Capabilities, agents, workflows, and mini apps

A service is a managed product area, such as research or memory. A capability is one action inside it, such as creating an analysis or recalling a memory.

The normal path is simple:

1. List the capabilities your key can use.
2. Pick a stable operation id.
3. Call it now, or create a durable run for longer work.

The registry covers agents, scheduled workflows, mini apps, research, memory, media, app hosting, communications, analytics, market intelligence, wallets, trading, compute, testnet assets, and managed databases. It also marks work that is not currently available.

Discover capabilities

const response = await hive.services.capabilities();
if (!response.ok) throw new Error(response.error);

for (const capability of response.capabilities) {
  console.log(
    capability.serviceId,
    capability.id,
    capability.operationId,
    capability.approval,
  );
}

Filter to one service with hive.services.capabilities("managed-agents") or GET /capabilities/{serviceId}. Results are filtered through the caller’s service and operation boundary.

Invoke by operation id

const result = await hive.services.invokeOperation(
  "hive-research",
  "analyses.create",
  { question: "Map the evidence for and against this decision." },
  { idempotencyKey: "analysis-decision-001" },
);

For a path template, pass values separately:

const result = await hive.services.invokeOperation(
  "managed-agents",
  "agents.chat",
  { message: "Prepare today's operating brief." },
  {
    pathParameters: { agentId: "agent_..." },
    idempotencyKey: "agent-brief-2026-08-25",
  },
);

Capabilities with requestFormat: "multipart" accept project files by id. Upload once, then attach the same managed file to direct invocations or durable runs:

await hive.services.invokeOperation(
  "distill",
  "runs.create",
  { skillName: "Customer interview analyst" },
  {
    fileIds: [uploaded.file.id],
    idempotencyKey: "distill-interviews-v1",
  },
);

The calling key needs files:read. A single invocation may attach up to 40 files with a combined 25 MB limit.

Capabilities with requestFormat: "binary" accept exactly one managed file by id. The published hive-compute.artifacts.upload operation uses this path for encrypted workload input. The SuperAgent API takes the ciphertext digest from the managed file record and validates the encryption metadata before forwarding the bytes. It never accepts a caller-supplied replacement digest.

Create a protected action approval

Capabilities marked approval: "always" require a matching Platform approval. The approval binds the service, operation, path parameters, query, input, managed files, and connection id, so approving one action cannot authorize changed input or a different credential.

const requested = await hive.approvals.createServiceAction(
  {
    serviceId: "app-hosting",
    operationId: "sites.publish",
    input: { siteId: "site_...", versionId: "version_..." },
  },
  { idempotencyKey: "approval-publish-version-17" },
);

if (!requested.ok) throw new Error(requested.error);

await reviewer.approvals.decide(
  requested.approval.id,
  "approve",
  { idempotencyKey: "decision-publish-version-17" },
);

const published = await hive.services.invokeOperation(
  "app-hosting",
  "sites.publish",
  { siteId: "site_...", versionId: "version_..." },
  {
    approvalId: requested.approval.id,
    idempotencyKey: "publish-version-17",
  },
);

An approval expires, can be rejected, and is consumed after one successful matching action. Capabilities marked policy rely on the owning service’s account policy and may still return an approval or confirmation requirement.

Capability families

The registry exposes these backend families:

• Persistent cloud agents, agent chat, app building, integrations, and governed computer actions through managed-agents.
• Scheduled, manual, and event-driven routines through managed-workflows.
• Research, multi-agent swarms, audience discovery, transcripts, skill distillation, market intelligence, and lead data.
• Managed model inference, media generation, photo analysis, GPU rentals, and Hive Compute jobs.
• Typed and evolving cloud memory, lexical and semantic recall, grounded answers, generations, capsules, and knowledge graphs through cloud-superbrain; portable application data through hivemind-database.
• App hosting, bookings, social publishing, email, analytics, X operations, calendar access, and remote tool servers.
• Token research, wallet risk, bridge quotes, copy-trading monitors, managed wallets, and managed spot trading.

hive-compute and testnet-faucet are available through registered operations and settle against the same authenticated HivemindOS credit account as the rest of the managed suite. Failed work releases its reservation; completed work reports the exact charge in the invocation result, response header, usage records, and audit records.

Available managed mini-app backends use the same account credit balance and service-owned policies. The API exposes their backend logic without requiring a desktop session or copying their protected credentials into your product.

Next: use managed memory and knowledge, use Hive Compute and the testnet faucet, browse the service catalog, or open the exact endpoint map.