A small XML file, linked from a page with rel="search", that gives a URL template for searching the site so browsers and agents can run a search without filling in a form. CrawlCheck's agent task canaries look for one when they try to search a store's catalogue. It declares how to search; it does not show the search returns useful results.
The W3C Trace Context header that carries a trace id and a span id across services, so one request can be followed through every system it touches. CrawlCheck accepts it on incoming requests and returns one with every error, so a reported problem can be matched to its record. It links the steps; it does not record what happened in them.
A vocabulary for describing the shape of JSON data: which fields exist, their types, which are required and which values are allowed. CrawlCheck publishes schemas for its API answers under /schemas/ so a client can validate a response before acting on it. A response that passes is well-formed; passing does not make the values correct.
A JSON document, defined by RFC 9728, that an OAuth-protected server publishes to say which authorisation servers issue its tokens, which scopes exist and which resource a token is for. MCP servers that require a login are expected to publish it, so a client can discover how to log in. A server without it leaves the client guessing, and a client that does not check its resource field can be pointed at another host.
The resource an access token is valid for. A narrow audience (exactly the endpoint) limits what a leaked or over-shared token can do; an audience of the whole host makes a token issued for one tool usable across the rest of the service. Clients should compare the audience a server declares with the endpoint they connected to and refuse a mismatch.
Proof Key for Code Exchange: an OAuth extension in which the client sends a hash of a secret when it starts a login and the secret itself when it collects the token, so an intercepted login code is useless to anyone else. OAuth 2.1 and the MCP authorisation specification expect the S256 method. A server that does not advertise it gives clients no way to know it is supported.
A long-lived credential that lets a client obtain new access tokens without the person logging in again, usually granted when a client requests the offline_access scope. Useful for software that runs for months; more than a one-off agent task needs. A server that requests it by default gives every agent a grant that outlasts the conversation.
The set of permissions a client asks for when the server gives it no narrower instruction. If a login challenge names no scope, many clients request every scope the server lists, including write and admin rights, and the person approving rarely reads the list. Servers can prevent this by naming the minimum scope in their WWW-Authenticate challenge.
An attack in which an MCP server hides instructions for the model inside a tool's description, name or schema, so an agent that reads the tool list is steered without the user seeing it: to read files, send data elsewhere or call another tool. It works because agents treat tool descriptions as trusted context. A scan of the tool list can flag concealed instructions before any tool is called; it cannot prove a tool is safe, only that its published text carries no known pattern.
A record of the exact tool list an agent approved from an MCP server: a hash of every tool's name, description, input schema and annotations, plus one hash over the set. Before each session the agent compares the server's current tools with the lockfile and refuses or asks again when anything moved. It catches a changed tool; it does not judge whether the approved version was safe in the first place.
A server that behaves well while it is reviewed and approved, then changes a tool's description or behaviour afterwards, relying on clients never checking again. The defence is pinning: keep a hash of what was approved and compare on every connection, which is what a tool lockfile does. Pinning detects a change in the published tool text; a change made only on the server side, behind an unchanged description, is not visible to it.
Two MCP servers exposing tools with the same names, so an agent connected to both can call the wrong one, or a server copies a known publisher's tools to borrow its trust. Shadowing is detected by comparing tool lists across servers and across the servers one client connects to. A shared generic name such as search or fetch is not shadowing on its own.
A check an MCP client runs before it connects to a server and before each tool call: it asks for a signed decision about the server's domain and the action, and proceeds, warns, asks a person, or refuses. CrawlCheck publishes preflight middleware for the official TypeScript and Python MCP clients. A preflight decides whether to act on what is known now; it is not a security review of the server's code.
A machine-readable file a service publishes so AI agents can learn, in one request, how to use it: where to ask, how to verify the answer, how to cite it and under what terms. CrawlCheck publishes a signed one at /.well-known/crawlcheck.json. A manifest describes what the service offers; it does not prove the service behaves as described, which is why the answers themselves are signed.
A signed verdict about a site that the site serves itself, in a response header or a well-known file, so an agent can check it without calling the issuer. The idea comes from OCSP stapling in TLS. The agent still verifies the signature, the domain it is bound to and its expiry; a stapled verdict past its fresh-until time is ignored and the agent asks directly.
A compact signed verdict about a domain published as a DNS TXT record, so agents and resolvers can fetch it with an ordinary DNS query and cache it for its stated lifetime. CrawlCheck answers <domain>._v.crawlcheck.io over DNS-over-HTTPS. It carries decisions and freshness, not the full evidence, which stays in the full signed answer.
The web as it is used by software agents rather than people: programs that fetch pages, read machine files, call tools and complete tasks on a user's behalf, and that need to know what a site permits before they act. It differs from the crawled web in that an agent acts in the moment, so it needs an answer about a site now, not an index built last month. A site can prepare for it by publishing its rules and capabilities in machine-readable form; it cannot control what an agent does with them, which is why verification and conduct records exist.
A check an agent runs against a site before it acts there: can it read, cite, connect to or transact with this domain, under the site's own published rules and its measured state. A preflight answers from evidence already gathered, so it costs the site nothing and takes well under a second; the answer is a signed decision receipt the agent can carry. It can tell an agent whether the action is permitted and how fresh that judgement is; it cannot guarantee the site will behave the same way a day later, which is why every receipt carries an expiry.
A signed record of a preflight decision: which domain, which action, allow or refuse, the reasons, the evidence it rested on and when it expires. An agent keeps the receipt as proof that it checked before acting, and a site can verify it offline with a public key. It can show that a decision was made and what it said; it cannot prove the agent then obeyed it, which is what the agent conduct record covers.
The minimum level of evidence a decision demands before an action is allowed, scaled to what the action can do: reading a page needs less than citing it, citing less than connecting to a tool, connecting less than paying. A resolve answer states the assurance each action reached and the doubts that kept it from a higher one. It can make a refusal explainable; it cannot make a low-risk action safe when the site itself is unreliable.
A signed, freshness-bounded statement about a domain that an agent can act on: its crawler policy, machine files, capabilities, measured state, and a decision per action with the assurance it reached. It is computed from evidence already held, served in milliseconds, and carries the time after which it should be treated as stale. It can replace a crawl of the site for the purpose of deciding what to do there; it cannot describe pages that were never measured, and it says so rather than guessing.
The period during which a stored answer about a site is still considered current, stated as a fresh-until time on the answer itself. Past that time the answer is served as stale, with the staleness marked, until the site is measured again. It can tell a consumer how much to trust an answer right now; it cannot extend the life of an answer by itself, since only a new measurement does that.
Software that uses a model plus tools, state, and control logic to pursue a task across multiple steps. The label does not imply autonomy, reliability, or legal agency.
Measured: ROBOTS_LEGACY_AGENT_ONLY 0.2%
The repeated cycle of observing state, choosing an action, executing it, and evaluating the result. Without limits it can repeat errors, consume budgets, or act beyond the original intent.
The component that decomposes a goal into proposed steps or dependencies. A coherent plan is not evidence that the steps are feasible, authorized, or complete.
The component that performs selected tool calls or actions. Separating it from planning allows policy checks but does not guarantee they are correctly implemented.
Stored information carried across steps, sessions, or users to support continuity. It introduces retention, privacy, staleness, and incorrect-association risks.
Temporary task state kept within the active context or run. It is constrained by context and can disappear during compaction, retries, or handoffs.
State persisted beyond one interaction for later retrieval. Persistence does not make a memory accurate, relevant, current, or appropriate to expose.
Stored records of prior events, interactions, or runs. Recalling an episode can guide behavior while importing a past error into a new context.
Stored facts or generalized knowledge extracted from earlier material. It removes event detail and may lose when, where, or how a claim was established.
Coordinating models, tools, agents, approvals, and state through a workflow. More components expand capability and the number of failure boundaries.
A system in which several agents communicate or divide work. Agreement among agents is not independent corroboration when they share models, prompts, or sources.
Transferring control and task context from one agent to another. Missing constraints or provenance at the boundary can change the task silently.
Assigning another agent responsibility for a task or subtask. The delegating agent still needs to validate authority, inputs, outputs, and side effects.
Data returned after an external operation invoked by a model or workflow. It should be treated as untrusted input unless the source and schema are verified.
The machine-readable definition of a tool’s parameters and result shape. It constrains syntax but not business rules, permissions, or safe combinations of values.
Natural-language guidance telling a model when and why to use a tool. Small wording changes can alter selection, making descriptions part of application behavior.
The decision to answer directly or invoke one or more available functions. A correct tool can still be called with the wrong arguments or at the wrong time.
A policy determining which actions an agent may invoke in a context. Exposure in a schema is not the same as authorization to execute.
A workflow requiring human review, input, or approval at specified points. It reduces some risks only if the reviewer receives adequate context and can meaningfully refuse.
Human supervision of mostly autonomous execution with the ability to intervene. Oversight after rapid or irreversible actions may be nominal rather than effective.
A control that pauses execution before a consequential step. It should bind the approved parameters and target; approving a vague plan is not approving every resulting action.
An isolated environment restricting code, network, files, credentials, or system calls. Isolation limits impact but is only as strong as its escape boundaries and configuration.
The verifiable principal on whose behalf an agent connects or acts. A product name in a request is not authentication.
The permissions granted to an authenticated agent or delegated user. Identity answers who; authorization answers what that principal may do.
Finding machine-readable descriptions of available agents and capabilities. Discovery exposes claims that clients must still validate before delegation.
Client and server agreeing on supported versions, modalities, streaming, or interaction features. It prevents incompatible use but does not assess output quality.
A program exposing resources, prompts, tools, or related capabilities through Model Context Protocol. Availability through MCP does not make its data trusted or its actions safe.
The protocol component maintaining a connection between a host application and an MCP server. It transports capabilities and messages but need not make model decisions itself.
The user-facing AI application coordinating one or more MCP clients. Host policy determines which server capabilities are exposed to the model or user.
Context or data a server makes available for a user or model to read. A resource is not necessarily static, public, authoritative, or side-effect free.
A server-provided message template or workflow intended for explicit user selection. It supplies instructions, not an independently executable capability.
A callable function an MCP server allows a model to request. Tool metadata describes expected use; the server remains responsible for validation and access control.
The communication mechanism carrying MCP protocol messages between components. Transport security and application authorization are separate requirements.
A JSON-based remote procedure call format used by protocols including MCP. Well-formed messages do not authenticate sender, authorize methods, or encrypt transport.
A protocol for a client agent to discover and exchange task-related messages with a remote agent. It complements tool protocols but does not make autonomous delegation trustworthy.
The application or agent initiating requests to an A2A server for a user or another system. It must decide whether the remote agent is suitable and authorized.
A remote agent exposing an A2A-compatible interface for tasks and responses. Protocol compliance does not establish competence or honesty.
A capability unit advertised by an agent with identifiers, descriptions, examples, and supported modes. It is a declared ability, not an independently tested certification.
A stateful unit of delegated work exchanged through an agent protocol. Task completion status does not prove that the result is correct or accepted by the user.
A file, structured object, or other output produced during an agent task. It should retain provenance and content type rather than relying only on the surrounding conversation.
A proposal for a web page to expose tools to a browser-based agent through a small script bridge, so the agent can act on the page's functions rather than scraping its markup. It runs client-side, so it is invisible to a server-side scanner and to any agent that does not execute scripts.
Agent-to-Agent, a protocol for one agent to discover another's capabilities through an agent card and delegate tasks to it over HTTP. It standardises the handshake between agents, not what they may do; an agent card is a claim about capability that a caller still has to test.