Everest SDK

Compress context before it reaches your AI agent.

Compression will remove unnecessary context and leave the important details.

For example, a coding agent makes calls to read files, execute code, and run tests. The file output, code results, and test logs might have a lot of noise. Pass the code, logs and other output to compress which will return only the content for the agent to focus on.

Install

npm install @everestagi/sdk

Requires Node.js 18 or later. ESM and CommonJS are supported, and the package has no runtime dependencies.

Usage

// Keep EVEREST_API_KEY on the server, out of client code and tool environments.
import { EverestClient } from '@everestagi/sdk'

const everest = new EverestClient({
  apiKey: process.env.EVEREST_API_KEY!,
  downstreamInputUsdPerMillionTokens: 4,
})

export async function compressOutput(output: string, goal: string) {
  const result = await everest.compress({ output, goal })

  console.log({
    applied: result.applied,
    reason: result.reason,
    tokensSavedEstimate: result.usage.tokensSavedEstimate,
    reductionRatio: result.usage.reductionRatio,
    inputCostSavedEstimateUsd: result.cost?.inputCostSavedEstimateUsd,
    requestId: result.request_id,
  })

  return result.text
}

output is the completed tool output; goal is the agent’s overall task; focus is optional, output-specific guidance about what to preserve; and command is optional context only and is never executed. Skips, timeouts and API failures keep the original text. The SDK never runs or reruns tools.

You can optionally pass focus, command and signal to compress, or set timeoutMs when creating the client. result.usage contains local size, token and latency estimates. If you provide downstreamInputUsdPerMillionTokens, result.cost also estimates downstream input cost savings. These are estimates, not billing records.

Savings estimates

Pass a local sessionId to compression calls, then use getSavingsTotals(sessionId) for that session or getSavingsTotals() for the client-wide total. To include expected cached reuse, configure downstreamCachedInputUsdPerMillionTokens and pass estimatedCacheReadCount; the SDK reports uncached, cached and combined estimated savings. Session IDs and cache-read estimates are not sent to Everest.

request_id, timestamp, attempts and other bounded diagnostics may be present when the API supplies them. Their absence does not make a valid compression result fail.

Error-handling behavior

When result.applied is false, result.text is the exact original output. Stable SDK failure reasons include sdk_authentication_failed, sdk_payload_too_large, sdk_timeout, sdk_cancelled, sdk_network_error, sdk_api_error and sdk_invalid_response. API skip reasons include below_threshold and raw_requested.

Keep EVEREST_API_KEY only on your trusted server. Do not put it in browser code, mobile apps, distributed CLIs, agent tool environments, source control or logs.

Using Everest directly with Codex instead of integrating a product? See the CLI guide.