npm i -g billion-context MIT runs on your machine, sees nothing else
billion-context is a local context-compression proxy for AI coding agents — billion-context is all you need. Small context windows are enough: it keeps a 100K-window model working for days at a time, cuts token spend by up to 5×, and accounts for every token it moves.
×5
fewer tokens per task
174,327
model calls in the longitudinal study
18.8B
cumulative input tokens compressed
0
window violations on 204,800-token models
Figures from the open-sourced study (4.5 months, 3 hosts, marathon sessions of 8,584–12,049 calls) — see the paper below.
Works with the agents you already run
Not a bigger window — a better layout. Four mechanisms keep months of work inside an ordinary context limit:
When history outgrows its budget, only the newly consumed slice is folded into a compact summary block — once per block, ever. The window oscillates in a bounded band instead of growing without limit; production sessions ran months at ten-thousand-plus calls. Any block decompresses back to full text when you need the detail.
A fold deactivates old ranges without touching the front of the conversation, so the prompt-cache prefix survives turn after turn — 94.2% of active input was served from cache in the study. Rewrite-style compaction invalidates the whole prefix and re-bills it. This does not.
A versioned compression doctrine lives in the system prompt; a growth-gated nudge says when folding pays off. The agent judges what the task still needs, may refuse, and compresses between steps — never mid-step.
Each request is sampled into a per-session ledger — new content, cache reads, compression repay, TTL expiry — plotted in percentage points on a trajectory chart. Audit real savings with /acp in the terminal or the built-in web UI at /__bili/.
One npm package, Node 22+. Then pick your connection style:
Installs straight into your client for the deepest integration: bili plugin install pi. Covers pi · omp · claude · codex · opencode · dsh · kimi · hermes · zcode.
Wraps your client's command and injects bili automatically, zero config edits: bili pi, bili codex, bili claude — about twenty agents.
Any client that accepts a custom endpoint works out of the box: point its base URL at http://localhost:8787/bili/<your upstream>.
npm install -g billion-context --prefix=~/.local
bili pi # or: bili codex / claude / omp / opencode …Connected? Type /acp in your agent for the live panel and /acp-cache for the token economics. Per-agent details, MITM mode and troubleshooting: the connect guide.
Peer-reviewable by anyone
The method behind the product, open-sourced under MIT inside the repository and updated as a living document. A 4.5-month production study across three hosts: 174,327 model calls, 18.76B cumulative input tokens, zero window violations.