Problem
Agents pay for the same context, again and again.
Agents repeatedly send conversation history, tool outputs, code and earlier reasoning back to the model. Every step makes the next one more expensive.
Squeeze the most out of your enterprise AI.
Context compression for AI agents, on the way to the model.
Solution
Compress on the way to the model.
SOMA compresses the context and the chain of thought before they reach the model, and keeps what the task needs. The agent and the model stay as they are.
Tokens sent per turn
- Turn
- 01 / 12
- Sent this turn
- 5.1k tokens
- Session cost
- $0.00
Illustrative 12-turn session at $3 per 1M input tokens. Savings shown at 40%, SOMA's current estimate for Mistral Large 4 on GitHub Copilot CLI.
Why SOMA
Built for the way agents work.
Agent-native
Designed for multi-step workflows, where context grows with every turn. SOMA compresses each request as it happens, not once at the end.
- Every turnEach request is compressed before it reaches the model, from the first step to the last.
- Longer runs, bigger savingsThe more turns an agent takes, the more context it carries, and the more there is to trim.
- Same answersThe aim is the same output quality from a leaner context.
Drop-in
Sits between the agent and the model, without redesigning the workflow. Point the agent at SOMA and keep everything else.
- One endpointThe agent calls SOMA instead of the model. Prompts, tools and code stay as they are.
- Nothing to retrainNo fine-tuning and no new framework to adopt.
- Easy to switch offRemove SOMA and the agent talks to the model directly again.
Model-independent
Works with the coding agents teams already run and the models behind them. SOMA speaks the OpenAI API, so the same layer sits under each of them.
- Your harnessGitHub Copilot CLI, Codex, Claude Code and OpenCode, with the same layer under each.
- Your modelFrontier APIs, open-weight or self-hosted models behind the same layer.
- No lock-inChange the harness or the model without changing how compression works.
Contact
Let's talk.
For enterprise evaluations and investor conversations, book a short call with Oli.
Intro call with Oli
