{"protocolVersion":"0.3.0","name":"HALOWERK agentenwerk","description":"HALOWERK agentenwerk. Bezahlung über x402 in USDC auf Base Mainnet.","url":"https://agent.halowerk.com","version":"1.0.0","preferredTransport":"JSONRPC","capabilities":{"streaming":false,"pushNotifications":false,"stateTransitionHistory":false},"defaultInputModes":["application/json"],"defaultOutputModes":["application/json"],"skills":[{"id":"webhook_inbox","name":"Create a short-lived public receive URL, then poll it for the events that arrived, parsed and in order.","description":"Gives an agent without a reachable server a place to receive callbacks. create returns a public URL that accepts any POST, GET or PUT from a third party; poll returns the events that arrived since a cursor you carry forward, with headers, parsed body and arrival time. The inbox expires after at most seven days and holds at most 500 events, oldest dropped first. Reading requires the secret returned at creation. Do not point payment or authentication callbacks here: anyone who learns the URL can post to it, and stored bodies are readable with the secret alone.","tags":["agentenwerk"],"examples":[],"inputModes":["application/json"],"outputModes":["application/json"]},{"id":"state_store","name":"Store, read, list and delete small pieces of agent state under a namespace, with a mandatory expiry so nothing is kept forever.","description":"Key-value state for agents that need to survive between calls. Every write carries a time-to-live; the maximum is 30 days and there is no way to store something permanently. Values are opaque JSON up to 64 KB, at most 1000 live keys per namespace. Actions: set, get, list, delete. A get on a missing or expired key answers found=false, it is not an error. The namespace is a plain string chosen by the caller and is not authenticated — anyone who knows it can read it, so do not put secrets in it. Expired rows are removed on access, not on a timer, so a list can lag a few seconds behind the exact expiry moment.","tags":["agentenwerk"],"examples":[],"inputModes":["application/json"],"outputModes":["application/json"]},{"id":"wakeup_schedule","name":"Schedule one HTTP callback to your own URL at a fixed time, then check or cancel it — for agents that are not running when the moment arrives.","description":"Books a single POST to a URL you own, at a time you name, with a payload you supply. Delivery is retried up to five times with growing gaps and every attempt is recorded with its status. Lead time is between 10 seconds and 30 days, and at most 200 pending wakeups per target URL. The target must be a public https address: private, loopback and link-local addresses are refused, so this cannot be pointed at an internal network. Delivery is at-least-once and the time is a floor, not a guarantee — the dispatcher checks every 15 seconds, so expect a few seconds of lag. Actions: schedule, status, cancel, list.","tags":["agentenwerk"],"examples":[],"inputModes":["application/json"],"outputModes":["application/json"]},{"id":"job_runner","name":"Hand over a list of up to 50 HTTP calls, get an immediate job id, and have the collected results posted to your callback when they are done.","description":"For work that outlives one HTTP response. You submit steps, each an HTTP request against a public address; the job is acknowledged at once and runs in the background, and the collected results are POSTed to your callback URL when finished. Every step reports its status, timing, response headers and body up to 256 KB, and a failing step never aborts the others. Price is 0.003 USDC per step, quoted before payment. Only http and https against public hosts are performed — private and loopback addresses are refused and no supplied code is ever executed. Steps run one after another with a 20 second cap each, so a 50-step job can take several minutes. Actions: submit, status, cancel.","tags":["agentenwerk"],"examples":[],"inputModes":["application/json"],"outputModes":["application/json"]},{"id":"file_vault","name":"Park a file for a limited time and get back an unguessable download URL that anyone you send it to can fetch without a wallet.","description":"Takes a base64 file up to 25 MB and stores it for as long as you say, at most seven days. You get a download URL carrying a 32-byte token; whoever holds the link can fetch the file once or up to a limit you set, with no payment and no account. Price is 0.004 USDC per started megabyte, quoted before payment from the size you declare. Actions: store, status, delete. Anything expired is deleted with its contents, and a link is only as private as you keep it, so this is not a place for personal or secret material. Content is stored and served unchanged and is never inspected or scanned.","tags":["agentenwerk"],"examples":[],"inputModes":["application/json"],"outputModes":["application/json"]},{"id":"a2a_card_validate","name":"Check an A2A Agent Card for required fields, skill definitions, authentication declarations and transport hygiene, and get every deviation listed with its JSON path.","description":"Validates an Agent Card document against the A2A convention and returns findings by severity: error (the card is unusable as published), warning (works, but a consuming agent will guess), info (recommended field absent). Checked: the required fields name, description, url, version, capabilities and skills; that url is absolute and https; that every skill carries id, name, description and tags and that skill ids are unique; that declared securitySchemes use a known type and that every entry in security refers to a defined scheme; that input and output modes are declared; that streaming and pushNotifications are booleans. This endpoint does not fetch the card and never contacts the agent: it judges only the document handed to it, so a passing card is not proof that the endpoint is reachable or honest.","tags":["agentenwerk"],"examples":[],"inputModes":["application/json"],"outputModes":["application/json"]},{"id":"mcp_inspect","name":"Look at a remote MCP server from the outside: whether it answers, which protocol it speaks, and the tools, resources and prompts it offers with their schemas.","description":"Performs the MCP handshake against a streamable-HTTP endpoint and then reads the catalogue: tools with their input schemas, resources with their URIs and mime types, and prompts with their arguments. Servers answering as an event stream are handled as well as plain JSON, and a session id is carried across calls. Only listing methods are sent — no tool of the inspected server is ever executed, nothing is written, and no resource contents are read. The target must be a public address; private and loopback addresses are refused. An unreachable or non-conforming server yields a reachable=false report with the reason, not an error.","tags":["agentenwerk"],"examples":[],"inputModes":["application/json"],"outputModes":["application/json"]},{"id":"mcp_normalize","name":"Bring MCP tool definitions from different servers into one shape: uniform input schemas, a field list, a synthetic example call and a common set of error codes.","description":"Takes a list of MCP tool definitions, as returned by tools/list on any number of servers, and rewrites them into one common form. Servers disagree on where the schema lives — inputSchema, input_schema, parameters, schema or arguments are all in use — and a consumer reading only one of them loses the others silently; every spelling is accepted and the one found is reported. For each tool the result carries a flat field list with type, required flag and description, a synthetic example call built from the schema, and the tool name normalised to snake_case. Name collisions across servers are detected and reported with the servers involved. The example call shows the shape only: values like \"string\" and 0 are placeholders and must never be sent to a live server. No network access happens here — only the definitions handed in are read.","tags":["agentenwerk"],"examples":[],"inputModes":["application/json"],"outputModes":["application/json"]},{"id":"fallback_resolve","name":"Find paid x402 endpoints that can stand in for one that stopped answering, ranked by how well they match the task and what they cost.","description":"Searches the public x402 discovery catalogue for services able to replace a broken dependency. Give the failed URL and, better still, a short description of what it did and the fields you need back; matches are scored on overlap between your task words and the candidate name, description and path, then ordered by score and price. Every hit carries the reason it matched, its price in USDC, its network and when the catalogue last saw it. The endpoint that failed is excluded from its own replacement list. This reads a public index and never calls a candidate: a listed service is not proof that it works today, and the input schema of a replacement will differ, so treat the result as a shortlist to test, not a drop-in.","tags":["agentenwerk"],"examples":[],"inputModes":["application/json"],"outputModes":["application/json"]},{"id":"task_quote_compare","name":"Compare competing agent offers for one task on price, delivery time, required capabilities and output format, and get a ranked list with the reason for every position.","description":"Normalises offers that describe the same task in different shapes and ranks them. Each offer may carry a price, a delivery time, a list of capabilities and one or more delivery formats. You state which capabilities and formats the task requires; offers missing a required capability or format are marked ineligible and ranked last, never silently dropped. The ranking is a weighted score over price, speed and capability coverage, with weights you supply and which are echoed back. Every ranked entry carries the sub-scores it was built from, so the order can be recomputed by hand. Prices must already be in one currency: no conversion happens here, and a mixed-currency input is rejected rather than compared wrongly.","tags":["agentenwerk"],"examples":[],"inputModes":["application/json"],"outputModes":["application/json"]},{"id":"result_consensus","name":"Compare several agent answers to the same question, report where they agree, and list every field where they contradict each other with the competing values.","description":"Takes two to twelve answers and returns a field-by-field comparison. For JSON answers every leaf path is compared across responders: a path is agreed when all responders that supplied it carry the same value, conflicting when they differ, and partial when some omitted it. Numbers count as equal inside a relative tolerance you set. For plain-text answers the comparison is word-overlap similarity, which measures shared vocabulary and not meaning — two correct answers in different words will score low, and that is reported as low agreement, not as a contradiction. The result names a majority value per conflicting path together with how many responders backed it. This endpoint does not decide who is right: it has no ground truth and no model, it only shows where the answers part company.","tags":["agentenwerk"],"examples":[],"inputModes":["application/json"],"outputModes":["application/json"]},{"id":"twin_diff","name":"Compare a desired state document against an observed one field by field and get every deviation with its absolute and relative size, the tolerance that was applied and the declared unit.","description":"Takes two state documents — target (what the twin should look like) and observed (what was measured) — and returns a leaf-by-leaf comparison. Every leaf path carries one of five states: in_tolerance, out_of_tolerance, missing_in_observed, unexpected_in_observed or type_mismatch. Numeric leaves are judged against a tolerance you set per path, either absolute (tolerance_abs) or relative (tolerance_ratio); paths without an entry fall back to default_tolerance_ratio, which is 0, meaning exact equality. Strings, booleans and null are compared for equality only, with optional case folding for strings. Arrays are addressed by index, so a reordered list reads as a set of deviations and not as a match. Units are never inferred from field names: a unit appears in the result only where you declared one, otherwise unit is null and unit_source says not_declared. This endpoint is stateless — it stores no twin, tracks no history and performs no synchronisation; it reports the gap between the two documents handed to it at this moment and nothing else.","tags":["agentenwerk"],"examples":[],"inputModes":["application/json"],"outputModes":["application/json"]},{"id":"memory_compact","name":"Compact a long agent conversation to a character budget in four declared stages and get back exactly which messages stayed verbatim, which were trimmed and which were dropped, with the loss quantified.","description":"Takes a message history and a character budget and returns a shortened history together with a full account of what it cost. Four stages run in fixed order: pin protects messages that must survive (roles listed in keep_roles, the last keep_last_n messages, and anything you marked pin), dedupe removes later near-duplicates of an earlier message of the same role using word-overlap similarity, trim shortens unprotected survivors to max_message_chars at the last sentence boundary, and drop removes the lowest-scoring unprotected messages until the budget is met. The score is disclosed in the response and is a fixed weighting of recency, role and overlap with the vocabulary of the whole history. This endpoint uses no language model and does not summarise, paraphrase or rewrite anything: every character in the output appears verbatim in the input, and nothing is appended, not even an ellipsis. Consequences follow from that and are reported rather than hidden — word overlap measures shared vocabulary and not meaning, so two differently worded messages about the same fact are not recognised as duplicates, and trimming cuts at a sentence boundary and can therefore remove the point of a long message. Protected messages are never dropped, so a budget smaller than the protected content cannot be met; the response then says budget_met false and reports the shortfall instead of discarding something it promised to keep.","tags":["agentenwerk"],"examples":[],"inputModes":["application/json"],"outputModes":["application/json"]},{"id":"state_machine_step","name":"Apply one event to a state machine definition and get back the next state, the actions it triggers, whether a deadline was breached and which events the machine will accept next.","description":"Takes a machine definition, the current state, one event and an optional context, and computes a single step. Transitions may carry a guard, which is a declared comparison — field, operator, value — and never an expression: the field is a dotted path into context or payload, so a machine handed in by a stranger cannot execute anything on this server. Supported operators are eq, ne, gt, gte, lt, lte, in, not_in, exists and not_exists; a guard whose field does not resolve evaluates to false and says so rather than passing silently. Deadlines are computed from elapsed_in_state_s, which you supply, so the result does not depend on this server clock and the same request always yields the same answer. When a state carries timeout_s and the elapsed time exceeds it, timeout_precedence decides whether the timeout transition or a matching event transition wins; the default is timeout. A breached deadline is reported even when the state defines no timeout_target, in which case the machine does not move on account of it. The definition is checked for referential integrity before the step runs: transitions pointing at states that do not exist, and an initial state that does not exist, are rejected as a bad machine rather than silently followed. This endpoint is stateless: it stores no machine, no state and no history, and it applies exactly one step per call — use /v1/state-store to keep the state between calls.","tags":["agentenwerk"],"examples":[],"inputModes":["application/json"],"outputModes":["application/json"]},{"id":"intent_clarify","name":"Given several candidate intents and what tells them apart, get the single follow-up question that narrows the set fastest, with the expected and worst-case number of candidates left after each possible answer.","description":"Takes a list of candidate intents, each described by attributes, plus whatever is already known, and returns the one attribute worth asking about next. Candidates contradicted by the known values are ruled out first and listed with the attribute that excluded them; a candidate that simply does not declare a known attribute is kept, because silence is not a contradiction. For every remaining attribute the endpoint builds the branches an answer would create and ranks them by the prior-weighted expected number of candidates left, then by the worst case, then by the number of options; a candidate that does not declare the attribute survives every branch, which is counted rather than ignored. Attributes on which all remaining candidates agree carry no information and are returned separately as unusable_attributes, so they are not asked by mistake. When an attribute has more distinct values than max_options, the largest options are shown and the omitted values are named explicitly — nothing is capped silently. The answer also carries a greedy worst-case estimate of how many further questions would be needed. This endpoint uses no language model and does not read natural language: it never invents a candidate, an attribute or a value, and the question wording is a fixed template over the attribute name unless you supply your own phrasing in attribute_prompts. It is stateless — it remembers no dialogue, so pass the accumulated answers back in known on the next call.","tags":["agentenwerk"],"examples":[],"inputModes":["application/json"],"outputModes":["application/json"]},{"id":"ephemeral_rag","name":"Create, update, query, inspect or delete an isolated micro vector store whose documents and embeddings are automatically erased after 24 hours.","description":"Stores at most 20 short documents in an isolated SQLite row, embeds them with the local nomic-embed-text model, and performs cosine-similarity retrieval. Each store has a random identifier and a fixed 24-hour lifetime that is never extended by reads or writes. An in-process sweeper deletes expired rows every minute and each request also purges expiry. Input text is not written to request logs. This is retrieval infrastructure, not a durable database or an answer-generating model.","tags":["agentenwerk"],"examples":[],"inputModes":["application/json"],"outputModes":["application/json"]}],"payment":{"protocol":"x402","network":"eip155:8453","asset":"USDC","recipient":"0x2880EdfFF13100677Bf97A3CBdF3Bc34771C4E5E","manifest":"https://agent.halowerk.com/.well-known/x402"}}