HALOWERK agentenwerk

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

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

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

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

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

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

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

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

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, t

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 mar

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 insi

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 j

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 remove

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 b

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 d

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 r

Takes an OpenTofu state snapshot and a passphrase, encrypts the snapshot through the pinned OpenTofu state-encryption code, then proves the result three ways: the authorized key recovers the state with an identical semantic hash, a wrong key is rejected, and a tampered payload is rejected. What come

Relocates a Java package across a bounded set of source files using OpenRewrite, which works on a type-attributed syntax tree rather than on text. That is the difference that matters: a textual search and replace hits strings, comments and same-named symbols from other packages, while this changes o

Clusters a key column with native OpenRefine fingerprinting and returns a consolidation plan tied to actual row ids. The plan distinguishes the two cases that matter: rows that repeat an identical value and can simply be merged, and rows that collide on the key but disagree in other fields — those a

Executes a caller-directed sequence of table operations in native OpenRefine and returns the transformed table together with the exported recipe. What distinguishes it from simply transforming data is the evidence: the recipe is persisted, the project is reopened, and the whole thing is replayed fro

Reconciles a column of names against a reference catalog that you supply — your catalog, not a public knowledge base — using native OpenRefine reconciliation with a caller-set edit distance and an optional required type. The result separates rows with a single candidate, rows with several, and rows

Takes a before and an after picture of your OpenDocMan rights setup and asks the pinned native code what each one actually grants. Every user-document pair is evaluated in both states, and what comes back is which decisions changed, which of your declared expectations were violated in each state, an

Runs your simple-product checkout cases against the pinned OpenCart source in a private, non-root database that exists only for the length of the request. For each case it reports the totals the native code actually produces, which expectations were missed, and where a cart path diverged, rather tha

Walks a staged product-catalog import through the pinned OpenCart source in a private, non-root database and reports what the native code makes of it at each stage: how many masters and variants survive, whether option values stay attached to the right products, and how the resulting catalog behaves

Rebuilds the orders you supply inside a private, non-root OpenCart database and checks your export against what the native code computes. Each export row is matched to a ledger order, and every difference is named: a total that does not add up under the stated arithmetic policy, a currency conversio

Replays a grading scale through the native OpenOLAT implementation and compares, for each assessment, the grade that was recorded against the grade the scale actually yields. Every discrepancy comes back with the score, the band it falls into, the expected value and the recorded one, so a wrong entr

Executes a question-bank assessment against answer fixtures you supply — an answer plus the score it should produce — using the native OpenOLAT QTI engine rather than a reimplementation of the scoring rules. Each fixture comes back with expected and actual score, and the discrepancy count says at a

Runs a course structure and a set of learners through the pinned OpenOlat learning-path implementation and reports, per learner, which nodes the original code actually unlocks and in what order. Your declared expectations are checked against that run, and every discrepancy is named rather than summa