide explains why tokens alone do not capture workflow economics. For a separately configured model backend, compare Novita inference options. Not a required purchase or verified Beacon integration. Affiliate disclosure: we may earn a commission.
Labs availability and ongoing verification
Beacon is not a callable inference model. Our Labs coverage notes explain the missing authorized corpus, pinned versions, consented traces, quality grader and cost ledger. No installation, trace upload, purchase or paid benchmark was performed.
The separate snapshot is checked daily against official sources; changed or unavailable sources freeze its verification clocks pending review.
Sources checked October 22, 2026: official pricing, the linked HN creator post, official repository and license, memory documentation, Beacon Skills documentation, and plan documentation. We have not independently tested integration coverage or savings.