HTTP API

The Dropstone HTTP API gives you programmatic access to the same three models the CLI uses — Dropstone Fast, Pro, and Heavy — over an OpenAI-compatible interface. One key, one bill, three model families.

Use the API when you want to call Dropstone from your own code: CI pipelines, internal tools, automation, or third-party apps. For interactive coding, use the CLI instead.

Base URL

https://api.dropstone.io/api/v1

All endpoints are mounted under /api/v1. The path is versioned so future breaking changes will land under /api/v2 without disrupting your code.

Authentication

Every request must include an API key in the Authorization header:

Authorization: Bearer dsk_live_<your-key>

Generate a key

1. Sign in to dropstone.io/dashboard
2. Open Settings → API
3. Click Create key, give it a name (e.g. Production CI)
4. Copy the full key — you will only see it once

Keys look like dsk_live_<43 chars>. Store them in a secrets manager or in the DROPSTONE_API_KEY environment variable.

Revoking a key

Revoke from the same Settings → API page. Revocation is immediate; in-flight requests with the key continue, new requests get 401.

Credits & billing

The API is pay-per-use against a credit balance. There is no free tier and no subscription on the API surface.

• Buy credits at dropstone.io/dashboard/billing. Stripe handles checkout.
• Each request deducts its cost from your creditBalance.
• When creditBalance drops to $0, the API returns 402 Insufficient credits until you top up.
• Subscription credits (Pro/Teams monthly allowance) and free-request quotas do not apply to API key requests.

Pricing

Pricing is real OpenRouter cost passed through with a 30% markup (1.3x). The full per-request cost is returned in the response usage.cost field, so you can verify every charge.

Cached prompt tokens are billed at the provider's cached rate (typically ~5–10% of the normal input rate), so multi-turn conversations get progressively cheaper.

Models

GET /api/v1/models

List the three available tiers.

curl https://api.dropstone.io/api/v1/models \
  -H "Authorization: Bearer $DROPSTONE_API_KEY"

Response:

{
  "object": "list",
  "data": [
    { "id": "dropstone-fast",  "object": "model", "display_name": "Dropstone Fast",  "owned_by": "dropstone" },
    { "id": "dropstone-pro",   "object": "model", "display_name": "Dropstone Pro",   "owned_by": "dropstone" },
    { "id": "dropstone-heavy", "object": "model", "display_name": "Dropstone Heavy", "owned_by": "dropstone" }
  ]
}

Chat completions

POST /api/v1/chat/completions

OpenAI-compatible chat completions. If you've used the OpenAI, Anthropic, or DeepSeek APIs, this looks identical.

Request body

Example: simple chat

curl https://api.dropstone.io/api/v1/chat/completions \
  -H "Authorization: Bearer $DROPSTONE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "dropstone-fast",
    "messages": [
      {"role": "user", "content": "Write a haiku about debugging."}
    ]
  }'

Response

{
  "id": "gen-1779530142-EfBhlhO1U2frV6tvMgKV",
  "object": "chat.completion",
  "created": 1779530142,
  "model": "dropstone-fast",
  "choices": [
    {
      "index": 0,
      "message": { "role": "assistant", "content": "Stack trace at midnight..." },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 11,
    "completion_tokens": 23,
    "total_tokens": 34,
    "cost": 0.0000098
  }
}

The usage.cost field is the billed amount in USD — what was deducted from your credit balance for this request (real provider cost × 1.3 markup).

Streaming

Set "stream": true to get a Server-Sent Events stream of token chunks. The stream ends with a data: [DONE] line and a final chunk containing the full usage block.

curl https://api.dropstone.io/api/v1/chat/completions \
  -H "Authorization: Bearer $DROPSTONE_API_KEY" \
  -H "Content-Type: application/json" \
  -N \
  -d '{
    "model": "dropstone-fast",
    "stream": true,
    "messages": [{"role": "user", "content": "Count to 5."}]
  }'

OpenAI SDK compatibility

Because the surface is OpenAI-compatible, you can use the official OpenAI SDK by overriding base_url:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.dropstone.io/api/v1",
    api_key=os.environ["DROPSTONE_API_KEY"],
)

resp = client.chat.completions.create(
    model="dropstone-fast",
    messages=[{"role": "user", "content": "Hello"}],
)
print(resp.choices[0].message.content)

Error codes

402 response shape

{
  "error": "Insufficient credits",
  "balance": 0,
  "message": "Your credit balance is empty. Top up at https://dropstone.io/dashboard/billing to continue.",
  "topUpUrl": "https://dropstone.io/dashboard/billing"
}

Rate limits

There are no hard rate limits enforced on the API today. Per-key usage tiers and daily spend caps are planned — when they ship, your existing keys will be auto-tiered based on lifetime spend, similar to OpenAI's tier system.

For now, set per-key budgets yourself by tracking the usage.cost field in your application.

Recommended practices

• Use environment variables, never inline keys: DROPSTONE_API_KEY=dsk_live_....
• One key per service, not one key shared everywhere. Easier to revoke when a service is compromised.
• Watch usage.cost in responses to track spend in real time.
• Handle 402 gracefully — your app should detect it and surface a top-up CTA rather than retrying.
• Cache responses for repeated identical requests on your side — we cache at the model level, but you save the full markup by short-circuiting before reaching us.

Differences from the CLI and SDK

If you want headless programmatic access in CI or a server, the HTTP API is the right surface. The SDK is for embedding the interactive agent in a Node process where a human is still in the loop.

Coming soon

These are on the roadmap and will land under the same /api/v1 namespace:

• POST /api/v1/agent/run — agent loop endpoint. Send a task, get a finished diff. Server-side multi-turn loop with built-in tools (file editing, web search, code execution). Flat per-task pricing.
• POST /api/v1/memory/store + GET /api/v1/memory/query — stateful memory via Qdrant. Agent context that persists across calls.
• MCP tool injection — include your own MCP server URLs in agent requests, agent calls them as native tools.
• Per-key spend caps — set a daily $ limit per key from the dashboard. CI safety net.

Star the GitHub repo or watch the changelog to know when they ship.