condense.chat
condense.chat provides a drop-in proxy that compacts coding agent session history before it reaches Anthropic or OpenAI, reducing token costs by roughly two-thirds on long, multi-turn sessions. The service uses proprietary compaction models—Helene 1 for general traffic and Adeline 1 for long agent traces—and requires only a base URL change to integrate. On a real 938-turn session, costs dropped from $154 to $43 on Sonnet and from $771 to $214 on Opus.
- Developer Tools
- Artificial Intelligence
- Software Only
- AI Agents
Funding
Founders
Founder details are not available yet.
Product
Problem
Coding agents become expensive because every turn re-sends the entire growing session history, forcing the model to re-read the same old context hundreds of times. Most of the bill is not raw input but cache reads and writes, which compound as sessions lengthen and drive up costs disproportionately. In real sessions, cache reads account for 67.7% of spend and cache writes 18.9%, leaving only 0.2% for raw uncached input.
Solution
condense.chat provides a proxy that compacts session history before it reaches the model, shrinking what gets written to and re-read from the cache. The service runs proprietary compaction models—Helene 1 for general traffic and Adeline 1 for long agent traces—that remove low-signal tokens while preserving answer quality. It integrates with both Anthropic and OpenAI SDKs by simply pointing the base URL at a provider route, keeping the user's existing key and changing nothing else. On real sessions, condense removes about two-thirds of total spend, with a 938-turn session dropping from $154 to $43 on Sonnet and from $771 to $214 on Opus. Compaction calls run in tens of milliseconds, adding no noticeable latency to the session.
Target Audience
Primary customers are individual developers and teams using coding agents like Claude Code, Codex, and OpenCode who want to reduce token costs on long, multi-turn sessions without sacrificing answer quality.
Features
- Helene 1 compaction model with auto mode that reads each input and decides how much to remove, or a fixed ratio set by the user
- Adeline 1 model optimized for long agent traces where deep compaction matters more than latency
- Proxy compatible with both Anthropic and OpenAI SDKs, requiring only a base URL change
- Cache-aware compression that targets the cache write and read tiers, which account for 67.7% of spend
- Compaction calls run in tens of milliseconds, adding no noticeable latency to the session
- Open-source minmax-bench harness that replays real sessions to measure bill reduction against competitors like headroom
- EU data residency with AES-256 encryption, zero data retention on request, and no training on user data