{"info":{"author":"Headroom Contributors","author_email":null,"bugtrack_url":null,"classifiers":["Development Status :: 4 - Beta","Intended Audience :: Developers","Operating System :: OS Independent","Programming Language :: Python :: 3","Programming Language :: Python :: 3.10","Programming Language :: Python :: 3.11","Programming Language :: Python :: 3.12","Programming Language :: Python :: 3.13","Programming Language :: Python :: 3.14","Topic :: Scientific\/Engineering :: Artificial Intelligence","Topic :: Software Development :: Libraries :: Python Modules","Typing :: Typed"],"description":"<div align=\"center\">\n\n<img src=\".github\/assets\/hero.svg\" alt=\"Headroom \u2014 the context compression layer for AI agents. A 55,957 token agent prompt compresses to the 24,340 tokens actually sent to the model, and the FATAL line at item 67 survives byte for byte.\" width=\"880\">\n\n<a href=\"https:\/\/trendshift.io\/repositories\/20881\" target=\"_blank\"><img src=\"https:\/\/trendshift.io\/api\/badge\/repositories\/20881\" alt=\"headroomlabs-ai\/headroom | Trendshift \u2014 #1 Repository Of The Day\" width=\"250\" height=\"55\"\/><\/a>\n\n<p>\n  <a href=\"https:\/\/github.com\/headroomlabs-ai\/headroom\"><img src=\"https:\/\/img.shields.io\/github\/stars\/headroomlabs-ai\/headroom?style=flat&color=00F0B5&labelColor=0C1118&label=stars\" alt=\"GitHub stars\"><\/a>\n  <a href=\"https:\/\/github.com\/headroomlabs-ai\/headroom\/actions\/workflows\/ci.yml\"><img src=\"https:\/\/github.com\/headroomlabs-ai\/headroom\/actions\/workflows\/ci.yml\/badge.svg\" alt=\"CI\"><\/a>\n  <a href=\"https:\/\/pypi.org\/project\/headroom-ai\/\"><img src=\"https:\/\/img.shields.io\/pypi\/v\/headroom-ai.svg?color=00F0B5&labelColor=0C1118&label=pypi\" alt=\"PyPI\"><\/a>\n  <a href=\"https:\/\/www.npmjs.com\/package\/headroom-ai\"><img src=\"https:\/\/img.shields.io\/npm\/v\/headroom-ai.svg?color=00F0B5&labelColor=0C1118&label=npm\" alt=\"npm\"><\/a>\n  <a href=\"https:\/\/huggingface.co\/chopratejas\/kompress-v2-base\"><img src=\"https:\/\/img.shields.io\/badge\/model-kompress--v2--base-65D8FF?labelColor=0C1118\" alt=\"Model\"><\/a>\n  <a href=\"https:\/\/docs.headroomlabs.ai\/docs\"><img src=\"https:\/\/img.shields.io\/badge\/docs-online-00F0B5?labelColor=0C1118\" alt=\"Docs\"><\/a>\n  <a href=\"LICENSE\"><img src=\"https:\/\/img.shields.io\/badge\/license-Apache%202.0-98A4B3?labelColor=0C1118\" alt=\"License\"><\/a>\n<\/p>\n\n<p>\n  <b><a href=\"https:\/\/docs.headroomlabs.ai\/docs\/quickstart\">Quickstart<\/a><\/b> \u00b7\n  <a href=\"#get-started-60-seconds\">Install<\/a> \u00b7\n  <a href=\"#proof\">Proof<\/a> \u00b7\n  <a href=\"#agent-compatibility\">Agents<\/a> \u00b7\n  <a href=\"https:\/\/docs.headroomlabs.ai\/docs\">Docs<\/a> \u00b7\n  <a href=\"https:\/\/discord.gg\/yRmaUNpsPJ\">Discord<\/a> \u00b7\n  <a href=\"llms.txt\">llms.txt<\/a>\n<\/p>\n\n<sub><b>AI agents \/ LLMs:<\/b> read <a href=\"llms.txt\"><code>\/llms.txt<\/code><\/a> here, or fetch\n<a href=\"https:\/\/docs.headroomlabs.ai\/llms.txt\">the live index<\/a> \u00b7\n<a href=\"https:\/\/docs.headroomlabs.ai\/llms-full.txt\">full docs blob<\/a>.<\/sub>\n\n<\/div>\n\n<!-- mcp-name: io.github.headroomlabs-ai\/headroom -->\n\nHeadroom compresses everything your AI agent reads \u2014 tool outputs, logs, RAG\nchunks, files, and conversation history \u2014 before it reaches the LLM. Same\nanswers, fraction of the tokens. Compression runs on your machine; no prompt or\nfile content is sent anywhere to be compressed.\n\n<div align=\"center\">\n  <img src=\"HeadroomDemo-Fast.gif\" alt=\"Headroom compressing a 10,144 token log dump to 1,260 tokens while preserving the FATAL line\" width=\"820\">\n  <br><sub>10,144 \u2192 1,260 tokens. The same <code>FATAL<\/code> found.<\/sub>\n<\/div>\n\n## What it does\n\n- **Library** \u2014 `compress(messages)` in Python or TypeScript, inline in any app.\n- **Proxy** \u2014 `headroom proxy --port 8787`, zero code changes, any language.\n- **Agent wrap** \u2014 `headroom wrap claude|codex|grok|copilot|cursor|aider|opencode|cline|continue|goose|openhands|openclaw|vibe|omp|zcode` in one command; undo with `headroom unwrap <tool>`.\n- **MCP server** \u2014 `headroom_compress`, `headroom_retrieve`, `headroom_stats` for any MCP client.\n- **Cross-agent memory** \u2014 one shared store across Claude, Codex, Gemini and Grok, with automatic dedup.\n- **`headroom learn`** \u2014 mines failed sessions and writes corrections to `CLAUDE.local.md` (default, gitignored), `CLAUDE.md`, `AGENTS.md`, `GEMINI.md` or `GROK.md`.\n- **Output token reduction** \u2014 trims what the model *writes back*, not only what you send. See [below](#output-token-reduction).\n- **Reversible (CCR)** \u2014 originals are cached locally and retrieved on demand.\n\n## How it works\n\n```\n Your agent \/ app\n   (Claude Code, Cursor, Codex, LangChain, Agno, Strands, your own code\u2026)\n        \u2502   prompts \u00b7 tool outputs \u00b7 logs \u00b7 RAG results \u00b7 files\n        \u25bc\n    \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n    \u2502  Headroom   (runs locally \u2014 your data stays here)  \u2502\n    \u2502  \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500  \u2502\n    \u2502  CacheAligner  \u2192  ContentRouter  \u2192  CCR            \u2502\n    \u2502                    \u251c\u2500 SmartCrusher   (JSON)        \u2502\n    \u2502                    \u251c\u2500 CodeCompressor (AST)         \u2502\n    \u2502                    \u2514\u2500 Kompress-v2-base (text, HF)  \u2502\n    \u2502                                                    \u2502\n    \u2502  Cross-agent memory  \u00b7  headroom learn  \u00b7  MCP     \u2502\n    \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518\n        \u2502   compressed prompt  +  retrieval tool\n        \u25bc\n LLM provider  (Anthropic \u00b7 OpenAI \u00b7 Bedrock \u00b7 \u2026)\n```\n\n- **ContentRouter** detects the content type and selects a compressor for it.\n- **SmartCrusher \/ CodeCompressor \/ Kompress-v2-base** handle JSON, source code and prose respectively.\n- **CacheAligner** flags volatile content that would bust a provider KV-cache prefix. It never rewrites prompts.\n- **CCR** stores originals locally so the model can call `headroom_retrieve` when it needs the full text.\n\n\u2192 [Architecture](https:\/\/docs.headroomlabs.ai\/docs\/architecture) \u00b7\n[CCR](https:\/\/docs.headroomlabs.ai\/docs\/ccr) \u00b7\n[Kompress-v2-base model card](https:\/\/huggingface.co\/chopratejas\/kompress-v2-base)\n\n## Get started (60 seconds)\n\n```bash\n# 1 \u2014 Install\nuv tool install --python 3.13 \"headroom-ai[all]\"  # CLI in a self-contained env\npip install \"headroom-ai[all]\"                    # Python \u2014 ships the `headroom` CLI\nnpm install headroom-ai                           # TypeScript SDK only \u2014 no CLI\n\n# 2 \u2014 Pick a mode\nheadroom deploy                         # turnkey local deployment + agent config\nheadroom wrap claude                    # wrap a coding agent\nheadroom proxy --port 8787              # drop-in proxy, zero code changes\n# or: from headroom import compress     # inline library\n\n# 3 \u2014 Check it and watch the savings\nheadroom doctor                         # health check \u2014 confirms routing works\nheadroom perf\nheadroom dashboard                      # live savings (proxy must be running)\n```\n\nInline, in Python:\n\n```python\nfrom headroom import compress\nfrom openai import OpenAI\n\nmessages = [{\"role\": \"user\", \"content\": \"Analyze these results\"}]\nresult = compress(messages, model=\"gpt-4o\")\n\nclient = OpenAI()\nresponse = client.chat.completions.create(model=\"gpt-4o\", messages=result.messages)\nprint(f\"Saved {result.tokens_saved} tokens ({result.compression_ratio:.0%})\")\n```\n\nLaunch a wrapped agent session each time, so the setup runs. `headroom wrap`\nstarts a local proxy, installs **[Serena](https:\/\/github.com\/oraios\/serena)** for\nsemantic code navigation, and launches the agent configured to route through\nHeadroom. For Claude Code, Serena is registered for the wrapped project only\n(as a `local`-scope MCP server in `~\/.claude.json`). Use\n`--code-memory-scope user` to make it available in every project, or\n`--code-memory none` to skip it. `headroom unwrap` removes either registration.\n\nThe `headroom` CLI ships only in the PyPI package. The npm `headroom-ai` package\nis the TypeScript SDK \u2014 a library you import\n(`import { compress } from 'headroom-ai'`) \u2014 and provides no `headroom` command.\n\n## Proof\n\nFour scenarios built from real MCP server output formats, measured with the\nprovider tokenizer and the shipped `compress()`. Seeded and offline, so you get\nthe same numbers we did:\n\n```bash\nuv run python benchmarks\/index_proof_table.py --seed 20260902\n```\n\n| Scenario | Before | After | Saved |\n|---|---:|---:|---:|\n| Code search (100 results) | 17,199 | 13,597 | **21%** |\n| SRE incident debugging | 55,957 | 24,340 | **57%** |\n| Codebase exploration | 58,801 | 33,895 | **42%** |\n| GitHub issue triage | 46,067 | 32,429 | **30%** |\n\nSavings scale with how repetitive the payload is. Repeated JSON arrays and log\nlines clear 90% in `benchmarks\/bench_latency.py`; prose and already-dense output\ncompress very little. Run `headroom savings` against your own traffic for the\nnumber that applies to you.\n\nCompression costs **well under a millisecond** \u2014 0.21 ms p50 on a 10K-token JSON\nsearch result, 1.4 ms at 100K tokens \u2014 so it does not show up in agent latency.\n\n**Accuracy.** `python -m headroom.evals suite --tier 1`:\n\n| Benchmark | Category | N | Baseline | Headroom | Delta |\n|---|---|---:|---:|---:|---|\n| GSM8K | Math | 100 | 0.870 | 0.870 | \u00b10.000 |\n| TruthfulQA | Factual | 100 | 0.530 | 0.560 | +0.030 |\n| SQuAD v2 | QA | 100 | \u2014 | 97% | at 19% compression |\n| BFCL | Tools | 100 | \u2014 | 97% | at 32% compression |\n\nAt N=100 a delta of \u00b10.03 falls inside the confidence interval, so TruthfulQA\nshows no detectable difference rather than an improvement.\n[Methodology \u2192](https:\/\/docs.headroomlabs.ai\/docs\/benchmarks)\n\n## Output token reduction\n\nEverything above shrinks the prompt you **send**. You also pay for every token\nthe model **writes back**, and on Opus-class models output costs 5\u00d7 input. Much\nof that output is ceremony: \"Great, let me\u2026\" preambles, code re-printed straight\nback at you, and deep reasoning spent on routine steps like reading a file.\n\nHeadroom trims it from the proxy, with no change to your code:\n\n- **Verbosity steering** appends a short \"be terse, don't restate context\" note to the *end* of the system prompt, so your prompt cache still hits.\n- **Effort routing** dials thinking effort down when a turn is only the model resuming after a tool result \u2014 a file read, a passing test. New questions and errors keep full effort.\n\nBoth apply to Anthropic `\/v1\/messages` and to OpenAI-compatible\n`\/v1\/chat\/completions` and `\/v1\/responses`. Effort routing uses\n`reasoning_effort` on OpenAI and `thinking.budget_tokens` \/ `output_config.effort`\non Anthropic, with the same clamp-only invariant and the same `output_shaper:*`\nlabels on both paths.\n\n```bash\nexport HEADROOM_OUTPUT_SHAPER=1     # off by default\nheadroom proxy --port 8787\n```\n\n> **Already running a proxy?** These switches are read live on every request, so\n> a proxy that `headroom wrap` *reused* rather than started would not see a value\n> you export afterwards \u2014 its environment was snapshotted at launch. `headroom\n> wrap` hot-syncs your current settings to the running proxy over a loopback\n> `POST \/admin\/runtime-env`, so they take effect with no restart and no dropped\n> requests. On a shared proxy these overrides are global; the last explicit\n> setting wins.\n\n**Terseness you didn't have to configure.** People rarely state how terse they\nwant answers \u2014 they show it, by interrupting long replies or moving on before\nthey could have read them. `headroom learn --verbosity` reads past sessions and\npicks the level:\n\n```bash\nheadroom learn --verbosity            # dry run \u2014 preview what it found\nheadroom learn --verbosity --apply    # save it; the proxy picks it up\n```\n\n**Measuring it.** Output savings are counterfactual \u2014 we never see what the model\n*would* have written \u2014 so Headroom reports an estimate with a confidence range\nand labels it as one:\n\n```bash\nheadroom output-savings\n# Reduction: 31.7%  (95% CI 27.7% \u2026 35.7%)   [estimated]\n```\n\nFor a measured number instead, hold out 10% of conversations as an unshaped\ncontrol: `export HEADROOM_OUTPUT_HOLDOUT=0.1`. The dashboard's **Output Tokens\nSaved** card then reads `measured` rather than `estimated`, with the band.\n\n\u2192 [Output token reduction](https:\/\/docs.headroomlabs.ai\/docs\/savings)\n\n## Agent compatibility\n\n| Agent | `headroom wrap` | Notes |\n|---|:---:|---|\n| Claude Code | \u2705 | `--memory` \u00b7 `--code-graph` \u00b7 `--1m` \u00b7 `--tool-search` |\n| Codex | \u2705 | shares memory with Claude |\n| Grok CLI | \u2705 | routes via `GROK_MODELS_BASE_URL` |\n| Cursor | Manual setup | starts the proxy and prints base URLs for Cursor settings |\n| Aider | \u2705 | starts proxy + launches |\n| Copilot CLI | \u2705 | starts proxy + launches |\n| VS Code Copilot | \u2705 | transparent proxy; keeps the selected model |\n| OpenClaw | \u2705 | installs as a ContextEngine plugin |\n| OpenCode | \u2705 | injects config \u00b7 starts proxy + launches |\n| Cline | \u2705 | starts proxy + injects config |\n| Continue | \u2705 | starts proxy + injects config |\n| Goose | \u2705 | starts proxy + launches |\n| OpenHands | \u2705 | starts proxy + launches |\n| Mistral Vibe | \u2705 | starts proxy + launches |\n| Oh My Pi | \u2705 | injects config \u00b7 starts proxy + launches |\n| Cortex Code | Library only | 60\u201365% savings in library mode; no `wrap` |\n| Kimi CLI | \u2705 | OAuth bearer forwarded \u2014 log in once |\n| ZCode | \u2705 | starts the proxy and prints base URLs for ZCode settings |\n\nAny OpenAI-compatible client works through `headroom proxy`. MCP-native clients:\n`headroom mcp install`. Undo durable wrapping with `headroom unwrap <tool>`\n(`claude`, `copilot`, `codex`, `grok`, `kimi`, `omp`, `opencode`, `openclaw`,\n`zcode`). Registry authors should use the canonical [`server.json`](server.json)\nrather than reconstructing the `headroom mcp serve` contract from prose.\n\nFor Anthropic `\/v1\/messages`, `--mode cache` skips automatic `--memory` context injection so the provider prefix remains stable. OpenAI chat\/responses and Gemini append memory to the live-zone tail. Use `--mode token` when you need automatic memory context on the Anthropic path.\n\n<details>\n<summary><b>GitHub Copilot CLI subscription mode<\/b><\/summary>\n\nHeadroom can route Copilot CLI subscription traffic through the local proxy:\n\n```bash\nheadroom copilot-auth login\nheadroom wrap copilot --subscription -- --model gpt-4o\n```\n\nThe wrapper exchanges Headroom's reusable GitHub OAuth token for Copilot's\nshort-lived API token and prints the upstream endpoint as\n`COPILOT_PROVIDER_API_URL=...` at launch. `headroom copilot-auth login` stores a\nHeadroom-specific Copilot OAuth token, rather than relying on generic GitHub or\nCopilot CLI tokens that can read account metadata but are still rejected by\nCopilot's token-exchange endpoint.\n\nFor GitHub Enterprise Server or a custom-domain Copilot deployment, set one of\nthese before launching. If both are set, the URL wins:\n\n```bash\nexport GITHUB_COPILOT_ENTERPRISE_DOMAIN=ghe.example.com\nexport GITHUB_COPILOT_ENTERPRISE_URL=https:\/\/ghe.example.com\n```\n\nFor GitHub.com Enterprise Cloud URLs such as\n`github.com\/enterprises\/your-enterprise`, set neither \u2014 Headroom uses GitHub's\nnormal token-exchange endpoint and the Copilot API endpoint advertised for the\nsigned-in account.\n\n**Platform support.** macOS auth reuse through Copilot CLI Keychain storage and\nWindows device authentication are live-tested. Copilot CLI 1.0.81 does not expose\nits Windows login through the legacy Credential Manager schema Headroom reads, so\nrun `headroom copilot-auth login` on Windows. Linux Secret Service \/\n`secret-tool` reuse is implemented but not yet validated on a real desktop. In\nDocker and CI, pass an explicit `GITHUB_COPILOT_TOKEN` or\n`GITHUB_COPILOT_GITHUB_TOKEN` instead of relying on host keychain access.\n\n<\/details>\n\n<details>\n<summary><b>GitHub Copilot in VS Code<\/b><\/summary>\n\nHeadroom overrides Copilot's API proxy endpoint, so the VS Code model picker\nstays authoritative. GPT-5.5, GPT-5.6 Luna\/Sol\/Terra, Claude Sonnet\/Opus and\nother Copilot models keep their original model IDs while traffic passes through\nthe local compression proxy. Headroom does not patch VS Code or change Codex\nsettings.\n\n```bash\nheadroom copilot-auth login\nheadroom wrap vscode\n```\n\nKeep the command running and use Copilot normally. The short-lived upstream\nCopilot token is held only in the proxy process.\n[Full guide \u2192](https:\/\/docs.headroomlabs.ai\/docs\/vscode-copilot)\n\n<\/details>\n\n<details>\n<summary><b>Claude Code in VS Code<\/b><\/summary>\n\nThe official Claude Code extension embeds Claude Code and reads the same user\nsettings as the CLI. Install the proxy extra, then run the wrapper from the\nproject you will open in VS Code:\n\n```bash\npip install \"headroom-ai[proxy]\"\nheadroom wrap vscode-claude\n# Optional: persist Claude Code's client-owned 1M model selector\nheadroom wrap vscode-claude --1m\n```\n\nReload the VS Code window on first run. Keep the wrapper terminal running while\nyou use the Claude Code panel; the dashboard or proxy log printed at startup\nshows requests and savings. Your Anthropic authentication and selected model are\npreserved. `Ctrl+C` stops the proxy; `headroom unwrap vscode-claude` restores the\nsettings that existed before setup.\n\nWith `--1m`, Headroom writes the resolved `[1m]` model selector to Claude Code's\ntop-level user setting and restores the exact prior model on\n`headroom unwrap vscode-claude` or when setup runs again without `--1m`. Use\n`headroom wrap vscode-claude --no-configure --1m` to print the setting without\nwriting files.\n\nClaude Code and Anthropic own 1M model support and account eligibility. Local\ntests prove persisted settings and restoration, not a live entitled VS Code\nsession or the resulting context window.\n[Full guide \u2192](https:\/\/docs.headroomlabs.ai\/docs\/vscode-claude-code)\n\n<\/details>\n\n## When to use \u00b7 when to skip\n\n**Good fit if you** run coding agents daily and want savings without touching\nyour code, work across several agents and want one shared memory, or need\ncompression that is reversible \u2014 originals stay retrievable through CCR for the\nconfigured TTL.\n\n**Skip it if you** only use one provider's native compaction and don't need\ncross-agent memory, or work in a sandbox where local processes can't run.\n\nHeadroom pays off on long agent sessions with heavy tool output. Short\nconversational exchanges, prose, and already-dense payloads see little or no\nreduction, and blocks under `min_input_words` come back byte-identical.\n[Limitations](https:\/\/docs.headroomlabs.ai\/docs\/limitations) has the full list.\n\n<details>\n<summary><b>Integrations \u2014 drop Headroom into any stack<\/b><\/summary>\n\n| Your setup | Hook in with |\n|---|---|\n| Any Python app | `compress(messages, model=\u2026)` |\n| Any TypeScript app | `await compress(messages, { model })` |\n| Anthropic \/ OpenAI SDK | `withHeadroom(new Anthropic())` \u00b7 `withHeadroom(new OpenAI())` |\n| Vercel AI SDK | `wrapLanguageModel({ model, middleware: headroomMiddleware() })` |\n| LiteLLM | `litellm.callbacks = [HeadroomCallback()]` |\n| LangChain | `HeadroomChatModel(your_llm)` |\n| Agno | `HeadroomAgnoModel(your_model)` |\n| Strands | [Strands guide](https:\/\/docs.headroomlabs.ai\/docs\/strands) |\n| ASGI apps | `app.add_middleware(CompressionMiddleware)` |\n| Multi-agent | `SharedContext().put \/ .get` |\n| MCP clients | `headroom mcp install` |\n\n<\/details>\n\n<details>\n<summary><b>What's inside<\/b><\/summary>\n\n- **SmartCrusher** \u2014 universal JSON: arrays of dicts, nested objects, mixed types. It keeps error items, values outside the normal statistical range, and first\/last boundaries, selected from field-variance statistics rather than a keyword list.\n- **CodeCompressor** \u2014 AST-aware for Python, JS\/TS, Go, Rust, Java, C\/C++, C#, PHP.\n- **Kompress-v2-base** \u2014 our HuggingFace model, trained on agentic traces.\n- **Image compression** \u2014 40\u201390% reduction through a trained ML router.\n- **CacheAligner** \u2014 flags volatile content that would bust a provider KV-cache prefix; never rewrites prompts.\n- **Live-zone compression** \u2014 only new bytes are compressed (fresh tool output, the latest turn). The frozen prefix stays byte-identical, so the provider cache survives, and history is never dropped.\n- **CCR** \u2014 reversible compression; the model retrieves originals on demand.\n- **Cross-agent memory** \u2014 shared store with agent provenance and auto-dedup.\n- **SharedContext** \u2014 compressed context passing across multi-agent workflows.\n- **`headroom learn`** \u2014 plugin-based failure mining for Claude, Codex and Gemini.\n\n<\/details>\n\n<details>\n<summary><b>Pipeline internals<\/b><\/summary>\n\nOne request lifecycle is shared by `compress()`, the SDKs and the proxy:\n\n`Setup` \u2192 `Pre-Start` \u2192 `Post-Start` \u2192 `Input Received` \u2192 `Input Cached` \u2192\n`Input Routed` \u2192 `Input Compressed` \u2192 `Input Remembered` \u2192 `Pre-Send` \u2192\n`Post-Send` \u2192 `Response Received`\n\n- **Transforms** do the work: CacheAligner \u2192 ContentRouter \u2192 SmartCrusher \/ CodeCompressor \/ Kompress-base, live-zone only. IntelligentContext and RollingWindow were retired in PR-B1.\n- **Pipeline extensions** observe or customise lifecycle stages through `on_pipeline_event(...)`.\n- **Compression hooks** sit alongside the lifecycle as an additional extension seam.\n- **Proxy extensions** are the integration seam for ASGI middleware, routes and startup policy.\n\nProvider- and tool-specific behaviour lives under `headroom\/providers\/`, so core\norchestration stays focused on lifecycle, sequencing and policy:\n\n- CLI\/tool slices \u2014 `headroom\/providers\/claude`, `copilot`, `codex`, `grok`, `openclaw`\n- Provider runtime slices \u2014 `headroom\/providers\/claude`, `gemini`, with shared backend dispatch in `headroom\/providers\/registry.py`\n- `wrap.py`, `client.py`, `cli\/proxy.py` and `proxy\/server.py` delegate env shaping, API target normalisation, backend selection and transport dispatch\n\n<\/details>\n\n## Install\n\n```bash\nuv tool install --python 3.13 \"headroom-ai[all]\"  # CLI, isolated app env\npip install \"headroom-ai[all]\"                    # Python, everything \u2014 includes the CLI\nnpm install headroom-ai                           # TypeScript SDK (library only)\ndocker pull ghcr.io\/headroomlabs-ai\/headroom:latest\n```\n\nGranular extras: `[proxy]`, `[mcp]`, `[ml]` (Kompress-v2-base), `[code]`,\n`[memory]`, `[vector]` (optional HNSW backend \u2014 needs a C++ toolchain, not in\n`[all]`), `[relevance]`, `[image]`, `[agno]`, `[langchain]`, `[evals]`,\n`[pytorch-mps]` (Apple-GPU memory-embedder offload \u2014 set\n`HEADROOM_EMBEDDER_RUNTIME=pytorch_mps`). Requires **Python 3.10+**.\n\n> `[all]` covers the core stack but not the framework adapters. Install those\n> separately: `pip install \"headroom-ai[langchain]\"`, and likewise `[agno]`,\n> `[strands]`, `[anyllm]`, `[bedrock]`.\n\n\u2192 [Installation guide](https:\/\/docs.headroomlabs.ai\/docs\/installation) \u2014 Docker\ntags, persistent service, PowerShell, devcontainers.\n\n<details>\n<summary><b>uv, pipx, and MCP clients that don't inherit your PATH<\/b><\/summary>\n\nPrefer `uv tool install` for the CLI so the command lives in an isolated app\nenvironment. On macOS, pass `--python 3.13` if your default `python3` is newer\nthan the current wheel set:\n\n```bash\nbrew install python@3.13  # if 3.13 is not already available\nuv tool install --python 3.13 \"headroom-ai[all]\"\nuv tool update-shell      # if ~\/.local\/bin is not on PATH\nheadroom --version\n```\n\nCodex and other MCP clients often cannot inherit an interactive shell `PATH`.\nConfigure the absolute path returned by `command -v headroom`:\n\n```toml\n[mcp_servers.headroom]\ncommand = \"\/Users\/you\/.local\/bin\/headroom\"\nargs = [\"mcp\", \"serve\"]\n```\n\n`command = \"headroom\"` only works when the client starts with a `PATH` that\nalready includes the uv tool directory.\n\nWith pipx, choose the interpreter explicitly:\n\n```bash\npipx install --python python3.13 \"headroom-ai[all]\"\n```\n\nNative wheels currently cover macOS Apple Silicon and Linux. On Intel macOS, use\nthe Docker-native install until native wheel support lands.\n\n**CPU requirement (x86\/x86_64).** The ONNX-backed features \u2014 Magika content\ndetection and embedding relevance \u2014 use a precompiled ONNX Runtime that needs\n**AVX2**. On x86 hosts without AVX2 (some Docker\/QEMU setups, older cloud VMs)\nHeadroom falls back to its non-ONNX paths \u2014 BM25 relevance, heuristic detection \u2014\nrather than crashing. `arm64` and Apple Silicon need no AVX2.\n\n<\/details>\n\n<details>\n<summary><b>Updating<\/b><\/summary>\n\n```bash\nheadroom update          # detects pip \/ pipx \/ uv tool and upgrades in place\nheadroom update --check  # report the latest release without upgrading\nheadroom update --pre    # include pre-releases\n```\n\n`headroom update` works out how Headroom was installed (pip\/venv, `pip --user`,\npipx, uv tool) and runs the matching upgrade on macOS, Linux and Windows. For git\ncheckouts, editable installs, Docker images and externally-managed system Pythons\n(PEP 668) it prints the correct manual step instead of guessing.\n\nThe proxy also prints a one-line \"update available\" notice at startup. It checks\nPyPI at most once a day, in the background, and never blocks. Opt out with\n`HEADROOM_UPDATE_CHECK=off`; it is also skipped in `--stateless` mode and CI.\n\n<\/details>\n\n<details>\n<summary><b>Corporate networks and SSL inspection<\/b><\/summary>\n\nIf `pip install \"headroom-ai[all]\"` fails with `CERTIFICATE_VERIFY_FAILED`\n(`unable to get local issuer certificate`), your network runs SSL inspection \u2014 a\nMITM proxy presenting a company CA. The build backend (`maturin`) downloads\n`rustup` over a connection your TLS stack does not trust. Install Rust first so\nthe build never fetches it:\n\n```bash\n# macOS \/ Linux\ncurl --proto '=https' --tlsv1.2 -sSf https:\/\/sh.rustup.rs | sh && rustup default stable\n# Windows\nwinget install Rustlang.Rustup && rustup default stable\n```\n\nRestart your shell, then install. A prebuilt wheel avoids the Rust build\nentirely: `pip install --only-binary headroom-ai headroom-ai`. Wheels are\npublished for Windows (`win_amd64`), Linux (`x86_64` \/ `aarch64`) and macOS\n(Apple Silicon and Intel), so those platforms never need a local Rust toolchain \u2014\nthe Rust-first step above is only for the sdist fallback when no wheel matches.\n\n**At runtime** Headroom verifies upstream TLS against the operating system's\ncertificate store \u2014 where IT installs the Zscaler \/ Netskope \/ Palo Alto root \u2014\nplus certifi's public roots, so a managed machine normally needs no setup.\n`headroom doctor --network` shows who signed each upstream certificate, whether\nHeadroom trusts it, and whether a gateway block page is in the way. If the root\nis not in the OS store, export it as PEM and set\n`HEADROOM_CA_BUNDLE=\/path\/to\/root.pem` (`NODE_EXTRA_CA_CERTS` also works). Full\nguide and the domain list for IT: [Corporate networks](https:\/\/docs.headroomlabs.ai\/docs\/corporate-networks).\n\nTwo download hosts can still need attention:\n\n- **`cdn.pyke.io`** \u2014 the ONNX Runtime for the Rust core, fetched with Mozilla's roots only. Pre-provide it with `ORT_STRATEGY=system` and `ORT_LIB_LOCATION=\/path\/to\/onnxruntime`.\n- **`huggingface.co`** \u2014 the `kompress-base` model (Python, OS store) and the Rust core's embedding model (Mozilla's roots only). If the domain is blocked, pre-download and run with `HF_HUB_OFFLINE=1`, or point `HF_ENDPOINT` at a trusted mirror.\n\nRunning with compression disabled (pure gateway) needs neither asset.\n\n**Intel macOS: no prebuilt ONNX Runtime ([#941](https:\/\/github.com\/headroomlabs-ai\/headroom\/issues\/941)).**\n`ort-sys` ships no prebuilt binary for `x86_64-apple-darwin`, so a source build\nfails by default even outside a corporate proxy. Point it at a system runtime:\n\n```bash\nbrew install onnxruntime\nORT_STRATEGY=system \\\nORT_LIB_LOCATION=\"$(brew --prefix onnxruntime)\/lib\" \\\nORT_PREFER_DYNAMIC_LINK=1 \\\n  pip install \"headroom-ai[all]\"\n\n# ORT is dlopen'd at runtime too:\nexport ORT_DYLIB_PATH=\"$(brew --prefix onnxruntime)\/lib\/libonnxruntime.dylib\"\n```\n\n`ORT_LIB_LOCATION` must point at `lib\/`, not the bare prefix, and\n`ORT_PREFER_DYNAMIC_LINK=1` is required \u2014 without it `ORT_STRATEGY=system` still\nattempts static linking, which the Homebrew keg does not provide.\n\n**\"Basic Constraints of CA cert not marked critical\"** only occurs with\n`HEADROOM_CERT_STORE=bundled` (the OS store verifies these roots fine). If TLS\nfails with:\n\n```\n[SSL: CERTIFICATE_VERIFY_FAILED] certificate verify failed:\nBasic Constraints of CA cert not marked critical\n```\n\nthen the corporate CA *is* found and trusted, and adding it to a CA bundle\nchanges nothing. Python 3.13 with OpenSSL 3.x enables `VERIFY_X509_STRICT` by\ndefault, which enforces RFC 5280 \u00a74.2.1.9: a CA cert's `basicConstraints` must be\nmarked critical. Inspection roots such as Zscaler set `CA:TRUE` without the\ncritical bit, so the chain is rejected.\n\n`HEADROOM_TLS_STRICT=0` clears only the strict flag, from every TLS context\nHeadroom controls \u2014 the proxy's httpx upstream client and the\nurllib3\/`huggingface_hub` path used for model downloads. Chain validation,\nsignature, expiry and hostname checks all stay on.\n\n```bash\nHEADROOM_TLS_STRICT=0 headroom proxy --port 8787\n```\n\nThe Rust core's downloads use rustls with Mozilla's roots and are unaffected by\n`HEADROOM_TLS_STRICT` and the CA variables; pre-provision ONNX Runtime with\n`ORT_STRATEGY=system` and models via `HF_ENDPOINT` as above.\n\n<\/details>\n\n## headroom learn\n\n<div align=\"center\">\n  <img src=\"headroom_learn.gif\" alt=\"headroom learn mining failed sessions and writing corrections\" width=\"720\">\n<\/div>\n\n`headroom learn` mines failed sessions and writes corrections to\n`CLAUDE.local.md` (default, gitignored; use `--target CLAUDE.md` for the shared\nteam file), `AGENTS.md` or `GEMINI.md`.\n\u2192 [Failure learning](https:\/\/docs.headroomlabs.ai\/docs\/failure-learning)\n\n## Telemetry\n\nAn anonymous beacon is **on by default**. It reports how compression behaved:\nratios, counters, provider and model IDs, OS and architecture. It never sends\nprompts, completions, code or file paths. It exists so we can see when a release\nregresses a compression ratio across real workloads rather than only our own test\ncorpus.\n\nTurn it off with `HEADROOM_BEACON=off`, the `DO_NOT_TRACK=1` convention, or\n`--offline`. The full field list is in\n[the proxy docs](https:\/\/docs.headroomlabs.ai\/docs\/proxy).\n\n## Headroom for teams\n\nHeadroom OSS is built for individual developers: run `headroom proxy` or\n`headroom wrap` on your laptop and start cutting tokens in minutes, free and\nlocal-first.\n\nRunning it across an engineering org is a different job \u2014 a shared always-on\ndeployment, centralised config and version rollout, org-wide savings dashboards,\nSSO and access control, air-gapped and VPC installs, and someone to call. We help\ncompanies with that, self-hosted with support or fully managed.\n\nIf your team is spending real money on LLM tokens \u2014 Claude Code, Codex, Cursor,\nor agents running in CI \u2014 email **[hello@headroomlabs.ai](mailto:hello@headroomlabs.ai)**\nwith your stack and rough monthly LLM spend.\n\nEverything in this repo stays open source under Apache 2.0. The managed offering\nis for teams that would rather have it deployed, supported and scaled for them.\n\n## Documentation\n\n| Start here | Go deeper |\n|---|---|\n| [Quickstart](https:\/\/docs.headroomlabs.ai\/docs\/quickstart) | [Architecture](https:\/\/docs.headroomlabs.ai\/docs\/architecture) |\n| [Proxy](https:\/\/docs.headroomlabs.ai\/docs\/proxy) | [How compression works](https:\/\/docs.headroomlabs.ai\/docs\/how-compression-works) |\n| [MCP tools](https:\/\/docs.headroomlabs.ai\/docs\/mcp) | [CCR \u2014 reversible compression](https:\/\/docs.headroomlabs.ai\/docs\/ccr) |\n| [Memory](https:\/\/docs.headroomlabs.ai\/docs\/memory) | [Cache optimization](https:\/\/docs.headroomlabs.ai\/docs\/cache-optimization) |\n| [Failure learning](https:\/\/docs.headroomlabs.ai\/docs\/failure-learning) | [Benchmarks](https:\/\/docs.headroomlabs.ai\/docs\/benchmarks) |\n| [Configuration](https:\/\/docs.headroomlabs.ai\/docs\/configuration) | [Limitations](https:\/\/docs.headroomlabs.ai\/docs\/limitations) |\n| [Persistent installs](https:\/\/docs.headroomlabs.ai\/docs\/persistent-installs) | [Savings analytics](https:\/\/docs.headroomlabs.ai\/docs\/savings) |\n\n## Compared to\n\nHeadroom runs locally, covers every content type, works with every major\nframework, and is reversible.\n\n| | Scope | Deploy | Local | Reversible |\n|---|---|---|:---:|:---:|\n| **Headroom** | All context \u2014 tools, RAG, logs, files, history | Proxy \u00b7 library \u00b7 middleware \u00b7 MCP | Yes | Yes |\n| [Compresr](https:\/\/compresr.ai), [Token Co.](https:\/\/thetokencompany.ai) | Text sent to their API | Hosted API call | No | No |\n| OpenAI Compaction | Conversation history | Provider-native | No | No |\n\nHeadroom is the proxy, and it compresses everything flowing through it whatever\nsits upstream. Our recommended companion is\n**[Serena](https:\/\/github.com\/oraios\/serena)** for semantic code navigation,\ninstalled by default when you wrap an agent, plus **Ponytail** if you want leaner\nmodel output. Everything else is your call \u2014 attach a code-memory MCP, Graphify,\nCaveman, or any other MCP server, and Headroom compresses downstream of all of it.\n\n## Contributing\n\n```bash\ngit clone https:\/\/github.com\/headroomlabs-ai\/headroom.git && cd headroom\nuv sync --extra dev && uv run pytest\n```\n\nDevcontainers in `.devcontainer\/` (default, plus `memory-stack` with Qdrant and\nNeo4j). See [CONTRIBUTING.md](CONTRIBUTING.md).\n\n## Community\n\n- **[Discord](https:\/\/discord.gg\/yRmaUNpsPJ)** \u2014 questions, feedback, war stories.\n- **[Kompress-v2-base on HuggingFace](https:\/\/huggingface.co\/chopratejas\/kompress-v2-base)** \u2014 the model behind text compression.\n- **[Claude Code status-line plugin](https:\/\/github.com\/Ship-Wright\/headroom-plugin)** \u2014 live token savings in your status line, by [@Ship-Wright](https:\/\/github.com\/Ship-Wright).\n\n## License\n\nApache 2.0 \u2014 see [LICENSE](LICENSE).\n\n","description_content_type":"text\/markdown; charset=UTF-8; variant=GFM","docs_url":null,"download_url":null,"downloads":{"last_day":-1,"last_month":-1,"last_week":-1},"dynamic":null,"home_page":null,"keywords":"llm, openai, anthropic, claude, gpt, context, token, optimization, compression, caching, proxy, ai, machine-learning","license":null,"license_expression":"Apache-2.0","license_files":["LICENSE","NOTICE"],"maintainer":"Headroom Contributors","maintainer_email":null,"name":"headroom-ai","package_url":"https:\/\/pypi.org\/project\/headroom-ai\/","platform":null,"project_url":"https:\/\/pypi.org\/project\/headroom-ai\/","project_urls":{"AI \/ LLM Index":"https:\/\/docs.headroomlabs.ai\/llms.txt","Changelog":"https:\/\/github.com\/headroomlabs-ai\/headroom\/blob\/main\/CHANGELOG.md","Documentation":"https:\/\/docs.headroomlabs.ai\/docs","Homepage":"https:\/\/docs.headroomlabs.ai","Issues":"https:\/\/github.com\/headroomlabs-ai\/headroom\/issues","Repository":"https:\/\/github.com\/headroomlabs-ai\/headroom"},"provides_extra":["agno","all","anyllm","autogen","bedrock","code","dev","evals","html","image","langchain","langgraph","mcp","memory","memory-stack","ml","otel","proxy","proxy-prod","pytorch-mps","relevance","reports","sandbox","spreadsheet","strands","vector","voice","voice-train"],"release_url":"https:\/\/pypi.org\/project\/headroom-ai\/0.40.0\/","requires_dist":["tiktoken>=0.5.0","pydantic>=2.0.0","litellm<2.0,>=1.96.2","click>=8.3.3","rich>=13.0.0","opentelemetry-api>=1.24.0","ast-grep-cli!=0.44.1,>=0.30.0","pyyaml>=6.0","tomli>=2.0.0; python_full_version < \"3.11\"","tomlkit<1.0,>=0.13.0","truststore>=0.10.0","agno>=1.0.0; extra == \"agno\"","headroom-ai[code,evals,html,image,mcp,memory,ml,otel,proxy,relevance,reports,spreadsheet,voice]; extra == \"all\"","any-llm-sdk>=1.0.0; python_full_version >= \"3.11\" and extra == \"anyllm\"","autogen-agentchat>=0.7; extra == \"autogen\"","boto3>=1.41.0; extra == \"bedrock\"","botocore[crt]>=1.41.0; extra == \"bedrock\"","tree-sitter-language-pack<1.0,>=0.10.0; extra == \"code\"","tree-sitter<0.27,>=0.25.2; extra == \"code\"","pytest>=7.0.0; extra == \"dev\"","pytest-cov>=4.0.0; extra == \"dev\"","pytest-asyncio>=0.21.0; extra == \"dev\"","ruff==0.16.8; extra == \"dev\"","mypy>=1.0.0; extra == \"dev\"","pre-commit>=3.0.0; extra == \"dev\"","openai>=1.0.0; extra == \"dev\"","anthropic>=0.18.0; extra == \"dev\"","litellm<2.0,>=1.96.2; extra == \"dev\"","fastapi>=0.100.0; extra == \"dev\"","uvicorn<1.0,>=0.23.0; extra == \"dev\"","httpx[http2,socks]>=0.24.0; extra == \"dev\"","websockets>=13.0; extra == \"dev\"","opentelemetry-sdk>=1.24.0; extra == \"dev\"","opentelemetry-exporter-otlp-proto-http>=1.24.0; extra == \"dev\"","ollama>=0.4.0; extra == \"dev\"","langchain-ollama>=0.2.0; extra == \"dev\"","hnswlib>=0.8.0; extra == \"dev\"","sqlite-vec>=0.1.6; extra == \"dev\"","sentence-transformers<7.0,>=2.2.0; extra == \"dev\"","numpy>=1.24.0; extra == \"dev\"","openpyxl>=3.1.0; extra == \"dev\"","xlrd>=2.0.1; extra == \"dev\"","xlwt>=1.3.0; extra == \"dev\"","cryptography>=50.0.0; extra == \"dev\"","respx>=0.20.0; extra == \"dev\"","datasets>=5.0.1; extra == \"evals\"","sentence-transformers<7.0,>=2.2.0; (platform_machine != \"x86_64\" and extra == \"evals\") or (sys_platform != \"darwin\" and extra == \"evals\")","numpy>=1.24.0; extra == \"evals\"","scikit-learn>=1.3.0; extra == \"evals\"","anthropic>=0.18.0; extra == \"evals\"","openai>=1.0.0; extra == \"evals\"","trafilatura>=1.6.0; extra == \"html\"","pillow>=12.3.0; extra == \"image\"","sentencepiece>=0.1.99; extra == \"image\"","rapidocr-onnxruntime<2,>=1.4.0; python_full_version < \"3.13\" and extra == \"image\"","rapidocr<4,>=3.0; python_full_version >= \"3.13\" and extra == \"image\"","onnxruntime<2,>=1.7; python_full_version >= \"3.13\" and extra == \"image\"","langchain-core<4.0,>=1.3.3; extra == \"langchain\"","langchain-openai<2.0,>=1.1.14; extra == \"langchain\"","langchain-core<4.0,>=1.3.3; extra == \"langgraph\"","langchain-openai<2.0,>=1.1.14; extra == \"langgraph\"","langgraph<2.0,>=1.0; extra == \"langgraph\"","mcp<2.0.0,>=1.28.1; extra == \"mcp\"","httpx>=0.24.0; extra == \"mcp\"","starlette>=0.27.0; extra == \"mcp\"","uvicorn<1.0,>=0.23.0; extra == \"mcp\"","sqlite-vec>=0.1.6; extra == \"memory\"","sentence-transformers<7.0,>=2.2.0; (platform_machine != \"x86_64\" and extra == \"memory\") or (sys_platform != \"darwin\" and extra == \"memory\")","mem0ai<3.0,>=2.0.0; extra == \"memory-stack\"","qdrant-client<2.0,>=1.9.0; extra == \"memory-stack\"","neo4j<7.0,>=5.20.0; extra == \"memory-stack\"","torch>=2.12.1; (platform_machine != \"x86_64\" and extra == \"ml\") or (sys_platform != \"darwin\" and extra == \"ml\")","transformers<6.0,>=5.5.0; extra == \"ml\"","huggingface-hub<2.0,>=1.5.0; extra == \"ml\"","opentelemetry-sdk>=1.24.0; extra == \"otel\"","opentelemetry-exporter-otlp-proto-http>=1.24.0; extra == \"otel\"","fastapi>=0.100.0; extra == \"proxy\"","uvicorn<1.0,>=0.23.0; extra == \"proxy\"","orjson>=3.9.14; platform_python_implementation != \"PyPy\" and extra == \"proxy\"","httpx[http2]>=0.24.0; extra == \"proxy\"","openai>=2.14.0; extra == \"proxy\"","mcp<2.0.0,>=1.28.1; extra == \"proxy\"","magika>=0.6.0; extra == \"proxy\"","zstandard>=0.20.0; extra == \"proxy\"","websockets>=13.0; extra == \"proxy\"","onnxruntime>=1.24.0; python_full_version >= \"3.11\" and extra == \"proxy\"","onnxruntime<1.24.0,>=1.16.0; python_full_version < \"3.11\" and extra == \"proxy\"","transformers<6.0,>=5.5.0; extra == \"proxy\"","watchdog>=4.0.0; (python_full_version < \"3.14\" and extra == \"proxy\") or (sys_platform != \"darwin\" and extra == \"proxy\")","sqlite-vec>=0.1.6; extra == \"proxy\"","headroom-ai[proxy]; extra == \"proxy-prod\"","gunicorn>=21.0.0; sys_platform != \"win32\" and extra == \"proxy-prod\"","torch>=2.12.1; sys_platform == \"darwin\" and extra == \"pytorch-mps\"","sentence-transformers>=2.2.0; sys_platform == \"darwin\" and extra == \"pytorch-mps\"","fastembed>=0.4.0; extra == \"relevance\"","numpy>=1.24.0; extra == \"relevance\"","jinja2>=3.0.0; extra == \"reports\"","headroom-ai[code,html,mcp,otel,proxy,relevance,reports,spreadsheet]; extra == \"sandbox\"","openpyxl>=3.1.0; extra == \"spreadsheet\"","xlrd>=2.0.1; extra == \"spreadsheet\"","strands-agents>=0.1.0; extra == \"strands\"","hnswlib>=0.8.0; extra == \"vector\"","onnxruntime>=1.24.0; python_full_version >= \"3.11\" and extra == \"voice\"","onnxruntime<1.24.0,>=1.16.0; python_full_version < \"3.11\" and extra == \"voice\"","transformers<6.0,>=5.5.0; extra == \"voice\"","torch>=2.12.1; (platform_machine != \"x86_64\" and extra == \"voice\") or (sys_platform != \"darwin\" and extra == \"voice\")","headroom-ai[voice]; extra == \"voice-train\"","datasets>=5.0.1; extra == \"voice-train\"","accelerate>=0.20.0; extra == \"voice-train\""],"requires_python":">=3.10","summary":"The Context Optimization Layer for LLM Applications - Cut costs by 50-90%","version":"0.40.0","yanked":false,"yanked_reason":null},"last_serial":43435115,"ownership":{"organization":null,"roles":[{"role":"Owner","user":"chopratejas"},{"role":"Owner","user":"devanshivyas"}]},"urls":[{"comment_text":null,"core-metadata":{"sha256":"358932a259121318f7321e19b5d421983ef31ec47f264dfdefb004269f9291ee"},"digests":{"blake2b_256":"eec39d5bc6fb41ed8a0da89be842fe8b24bfc22d5bd5cb3663950822a91b8eb1","md5":"ccb560cc7e97400a369c99014c3ea4f2","sha256":"b7320996ea165bee1cbf250402b4d88dfd4b30abbe148664986e41b0ae09ca07"},"downloads":-1,"filename":"headroom_ai-0.40.0-cp310-abi3-macosx_10_12_x86_64.whl","has_sig":false,"md5_digest":"ccb560cc7e97400a369c99014c3ea4f2","packagetype":"bdist_wheel","python_version":"cp310","requires_python":">=3.10","size":15457946,"upload_time":"2026-10-06T03:15:49","upload_time_iso_8601":"2026-10-06T03:15:49.347418Z","url":"https:\/\/files.pythonhosted.org\/packages\/ee\/c3\/9d5bc6fb41ed8a0da89be842fe8b24bfc22d5bd5cb3663950822a91b8eb1\/headroom_ai-0.40.0-cp310-abi3-macosx_10_12_x86_64.whl","yanked":false,"yanked_reason":null},{"comment_text":null,"core-metadata":{"sha256":"358932a259121318f7321e19b5d421983ef31ec47f264dfdefb004269f9291ee"},"digests":{"blake2b_256":"f6cf2e7f237f343187bd4f1edba7f8117e6aa69953b46002a44f5136cc59f896","md5":"a0dc83a669e0804b8adaf5e82867118c","sha256":"f7b0186ad5e76d5c8f75e5de6c57e001637258ead75481575ccca4d57557f5fb"},"downloads":-1,"filename":"headroom_ai-0.40.0-cp310-abi3-macosx_11_0_arm64.whl","has_sig":false,"md5_digest":"a0dc83a669e0804b8adaf5e82867118c","packagetype":"bdist_wheel","python_version":"cp310","requires_python":">=3.10","size":15449719,"upload_time":"2026-10-06T03:15:52","upload_time_iso_8601":"2026-10-06T03:15:52.206081Z","url":"https:\/\/files.pythonhosted.org\/packages\/f6\/cf\/2e7f237f343187bd4f1edba7f8117e6aa69953b46002a44f5136cc59f896\/headroom_ai-0.40.0-cp310-abi3-macosx_11_0_arm64.whl","yanked":false,"yanked_reason":null},{"comment_text":null,"core-metadata":{"sha256":"358932a259121318f7321e19b5d421983ef31ec47f264dfdefb004269f9291ee"},"digests":{"blake2b_256":"c456c5d5ad7e9ace7206f2400e6dce84b98e3387f9ccfb61a882edbb406299dc","md5":"2d09a9fb2c64be3d18c81ca3990cf795","sha256":"140811066bfb7afe888f471be4a587375779af9de8dd76bb9a8126b52b8e930f"},"downloads":-1,"filename":"headroom_ai-0.40.0-cp310-abi3-manylinux_2_28_aarch64.whl","has_sig":false,"md5_digest":"2d09a9fb2c64be3d18c81ca3990cf795","packagetype":"bdist_wheel","python_version":"cp310","requires_python":">=3.10","size":15444416,"upload_time":"2026-10-06T03:15:54","upload_time_iso_8601":"2026-10-06T03:15:54.920123Z","url":"https:\/\/files.pythonhosted.org\/packages\/c4\/56\/c5d5ad7e9ace7206f2400e6dce84b98e3387f9ccfb61a882edbb406299dc\/headroom_ai-0.40.0-cp310-abi3-manylinux_2_28_aarch64.whl","yanked":false,"yanked_reason":null},{"comment_text":null,"core-metadata":{"sha256":"358932a259121318f7321e19b5d421983ef31ec47f264dfdefb004269f9291ee"},"digests":{"blake2b_256":"031c383f4a2d33ad18d40869037ca6221e1db24bc4da5e92befe70793220b138","md5":"f38d752cae6dc04cde57a1c2d3191e90","sha256":"49b6d8a01c8eceb5b9e9a675720ed5e4cb24e467282642653aac7f2fd8faf47f"},"downloads":-1,"filename":"headroom_ai-0.40.0-cp310-abi3-manylinux_2_28_x86_64.whl","has_sig":false,"md5_digest":"f38d752cae6dc04cde57a1c2d3191e90","packagetype":"bdist_wheel","python_version":"cp310","requires_python":">=3.10","size":15619942,"upload_time":"2026-10-06T03:15:57","upload_time_iso_8601":"2026-10-06T03:15:57.540323Z","url":"https:\/\/files.pythonhosted.org\/packages\/03\/1c\/383f4a2d33ad18d40869037ca6221e1db24bc4da5e92befe70793220b138\/headroom_ai-0.40.0-cp310-abi3-manylinux_2_28_x86_64.whl","yanked":false,"yanked_reason":null},{"comment_text":null,"core-metadata":{"sha256":"679a38e8c4f54c370d6357eab29e7868319d7c1e75e5543624754520e307162a"},"digests":{"blake2b_256":"97b9e14801dbcdf1d2dcf7f48675d49d6ed459445517177cf2ee3edae8062917","md5":"5f0c3e2cbdca051c5ca57f062ce02cee","sha256":"a803a60d90ec74806c02df1eb5d5e063af8a7aa3977c6aa7559693817490d193"},"downloads":-1,"filename":"headroom_ai-0.40.0-cp310-abi3-win_amd64.whl","has_sig":false,"md5_digest":"5f0c3e2cbdca051c5ca57f062ce02cee","packagetype":"bdist_wheel","python_version":"cp310","requires_python":">=3.10","size":15566384,"upload_time":"2026-10-06T03:16:00","upload_time_iso_8601":"2026-10-06T03:16:00.233394Z","url":"https:\/\/files.pythonhosted.org\/packages\/97\/b9\/e14801dbcdf1d2dcf7f48675d49d6ed459445517177cf2ee3edae8062917\/headroom_ai-0.40.0-cp310-abi3-win_amd64.whl","yanked":false,"yanked_reason":null},{"comment_text":null,"core-metadata":false,"digests":{"blake2b_256":"728aa3a878862b62b4bfc224dd6ea3b31be43b49513eafaf823f6a25288665b9","md5":"0421280c1c42eb30a22dcb4322aa88d4","sha256":"19403478224ae30e2599865de743051d0138f0e5d68d78418643c6f1f7892593"},"downloads":-1,"filename":"headroom_ai-0.40.0.tar.gz","has_sig":false,"md5_digest":"0421280c1c42eb30a22dcb4322aa88d4","packagetype":"sdist","python_version":"source","requires_python":">=3.10","size":3311116,"upload_time":"2026-10-06T03:16:02","upload_time_iso_8601":"2026-10-06T03:16:02.166419Z","url":"https:\/\/files.pythonhosted.org\/packages\/72\/8a\/a3a878862b62b4bfc224dd6ea3b31be43b49513eafaf823f6a25288665b9\/headroom_ai-0.40.0.tar.gz","yanked":false,"yanked_reason":null}],"vulnerabilities":[]}