Skip to content
AI Primer
release

Condense.chat opens Adeline 1 proxy for 9% agent-loop compaction

Condense.chat opened a compression proxy that strips tokens with Helene 1 and compacts settled agent loops with Adeline 1 to about 9% of their size. The service claims 100M saved tokens and 3× plan extension for Claude or Codex users, so test it on non-sensitive workflows first.

4 min read
Condense.chat opens Adeline 1 proxy for 9% agent-loop compaction
Condense.chat opens Adeline 1 proxy for 9% agent-loop compaction

TL;DR

Condense's homepage pitches the system as "Less context, same Claude Code," with typical savings listed at 64% to 70% and faithfulness at 94.2%. The API docs expose both proxy and rewrite modes, and the dense GitHub repo describes the local CLI as a Claude Code interceptor. Condense's token-bill post is the cost-model piece behind the launch.

Two-stage compression

Condense is selling the proxy as a context compactor, not a model router. The two named models split the job into extraction and summarization:

  • Helene 1: scores tokens and keeps a strict subset of the original agent context, according to TestingCatalog's report.
  • Adeline 1: rewrites aged, settled agent loops into short summaries that preserve intents, file paths, identifiers, errors, and code, according to the same report.
  • Untouched skeleton: system prompts, user messages, final answers, and paired tool calls with results are preserved byte for byte in TestingCatalog's description.
  • Compaction target: TestingCatalog's launch thread says Adeline 1 packs settled loops to about 9% of their size.

The claim lives or dies on old agent loops being compressible without corrupting the active working set. That is exactly where long coding sessions get weirdly expensive.

Token-bill math

Condense's token-bill post frames the problem as repeated cache reads and writes, not raw prompt text. TestingCatalog's report gives the launch numbers:

  • Dataset: 12 real coding sessions and 18,333 assistant turns, according to TestingCatalog's report.
  • Bill shape: cache reads made up 67.7% of a typical bill in the reported cost split.
  • Deep replay: one 938-turn session fell 72.3%, from $154 to $43 on Sonnet and from $771 to $214 on Opus, per the launch report.
  • Across sizes: dollar-weighted savings landed near 66%, while sessions above 400,000 tokens passed 53% savings, according to TestingCatalog's write-up.
  • Faithfulness: Condense reported 94.2% answer faithfulness against uncompressed transcripts in the same report.

The sharp bit is the cache line item: in a deep agent run, the model keeps paying to reread context the agent has already moved past.

Drop-in routes

The official docs describe Condense as a base-URL swap with two provider dialects:

  • Base URL: https://api.condense.chat.
  • Anthropic route: POST /anthropic/v1/messages, described in the docs as drop-in for the Anthropic SDK and Claude Code.
  • OpenAI route: POST /openai/v1/chat/completions, described as drop-in for the OpenAI SDK.
  • Pass-through: other /{provider}/... paths, including model lists and embeddings, are forwarded verbatim.
  • Function switch: X-Condense-Function: proxy | rewrite, with proxy as the default.

For Claude Code, the docs say the installer wraps the official claude CLI, stores the Condense token under ~/.condense/, and exports ANTHROPIC_BASE_URL=https://api.condense.chat/anthropic so Claude talks through Condense.

Keys and rewrite mode

The docs say each request carries two keys: a Condense API key in X-Condense-Auth-Token and the upstream provider key in the provider's normal header. The same page says Condense never stores the upstream key, only a SHA-256 fingerprint in the usage ledger.

rewrite mode is the inspection path. The API docs say it returns the rewritten request body as JSON, makes no upstream call, and incurs no upstream cost.

The dense repo is narrower than the proxy docs: its README calls dense a local CLI that routes Claude Code through the Condense proxy "today." TestingCatalog's report says the broader proxy currently drives Claude Code, Codex, and OpenCode across macOS, Linux, and Windows.

Share on X