---
title: "Register an agent for a sandbox key"
method: POST
path: "/api/v1/agents/register"
tags: ["Onboarding"]
---

# Register an agent for a sandbox key

`POST /api/v1/agents/register`

Self-serve agent onboarding: no account, no request body, no human step. Returns a sandbox API key (oxi_sk_test_...) and the path to live access. The key works only on the sandbox server (https://0xinsider.com/sandbox/api/v1), where it is optional: send it as Authorization: Bearer to exercise the credential path, and the sandbox answers a malformed key with the production 401. Nothing is stored, so the key cannot be listed or revoked and does not expire; register again for a new one. The live API answers a sandbox key with 401 invalid_api_key and error.reason sandbox_api_key. Live data needs an account with an active Pro subscription, and either an oxi_sk_live_ key from https://0xinsider.com/developers or an OAuth access token (https://0xinsider.com/auth.md). The request body is not read.

## Headers

- `X-Query-Validation` 'strict'

## Response `201`

A new sandbox key. Every call returns a different key, with Cache-Control: private, no-store.

- object
  - `object` 'agent_registration', required
  - `data` AgentRegistration, required — A sandbox key and the path to live access (#13959). Nothing is stored: the key cannot be listed or revoked and does not expire. Register again for a new one.
    - `api_key` string, required — The sandbox key. Send it as Authorization: Bearer <api_key> to the sandbox. The last 8 hex characters are a checksum (the first 4 bytes of SHA-256 over the rest of the key), so the sandbox can tell a mistyped key from a real one. It is not a secret and unlocks no production data.
    - `livemode` false, required — Always false: this key never reaches live data.
    - `environment` 'sandbox', required
    - `created_at` string, date-time, required
    - `sandbox` object, required
      - `api_base_url` string, uri, required — The sandbox V1 base URL. Every documented operation answers here with example data, except GET /api/v1/stream: a Server-Sent Events stream is a live connection rather than a body, so the sandbox answers it with 400.
      - `first_request_url` string, uri, required — A read to send with the key.
      - `openapi_url` string, uri, required — The OpenAPI document, served by the sandbox.
    - `live_access` object, required — What live data needs, and where each credential comes from. Both need a person: an account with an active Pro subscription.
      - `api_base_url` string, uri, required — The production V1 base URL.
      - `requirement` string, required — What a live credential needs, in one sentence.
      - `api_keys_url` string, uri, required — Where a signed-in Pro user creates a live key (oxi_sk_live_).
      - `oauth_authorization_server_metadata_url` string, uri, required — RFC 8414 authorization server metadata, for an OAuth 2.1 grant the user approves.
      - `oauth_registration_endpoint` string, uri, required — RFC 7591 dynamic client registration for public OAuth clients.
      - `pricing_url` string, uri, required — The Pro plan.
      - `auth_guide_url` string, uri, required — The authentication walkthrough for agents.
  - `meta` ResponseMeta, required
    - `request_id` string, required — Unique request ID (req_ prefix). The same value as the X-Request-Id response header, the request's usage accounting row and its log lines.
    - `cached` boolean, required
    - `cache_age_s` integer — Cache age in seconds. Omitted when the response was not cached, and also when it was cached but its age cannot be established (an entry stored before its cache carried a computed instant). Never a placeholder: an unknown age is reported as no value rather than as the cache TTL.
    - `cost` integer, required — Advisory request weight (relative compute cost). 1 for simple reads; higher for heavier endpoints. Not a credit/price.
    - `ranking_generation` integer — Committed PostgreSQL-owned leaderboard generation for the returned rows and cursor. Present on GET /api/v1/leaderboard; omitted on endpoints that do not read this ranking.
    - `ranking_as_of` string, date-time — Authoritative RFC3339 timestamp from cache_generations.updated_at for ranking_generation. It is read in the same repeatable-read snapshot as the leaderboard rows and is not request time, cache write time, or row insertion order.
    - `directional_source` 'live' | 'degraded' — Which path produced the team-directional read on this response. Only present on endpoints that compute one (today: GET /api/v1/sports-edge-signals). "live" means the read RAN. "degraded" means it FAILED, so nothing was measured and the ranking fell back to raw conviction. The flag describes the READ, not its consequence: a read that ran and found nothing groupable also leaves the directional fields null, and that is honestly "live" -- the per-signal nulls already say "nothing to enrich here", so this snapshot-level flag carries only what they cannot, namely whether the read ran at all. A degraded response is cached on the shorter degraded TTL so it self-heals. Reported SEPARATELY from ranking_source because the two degradations are independent -- a sharp-money DB miss weakens the ranking DATA, a directional failure removes a ranking WEIGHT -- and a consumer down-weighting a degraded response needs to know which input it lost. Omitted on endpoints that compute no directional read.
    - `ranking_source` 'live' | 'db_only' — Which ranking-data path produced this response. Only present on endpoints that can degrade a ranking (today: GET /api/v1/sports-edge-signals). "live" is the normal path (the current holder pile from the provider batch); "db_only" is the degraded fallback (a truthful but weaker trader_markets ranking) served when the live sharp-money ranking batch is unavailable (a sharp-money DB read failure, not a Polymarket outage) and cached on a shorter TTL, so a consumer can down-weight or skip it. Omitted on endpoints that never degrade.
    - `category_skill_source` 'live' | 'partial' | 'degraded' | 'unavailable' — Whole filtered snapshot category-evidence status before pagination. Operational live always remains partial source coverage.
    - `category_skill_model_version` string
    - `category_skill_taxonomy_version` string
    - `category_skill_platform` 'polymarket'
    - `category_skill_scope` 'observed_goldsky_primary_taker_fill'
    - `category_skill_source_coverage` 'partial_whale_threshold_fills' | 'graded_wallet_fills'
    - `category_skill_observation_started_at` string, date-time
    - `category_skill_model_operationally_degraded` boolean — Whole-model operational readiness captured with the category model snapshot. Present on category-enriched responses even when the filtered signal list is empty. When true, category_skill_source is degraded and sports-edge-signals uses the shorter degraded cache TTL.
    - `category_skill_status_counts` object
      - `live` integer, required
      - `insufficient` integer, required
      - `stale` integer, required
      - `unknown` integer, required
      - `degraded` integer, required
    - `category_skill_base_payload_hash` string — SHA-256 of the funded signal membership/order/rank/cursor vector immediately before category-skill enrichment. Sports-edge-signals only.
    - `category_skill_enriched_base_payload_hash` string — Independent SHA-256 recomputation over the same base fields immediately after category-skill enrichment. Equality with category_skill_base_payload_hash proves shadow enrichment did not change funded inputs. Sports-edge-signals only.

## Other responses

- `408` — The handler did not answer inside the server's 30-second timeout. error.code is request_timeout. On GET and HEAD the response carries Retry-After and error.retry_at; on a mutation it carries neither, because the request may have completed on the server: check its state before repeating it, and reuse its Idempotency-Key.
- `429` — Rate limit exceeded. Three independent budgets. (1) 100 requests/minute per user (sliding window), on every authenticated route. (2) On the BATCH routes only: 2500 batch item units/minute per user, reserved before any item is executed. A batch with N requested items costs N item units, including duplicate and invalid items. 2500 = 100 requests x 25 items per batch, which is the most item work a key can buy through the request limiter at all: a caller may spend their entire 100-request minute on full 25-item batches without the item budget being what stops them. The REQUEST budget is the effective ceiling, and batching is never the more expensive choice. The item budget can still deny at a sliding-window boundary (both counters carry the previous window forward with a floor, and the item counter runs 25x the request counter), so honor a 429 from either. Over-quota batches return 429 with Retry-After plus RateLimit-Limit, RateLimit-Remaining, and RateLimit-Reset before any item work is done. (3) The monthly quota: Pro includes 250,000 authenticated requests per UTC calendar month, whatever the billing cadence. Over 250,000: with pay as you go on, requests keep answering and the excess bills at USD 0.20 per 1,000 on a monthly invoice, up to 1,000,000 requests a month; without it, from October 1, 2026, the next request answers 429 rate_limited with error.reason monthly_quota_exceeded and a Retry-After to the month's reset, and from the same day a pay-as-you-go account answers the same past 1,000,000. A refused request is not counted. Every authenticated response carries X-Monthly-Quota-Limit, X-Monthly-Quota-Remaining, and X-Monthly-Quota-Reset (unix seconds, the first of next month). (4) The per-address budget: 1200 requests/minute per IP, shared by every caller behind one address and counted before authentication, on every route. A 429 from it carries error.reason ip_rate_limited and describes that bucket in RateLimit-*; a throttled address (sustained over-limit traffic) carries error.reason ip_throttled with a Retry-After of minutes to days, and a request before it does not shorten the cooldown. Every 429 is the standard error envelope with meta.request_id equal to X-Request-Id.
- `500` — Unexpected server error

## Changes

- **2026-09-23** `ece7a25b7a44` — 3 warning, 3 info
  - added the new `freshness_ceiling_unsatisfied` enum value to the `error/reason` response property for the response status `408`
  - added the new `freshness_ceiling_unsatisfied` enum value to the `error/reason` response property for the response status `429`
  - added the new `freshness_ceiling_unsatisfied` enum value to the `error/reason` response property for the response status `500`
  - added the optional property `error/freshness` to the response with the `408` status
  - …2 more
- **2026-09-23** `8462acf80f8c` — 3 warning
  - added the new `export_expired` enum value to the `error/reason` response property for the response status `408`
  - added the new `export_expired` enum value to the `error/reason` response property for the response status `429`
  - added the new `export_expired` enum value to the `error/reason` response property for the response status `500`
- **2026-09-22** `1c5e270b4124` — 3 warning, 1 info
  - added the new `unknown_query_parameter` enum value to the `error/reason` response property for the response status `408`
  - added the new `unknown_query_parameter` enum value to the `error/reason` response property for the response status `429`
  - added the new `unknown_query_parameter` enum value to the `error/reason` response property for the response status `500`
  - added the new optional `header` request parameter `X-Query-Validation`
- …earlier changes not shown

[Full history](https://skmtc.dev/0xinsider/apis/0xinsider-api/changes/api/v1/agents/register/post.md)

---

[API](https://skmtc.dev/0xinsider/apis/0xinsider-api.md) · [All operations](https://skmtc.dev/0xinsider/apis/0xinsider-api/llms.txt) · [OpenAPI document](https://skmtc.dev/0xinsider/apis/0xinsider-api/revisions/2035b3316329?raw)