# SeenRelay Autonomous agents should start with https://seenrelay.com/agent-bootstrap.json. It is a zero-contact machine descriptor for deciding whether to connect the remote MCP endpoint, run the local prescreen, or leave SeenRelay out. SeenRelay is execution-reuse infrastructure for agents, services, CI workers, browser jobs, scheduled processes and selected IoT/edge reads. Start locally: run the non-mutating prescreen, instrument only a plausible expensive read-only candidate, preserve every authoritative validation in shadow, prefer stronger native/local/source/provider controls, and keep SeenRelay only where the measured workload returns a justified USE rather than DO NOT USE or INSUFFICIENT EVIDENCE. Site and API owners can also use https://seenrelay.com/readiness for a free native-first quick audit of one public HTTPS root response. That surface checks bounded HTTP freshness/validator evidence and recommends native fixes first; it cannot determine SeenRelay workload fit from a surface scan. SeenRelay is currently free and requires no SeenRelay account or API key. The strongest current product fit is a fleet of workers or agents that repeatedly validates the same deterministic state while browser sessions, scraping/extraction, model work, paid search, rate-limited capacity or multi-step validation is materially expensive. SeenRelay should stay out of cheap one-off paths and paths already solved by an equivalent authoritative cache. SeenRelay is a provider-independent execution-reuse and revalidation layer below individual agent frameworks. It coordinates recent source-backed evidence across compatible callers while keeping local/private state, source-native confirmation and the authoritative source ahead of optional shared CHECK. It reports recent observations, not universal truth. It does not browse, search, externally verify arbitrary facts on demand, or use an LLM as a truth arbiter. ## Known-state decision boundary The core question is: `I already know X — do I need to validate X again now?` CHECK is not a lookup for somebody else's raw answer. The caller supplies the known value, deterministic source-backed fact identity and its own freshness window. Canonical starter fact descriptors are published at https://seenrelay.com/starter-facts.json; they contain no observed values, no recommended TTL and no reuse authorization. ## Owned verified update stream Human updates: https://seenrelay.com/updates Machine JSON: https://seenrelay.com/updates.json Atom feed: https://seenrelay.com/updates.atom This is generated from canonical SeenRelay public facts and verified public evidence so agents and operators can follow changes without relying on a social platform or third-party scheduler. ## Value handoff after measured savings Machine-readable status: https://seenrelay.com/commercial.json Current state is non-commercial bootstrap: SeenRelay is currently free, billing is OFF and no purchase is available. A Savings Report is measurement evidence, not an invoice, bill, charge or payment receipt. Agents with a real positive Savings Report may record non-binding willingness-to-pay or procurement intent through the published commercial handoff, but they may not create a payment or activate billing. ## Machine-first agent adoption Autonomous and coding agents do not need to scrape marketing copy to decide whether SeenRelay belongs on a project. Human-readable agent contract: https://seenrelay.com/agents Machine-readable adoption contract: https://seenrelay.com/agent-adoption.json Machine-readable trust posture: https://seenrelay.com/trust.json Machine-readable proof: https://seenrelay.com/proof.json The agent contract requires local prescreening, native-first comparison, shadow measurement with every authoritative call preserved, and explicit project authority before any active reuse. REMOVE is a valid successful outcome. ## First-party operating case study Eight department agents across two tenant identities kept independent role analyses while SeenRelay reduced paid web-search actions from 13 to 4. Measured provider cost fell from $0.16499684 to $0.05560972 in the paired first-party run. One tenant was real SeenRelay operating work and the other was an explicitly synthetic external-client rehearsal with real provider execution. This is not customer ROI or net customer savings. Human-readable case study: https://seenrelay.com/case-studies/company-fleet Machine-readable verified savings record: https://seenrelay.com/case-studies/company-fleet.json ## First proof: measure without changing application behavior ## Add SeenRelay to Cursor, Claude Code, or another coding agent Claude Code persistent repository-hosted plugin: `claude plugin marketplace add ovladon/seenrelay` `claude plugin install --scope user seenrelay@seenrelay` This self-hosted path is validated and available now. It does not imply Anthropic marketplace approval, does not automatically attach the hosted MCP endpoint, and does not enable reuse. GitHub Copilot CLI optional repository-hosted plugin: `copilot plugin marketplace add ovladon/seenrelay` `copilot plugin install seenrelay@seenrelay` The Copilot route is self-serve but has not been independently verified end to end. It is not a claim of inclusion in a default or curated marketplace, does not enable reuse, and requires normal project authority. Other compatible Agent Skills clients can install the published SeenRelay skill directly from the canonical domain: `npx skills add https://seenrelay.com --skill seenrelay --yes` Then ask the agent: `Run a SeenRelay shadow audit on this project. Find repeated expensive read-only validations, preserve every authoritative call, measure stronger local/source/provider-native controls first, do not enable reuse, and return USE / DO NOT USE / INSUFFICIENT EVIDENCE for each measured workload.` JavaScript / TypeScript 0.2.30: `npm install seenrelay` `import { ambientMcpClient } from 'seenrelay/ambient';` `const client = ambientMcpClient(rawMcpClient);` `console.log(client.seenRelayAmbient.getReport());` Python 0.2.30: `pip install seenrelay` `from seenrelay_ambient import ambient_mcp_client` `client = ambient_mcp_client(raw_mcp_client)` `print(client.get_report())` Both Ambient entry points are shadow-first: run the existing workload normally and inspect the local report before enabling bounded reuse. For full safety/economics evidence, use the current Shadow Proof and hostile-economics surfaces documented by the installed client. A native path that answers the same question more cheaply or more strongly wins. Audit method: https://github.com/ovladon/seenrelay/blob/main/docs/SHADOW_AUDIT.md ## Public install - JavaScript / TypeScript: npm install seenrelay - Python: pip install seenrelay - Client version: 0.2.30 - Required base runtime dependencies: 0 - Account/API key required by SeenRelay: no - Current SeenRelay API fee: $0 - Canonical direct MCP endpoint: https://seenrelay.com/mcp ## Product decision rule - Start in shadow mode and preserve every authoritative validation. - Measure exact recurrence on the caller's real workload. - Keep stronger local, source-native and provider-native controls ahead of SeenRelay. - Return USE, DO NOT USE or INSUFFICIENT EVIDENCE for each measured workload. - Enable reuse only where safety equivalence and net economics are demonstrated. - SeenRelay remains free while utility and repeatable workload ROI are being established. Canonical machine-readable product facts: https://seenrelay.com/product-facts.json Technical verification history: https://github.com/ovladon/seenrelay/blob/main/docs/VERIFIED_RESULTS.md ## Fleet path after a positive audit JavaScript / TypeScript and Python Zero-State private L1 are caller-owned storage paths for exact completed-result state across workers or process restarts. The store receives an opaque SHA-256 coordinate key and a sealed payload; the encryption key stays in the caller's secret-management boundary. This is not a claim that the public SeenRelay relay is a hosted private tenant store. A positive `privateMaxAgeMs` is an explicit caller freshness decision. With `privateMaxAgeMs = 0`, a private result does not suppress source validation, while retained ETag/Last-Modified state can still support conditional source confirmation. See https://seenrelay.com/fleet for the current deployment pattern. ## Preferred Zero-State order - exact in-flight / explicit-TTL local reuse; - optional encrypted caller-owned private L1 across workers/restarts; - source-native ETag / Last-Modified confirmation or a stronger authoritative mechanism; - provider-native cache when it solves the same semantics; - optional shared SeenRelay CHECK when share policy permits and it adds value; - original validation fallback; - OBSERVE only after fresh independent validation. For an existing MCP client, start with `seenrelay/ambient` for behavior-preserving measurement. After an eligible read-only tool is reviewed, `seenrelay/mcp-auto` provides the local-first bind-once protection path. Shared CHECK is off by default in Zero-State. Completed-result TTL defaults to zero. Provider-specific adapters are optional integrations and are not dependencies of SeenRelay Core. ## Python Zero-State / classic clients Python 0.2.30 includes an explicit provider-independent Zero-State path for caller-controlled read-only validation. The classic Python API and Python Ambient adapters remain shadow-first by default. The classic JavaScript / TypeScript API also remains available for CHECK-first measurement and explicit bounded shared-evidence reuse. ## Use SeenRelay when - the caller already retains a known source-backed value and naturally needs to decide whether to validate it again; - a repeated read-only validation has deterministic identity and meaningful cost or latency; - the same exact source-backed state or read-only execution repeats across runs, workers, agents, services, CI jobs or edge gateways; - local/private/source-native reuse can avoid work before shared evidence exists; - the validation consumes browser/proxy work, metered scraping or extraction, model work, paid search, rate-limited API capacity or a multi-step chain; - the application can preserve the authoritative fallback and define an explicit freshness policy. ## Do not use SeenRelay when - the operation is mutating or destructive; - it is a cheap one-off request with little repeat probability; - an equivalent authoritative cache or source-native validator already answers the same user-relevant question more cheaply; - policy requires brand-new live source confirmation every time and no conditional shortcut is useful. ## Canonical interfaces - Website: https://seenrelay.com/ - Shadow measurement method: https://github.com/ovladon/seenrelay/blob/main/docs/SHADOW_AUDIT.md - Site/API owner quick readiness audit: https://seenrelay.com/readiness - Execution substrate: https://seenrelay.com/substrate - AI agent cost optimization: https://seenrelay.com/ai-agent-cost-optimization - Duplicate tool calls / tool call deduplication: https://seenrelay.com/duplicate-tool-calls - Langfuse local TOOL census: https://seenrelay.com/langfuse - Fleet deployment: https://seenrelay.com/fleet - Quickstart: https://seenrelay.com/quickstart - Canonical starter facts for known-state revalidation: https://seenrelay.com/starter-facts.json - Client integrations: https://seenrelay.com/clients - Controlled Proof Atlas: https://seenrelay.com/proof - Use-case Atlas: https://seenrelay.com/use-cases - Machine-readable Use-case Atlas: https://seenrelay.com/use-cases.json - Machine-readable Proof Atlas: https://seenrelay.com/proof.json - Economics and measured examples: https://seenrelay.com/economics - Product facts: https://seenrelay.com/product-facts.json - Client source: https://github.com/ovladon/seenrelay/tree/main/clients - Machine descriptor: https://seenrelay.com/service.json - OpenAPI: https://seenrelay.com/openapi.json - MCP endpoint: https://seenrelay.com/mcp - Agent Skill index: https://seenrelay.com/.well-known/agent-skills/index.json - Agent Skill: https://seenrelay.com/.well-known/agent-skills/seenrelay/SKILL.md - Legacy Agent Skill discovery fallback: https://seenrelay.com/.well-known/skills/index.json - MCP Registry: io.github.ovladon/seenrelay - Trust: https://seenrelay.com/trust - Privacy notice: https://seenrelay.com/privacy - Data practices: https://seenrelay.com/data-practices.json - Public aggregate metrics: https://seenrelay.com/public-stats.json ## Hosted operations - CHECK / check_fact: compare a caller-known value with recent observations for the same deterministic source-backed fact. - OBSERVE / observe_fact: contribute a value only after the caller independently observed it while doing its own work. The hosted service has no third domain operation. Intermediary cache reuse must not be re-labeled as a fresh independent OBSERVE.