EnterpriseContact

Technology

Everything the task needs. Nothing it doesn't.

SOMA sits between your agents and the model. It removes redundant context while preserving the information needed to complete the task.

How it works

The noise goes. The work stays.

  1. Redundant tool outputs from earlier turns are compressed or removed when no longer needed.
  2. File reads that a later read covers are dropped.
  3. Output too long to be useful is trimmed.
  4. The task and the latest turns are kept as they are.
Request to the model47.2k tokens
  • System prompt3.0k
  • The task: fix the failing refund test0.8k
  • Read payments/refund.ts4.6k
  • Run the test suite6.2k
  • Search the repo for refund handling2.1k
  • Read payments/refund.ts, lines 1-1203.1k
  • Run the test suite, same output6.2k
  • Install dependencies, full log11.4k
  • Read payments/refund.ts4.8k
  • Edit payments/refund.ts2.6k
  • Latest turn: run the test suite2.4k
An illustrative request from a coding session.

Privacy

Your prompts are never stored.

Requests are compressed in memory and passed straight on. For each request we keep the token count, the model and the cost. Never the content.

What SOMA keeps

    A prompt with code passes through SOMA and disappears; only its token counts, model and cost are kept.

    Zero data retention

    We keep nothing. You pick where your code runs: any of our providers, or your own model.

    Never used for training

    Your code trains nothing. Choose the provider you trust, or bring your own model.

    Inside your network

    Requests are compressed inside your VPC before being forwarded to your selected model provider.

    Integration

    Two lines. Nothing else changes.

    Point any OpenAI- or Anthropic-compatible client at SOMA. Codex, Claude Code and your own agents keep working as they are.

    Specifications

    For your security review.

    Your VPC or SOMA EU cloudCodexClaude CodeYour agentOpenAI + Anthropic APISSOfail-openSOMAIn memorycounts, model, costProviderProviderProviderzero data retentionGDPREU
    API
    OpenAI-compatible (Chat Completions and Responses) and Anthropic Messages (/v1/messages)
    Clients
    Codex, Claude Code and any OpenAI- or Anthropic-compatible client
    Compression
    Algorithmic and deterministic, no language model in the loop, so your agents feel no added delay
    Reliability
    Fail-open: if compression fails, the original request goes through
    Data
    Processed in memory. Stored per request: token counts, model and cost
    Model providers
    Zero data retention, no training on your data
    Access
    SSO with SAML or OIDC, audit logs, limits per key and per team
    Deployment
    Managed in the EU, or self-hosted in your VPC
    Compliance
    GDPR with a data processing agreement, EU data residency

    Getting started

    Three ways to start.

    SOMA App

    Self-serve, with API keys, credits and a usage dashboard.

    Open SOMA App

    Pilot

    Two to four weeks on your own agents, with a report of the tokens SOMA saved.

    Plan a pilot

    Enterprise

    A dedicated SOMA, managed in the EU or in your VPC, with SSO and audit logs.

    Talk to us

    Contact

    Talk to our engineers

    A security questionnaire, an architecture review or a pilot on your agents: tell us what you need.

    Aleksander Muszyński

    Email usOpen SOMA App