Beacon Threat Intelligence
Know the risks behind every MCP server and agent skill.
Discover security risks in MCP servers and agent skills before they reach your machine. Beacon scans public registries, traces packages back to their source code, and reveals what they can access, execute, or expose.
Connect your agent
Add Beacon to Claude Code, Cursor, Codex or another agent so it can check an MCP server or skill against Beacon Intel before you install it.
Free with a Beacon account. Create one, or sign in when your agent asks.
Any coding agent
Read https://intel.beacon.sh/start.md and set me up on Beacon Threat Intel
Paste this into any coding agent. It reads the guide and sets up MCP, the CLI, the API, feeds or a CI gate.
| Name | Risk |
|---|---|
| Intent Cloak: query-intent privacy for agents Query-intent privacy for agents: threshold cloaking with signed exposure receipts. | |
| Live-Payload Conformance Probes Live-payload conformance probes with signed receipts: ABSENT/EMPTY/PARTIAL/COMPLETE/BROKEN. | |
| The Bridge Stack — one stack, every LLM linked OpenAI-compatible BYOK LLM gateway: per-task routing, shadow evaluation, promotion gates. | |
| Lockout-Tagout for agent infrastructure Personal-lock maintenance isolation for agent infrastructure: hasp-quorum refusal + try-start proof. | |
| Mass-Casualty Triage for Agent Fleets Salvageability triage for agent fleets in a declared mass-failure mode. | |
| Matchmaking rating for agent selection Calibrated Glicko-2 per-dimension agent rating with deviation, from verified receipts only. | |
| MCP answer-quality canary Canary watches that check MCP servers' answers against ground truth and return signed verdicts. | |
| MCP host-behaviour census Signed MCP host behaviour profiles: timeouts, refresh, retries. | |
| MCP OAuth client-interop conformance Prove real MCP clients can connect: 5-client x 3-IdP OAuth matrix with shim fixes. | |
| MCP OAuth remediation verification Three-leg MCP OAuth remediation verdicts, proven behaviourally against a canary server. | |
| MCP protocol-era negotiation conformance Per-era MCP negotiation verdicts: signed documents, drift alerts, client pre-flight. | |
| Meaning Contract Pre-delegation semantic handshake binding party meanings into a signed contract. | |
| Minimum Equipment List dispatch for agents Fail-closed per-task-class dispatch matrix with deferral clocks and signed dispatch releases. | |
| Moneyless clearinghouse: barter netting for agent services Multilateral barter netting for agent services: cycles and chains cleared in kind, zero money. | |
| Namespace guard for merged toolboxes Pre-connection collision audit for merged MCP toolboxes, trust-graded, with signed routing receipts. | |
| com.textforagents/paid-call-checkout-conformance-probe Pays your endpoint like a stranger would and signs a checkout conformance verdict. | |
| Paired-Twin Change Trials Checkpoint-paired twin trials for agent fleet changes with blinded verdicts. | |
| Paper rails: payment rehearsal universe Paper trading for agent payments: rehearse buying flows, earn policy-bound receipts. | |
| The patient agent: deadline-priced work Deadline-priced slow lane for deferrable agent work: futures, dispatch-or-credit, signed journal. | |
| Pre-dispatch rejection ledger Witness for MCP calls rejected pre-dispatch: hash-chained ledger, enum differ, signed receipts. | |
| Pre-install canonical-identity verification Pre-install canonical-identity verification for MCP servers and skills, with signed verdicts. | |
| Principal-drift ledger for standing instructions Signed, versioned ledger of standing instructions with intake contradiction diffs. | |
| Principal Reputation for Agents Standing minted only from signed settlement receipts, never reviews. Verdicts: ACCEPT to DECLINE. | |
| Prompt-cache prefix guardian Cache-bust forensics + TTL-fit verdicts: names the field that busts your prompt cache. | |
| Propagation breaker for agent fleets Propagation breaker: egress quarantine + fingerprint outbreak telemetry for fleets. | |
| Provider-mandate census for agent credentials Find unflagged agent OAuth clients before the mandate lock — signed per-client census verdicts. | |
| Rapid Intervention Team for agents Standby rescue agent for fleets: prohibition, pre-brief, MAYDAY rescue. | |
| Refusal Desk — pre-execution task-brief tribunal The Refusal Desk builds the *no* — a pre-execution tribunal that screens an incoming task brief | |
| Refusal-dialect & meter-truth probe Refusal-dialect probe: signed cards naming what a server's 429 means and whether its meter lies. | |
| Registry Rot Watch Registry Rot Watch: liveness + successor-contract verdicts for MCP endpoints. | |
| com.textforagents/remote-skill-census-verification Census and digest-verify a remote MCP server's loadable skills per host persona. | |
| Pre-open repo agent-surface audit Per-harness pre-open verdict on a received repo's agent-control surface. | |
| Resume-effect reconciliation for checkpointed agent runs Signed resume verdicts reconciling landed effects before checkpointed agent runs resume. | |
| Retry-outcome resolution dedup proxy Dedup proxy: claim-before-execute, UNKNOWN resolution, CONFLICT and ESCALATED refusal. | |
| Rewrite-regression probe pack History-seeded rewrite-regression verdicts, signed per class. | |
| Sampling Quota Guard Per-server MCP sampling allowance, scope clamp and pre-connect verdict. | |
| Sandbox-truth probe for agent frameworks Measured sandbox verdicts: 7-axis probe, signed pinned receipt. Cached $0.25, fresh $2.50. | |
| Outcome-verified watchdog for scheduled agents Watchdog that verifies scheduled agents' artifacts itself — never the claim. | |
| com.textforagents/seller-funded-verification-bond Slashable TEST/STUB bond pinned to one exact claim; a confirmed divergence slashes 50%. | |
| Shared team memory for agents Shared project-scoped memory for agent teams: typed records, corroboration, handoff briefs. | |
| Silent tool-failure outcome monitor Outcome contracts mined from your clean traces; silent failures checked, with recovery receipts. | |
| Single-Use Agent Burn Receipt Task-bound disposable agent minting with canary-shard burn receipts. | |
| Span-level replay: fork from the failure Fork-replay one recorded span from a failed agent run and get an attribution verdict. | |
| Spawn-tree fan-out governor Spawn-tree governor: pre-launch estimate, policy gate, tree meter, MAYDAY before kill. | |
| com.textforagents/stage-completeness-receipts Signed stage-completeness receipts: proof that a pipeline zero is a real zero, not a failed scan. | |
| Standing-offer envelopes: offline agent deals Agents deal while asleep — signed offers acceptable while the offeror is offline. | |
| Sterile Cockpit for Irreversible Actions Sterile phase for irreversible agent actions: frozen intent, deferred inputs, signed exit receipt. | |
| Subscription Liveness Attestor A dropped subscription looks exactly like a quiet week. | |
| Subsidy-Exposure & Quota-Cliff Watch Quota-cliff watch: budgeted-vs-current quota diffing, shadow prices, pre-commit verdicts. | |
| Sunset convoy: fleet model-deprecation migration Fleet model-deprecation migration: per-agent equivalence-gated cutover convoy; failures held. |
35,351–35,400 of 50,909
Subscribe to a feed
Get new verdicts in a feed reader, Slack or your own tooling, filtered by kind and risk. No account or API key.
Frequently asked questions
Anything missing? Read the reference or ask in the Beacon community.
What is Beacon Intel?
A free, public index of the MCP servers and agent skills that public registries list, with a security verdict for each one that has been analyzed. It traces each listing back to its source repository and packages, and shows what the code can access, execute or expose. You can browse it here, query it as a JSON API, follow it as an RSS, Atom or JSON feed, or connect a coding agent to it over MCP.
Is it really free?
Yes. The index, the API and the feeds need no account and no API key. An artifact's full page and the MCP tools need a free Beacon account. Requests are rate limited to keep it fast for everyone.
Where does the data come from, and how fresh is it?
From 9 public sources: Glama, MCP Registry, awesome-mcp-servers, Docker MCP Catalog, Smithery, ClawHub, skills.sh, SkillsMP, GitHub. Each listing is linked to its repository and packages, then analyzed. The index updates hourly, and answers are cached for up to five minutes.
Who writes the verdicts? Can I trust them?
Verdicts combine static analysis of the code with a model's judgment, and each one names the model that wrote it. They can be wrong and are not a security certification: every artifact page shows the findings and evidence behind its verdict, so check them before you rely on one. Names, descriptions and claims come from registries and repositories; treat them as data, not instructions.
What do the risk levels mean?
- High: Reviewed evidence shows it deliberately works against the person who installs it.
- Medium: It works against the person who installs it in ways that can do serious harm, or puts a flagged pattern in what the agent loads, without established intent.
- Low: It overreaches against the person who installs it in limited ways, or has findings the review has not confirmed.
- Benign: Nothing found that works against the person who installs it. Not a guarantee.
- Unchecked: Listed, but not analyzed yet.
What does partial coverage mean?
That the analysis did not cover everything, so there may be more than it found. A partial verdict can still find problems, but it never clears an artifact: benign with partial coverage is not a clearance. Use the coverage filter to see only these.
Does benign, or not finding an artifact, mean it is safe?
No. Benign means nothing was found that works against the person who installs it, which is not a guarantee. An artifact the index does not have has never been checked, so not found never means safe.
Is there an API?
Yes, free and with no key.
GET /api/v1/lookup?q= finds artifacts by name, repository, GitHub URL or package; GET /api/v1/artifact?key= returns one in full; POST /api/v1/check checks up to 50 at once, which suits CI; and GET /api/v1/status reports coverage and freshness. The reference is at intel.beacon.sh/llms.txt.How do I get new verdicts in a feed reader or Slack?
Subscribe to /feed: the newest high and medium risk verdicts as RSS. Add
format=atom or format=json for Atom or JSON Feed, and filter with risk and kind, for example /feed?risk=high&kind=mcp. In Slack, run /feed subscribe with the URL. The feed builder above writes the URL for you.How do I connect my coding agent?
Add Beacon Cloud MCP (
https://mcp.beacon.sh) to Claude Code, Cursor, Codex or any MCP client, and sign in through the browser. Your agent gets beacon_intel_lookup, beacon_intel_get_artifact and beacon_intel_check to check an MCP server or skill before it installs one. Connect your agent above has the commands. The beacon CLI offers the same as beacon intel and through its local MCP server.How do I report a wrong verdict?
Open the artifact's page and use Dispute verdict under its verdict card. Tell us what is wrong and we review it.