@coval/sdk

Typed TypeScript client for the Coval evaluation platform API.

The bulk of this SDK is generated from Coval's public OpenAPI specs (the same ones that power docs.coval.dev/api-reference). A small hand-written layer adds auth, retries with backoff, pagination iteration, and typed errors so the generated client is ergonomic for production use.

Install

npm install @coval/sdk

Requires Node 18+.

Quick start

import { CovalClient } from '@coval/sdk';

const coval = new CovalClient({ apiKey: process.env.COVAL_API_KEY! });

// List agents
const { agents } = await coval.agents.listAgents({ pageSize: 25 });

// Submit a production conversation for evaluation
await coval.conversations.submitConversation({
  covalConversationsAPISubmitConversationRequest: {
    agent_id: 'my-agent-id',
    external_conversation_id: 'call-2026-05-11-abc',
    occurred_at: new Date(),
    transcript: [
      { role: 'agent', content: 'Hi, thanks for calling.' },
      { role: 'user', content: "I'd like to update my account." },
    ],
  },
});

What the wrapper adds

Auth — x-api-key (lowercase)

Coval's gateway is case-sensitive about the API-key header. Uppercase X-API-Key returns Missing API Key. The CovalClient always sends lowercase x-api-key, so you don't have to think about it.

Retries with exponential backoff

For idempotent GET, HEAD, and OPTIONS requests, the client retries automatically on 408, 429, and 5xx responses, plus network errors. Defaults: 3 attempts, 200ms base delay with jitter, 5s cap. Retry-After headers (both seconds and HTTP-date) are respected. Mutating requests are not retried unless you explicitly add their method to retryableMethods.

const coval = new CovalClient({
  apiKey: process.env.COVAL_API_KEY!,
  retry: { maxAttempts: 5, baseDelayMs: 500, maxDelayMs: 10_000 },
});

// Disable retries entirely:
const coval2 = new CovalClient({ apiKey, retry: false });

Typed errors

Every non-2xx HTTP response becomes a CovalApiError after retries are exhausted. Network failures become CovalNetworkError. Both extend Error, so plain try/catch works.

import { CovalApiError } from '@coval/sdk';

try {
  await coval.agents.getAgent({ agentId: 'unknown' });
} catch (err) {
  if (err instanceof CovalApiError) {
    console.error(err.status, err.code, err.message, err.requestId);
  }
}

Pagination iterator

List endpoints return a single page plus a next_page_token. The paginate helper turns that into an async iterator so you don't have to manage tokens.

import { paginate, collectAll } from '@coval/sdk';

for await (const agent of paginate({
  fetchPage: (pageToken) => coval.agents.listAgents({ pageToken, pageSize: 50 }),
  items: (page) => page.agents,
  nextToken: (page) => page.next_page_token,
})) {
  console.log(agent.id, agent.display_name);
}

// Or collect everything into an array:
const all = await collectAll({
  fetchPage: (pageToken) => coval.simulations.listSimulations({ pageToken }),
  items: (p) => p.simulations,
  nextToken: (p) => p.next_page_token,
});

Diagnosing intermittent failures

A request that dies in transit never reaches the API, so there is no server-side trace of it. The client keeps live counters:

coval.stats;
// { calls: 12, requests: 14, retries: 2, networkErrors: 1 }

calls is logical API calls, requests counts every fetch including retries, and networkErrors counts failures where no response came back at all. Include these when reporting sporadic timeouts. For a line per retry and per in-transit failure:

const coval = new CovalClient({ apiKey, onDebug: console.debug });

Connection reuse

On Node the client uses the global fetch (undici), which expires idle keep-alive connections after roughly 4 seconds. That bound matters: some network paths silently drop an idle connection without closing it, and reusing one stalls the request until your timeout, because it never reaches the server.

Requests made within that window can still reuse a pooled connection. If you submit infrequently, or you have seen unexplained timeouts, pass an explicit dispatcher:

import { Agent, fetch as undiciFetch } from 'undici';

const coval = new CovalClient({
  apiKey,
  fetch: ((input, init) =>
    undiciFetch(input as never, { ...init, dispatcher: new Agent({ keepAliveTimeout: 1_000 }) })) as typeof fetch,
});

The Python SDK exposes the same bound directly as max_idle_seconds.

Customization

Override base URL (staging, self-host)

const coval = new CovalClient({
  apiKey,
  baseUrl: 'https://staging.api.coval.dev/v1',
});

Plug in custom middleware

const coval = new CovalClient({
  apiKey,
  middleware: [
    {
      async pre({ url, init }) {
        console.log('→', init.method, url);
        return undefined;
      },
    },
  ],
});

Override fetch (e.g., undici with keep-alive)

import { fetch as undiciFetch } from 'undici';

const coval = new CovalClient({ apiKey, fetch: undiciFetch });

Examples

Runnable in this repo. Set COVAL_API_KEY (and COVAL_AGENT_ID for the submit example) and run:

npm run example:list-agents
npm run example:submit-conversation

Versioning

The generated surface tracks the underlying Coval API. Spec changes that break the generated client increment the minor version. Hand-written wrapper changes that break the public TypeScript API also increment the minor version. Bug fixes increment the patch version.

Regenerating

If you want to regenerate the client locally:

# From the repo root, with the docs repo checked out as a sibling:
node scripts/bundle-spec.mjs
bash scripts/generate-sdks.sh

See Contributing for local generation and the release guide for scheduled regeneration setup.

License

MIT.