Stop hunting for commands. Describe what you want to do.
EveryCli is a natural-language command-line assistant. You describe your intent — in English or French — and it retrieves the matching shell command from a local corpus. Search runs entirely on your machine: a native Rust daemon combines lexical matching with semantic reranking through ONNX Runtime, with no API key and no network call. The move from the historical Python/PyInstaller flow to this native Rust stack means end users no longer need Rust, Cargo or Python — a prebuilt archive is enough.
🛟 Safety first. EveryCli never runs a command on your behalf. Results are shown for you to read;
--runalways asks for confirmation, and shell wrappers only place the command in an editable buffer.
EveryCli grew out of a small daily friction. Whenever a CLI command slipped my mind, I'd ask an online LLM — and each lookup added two or three seconds of waiting. That delay sounds small, but it can be enough to break context: you switch to another tab, start another task, and later have to reconstruct what you were doing. Digital attention is not split without cost; rapid task switching increases the effort of resuming context and the risk of losing the thread 1.
The idea was to keep that loop inside the terminal — describe the intent, get a command quickly from a local corpus, and gradually grow your own set of commands with add / list / remove. The goal is not to claim that a compact local model is universally better than a large remote one. It is to accept a smaller, controlled scope when that reduces latency, preserves context and avoids a detour to another tab. When local coverage is not enough, everycli ask remains available as an optional LLM assistant. The objective is simple: less waiting, fewer context switches, and continued control over what actually runs 1.
In this setting, a model that is fast and available locally can be more useful than a heavier model that imposes a round-trip and breaks attention.
EveryCli separates the user interface from the more expensive semantic computation. The Rust client receives your intent and communicates with a local daemon that keeps the model in memory and answers searches quickly. This avoids reloading the model for every command and helps you stay in the terminal on both Ubuntu/Linux and Windows.
User
│
▼
everycli-rs ── JSON/TCP localhost ──▶ everycli-daemon
│ ├── YAML corpus
│ ├── model.onnx
│ ├── tokenizer.json
│ └── native ONNX Runtime
└── local lexical fallback if the daemon is unavailable
The daemon listens only on 127.0.0.1:51821; it is not a public network API. It combines the YAML corpus, ONNX model and platform-native runtime. The client also keeps a local lexical fallback so search remains useful when the daemon is unavailable.
📖 Architecture deep dive: components, daemon role, protocol, model, runtime, corpus and lifecycle → docs/explanation_architecture.md.
- 🔎 Natural-language search — describe what you want to do instead of memorizing syntax, using a corpus organized by namespace (Git, Docker, Compose, npm, SSH, Python, Linux…).
- 🧬 Hybrid local ranking — lexical matching and ONNX semantic reranking are fused on your machine to surface relevant commands without sending your query to the cloud.
- 🌍 Bilingual search — write your request in English, French or a mixed query, without mentally translating a command or its description.
- 🛰️ Fast local daemon — the model stays loaded on 127.0.0.1:51821, allowing repeated searches to answer in a few hundred milliseconds after the first load on tested hardware, on Ubuntu/Linux and Windows, so you can stay in the work rhythm instead of waiting for a remote request.
- ✍️ Add your own commands — with add,listandremove, you gradually build a personal corpus; its files remain separate from the built-in corpus and survive updates.
- 🤝 Optional AI assistance — when no local command matches sufficiently, everycli askcan request a proposal from an OpenAI-compatible API and add it to your corpus, so you do not need to repeat a remote lookup for the same need later.
- 🛡️ Review before execution — Sentinel (everycli plan) can review a retrieved command and flag risks before you decide whether to run it.
- 🐚 Shell integration — the interactive picker, --json,--copyand deterministic--shellprotocol support human and scriptable workflows without automatic execution.
- 📦 No development dependencies for end users — a prebuilt release includes the binaries, model, tokenizer, runtime and corpus; end users do not need Rust, Cargo or Python.
The simplest path is to run the installer directly from the main branch. It downloads the latest release, verifies its integrity and configures the client, daemon, model, runtime, corpus and background startup automatically. Rust, Cargo and Python are not required.
⚠️ Use the next corrected release, v1.2.1 or later. The publicv1.2.0release shipped an ONNX Runtime library that is too old for theortcrate version used by the daemon; it should not be presented as a working semantic installation.
curl -fsSL https://raw.githubusercontent.com/HE11032006/EveryCli/main/install.sh | bashThe installer asks for the language, places EveryCli in ~/.local/share/everycli, creates links in ~/.local/bin and configures the systemd --user service. At the end, reload your profile:
source ~/.profile
everycli search "how to undo my last commit"If everycli is not recognized yet, open a new terminal. The first start may take longer while the daemon loads the model and computes corpus embeddings; later searches benefit from the warm daemon and disk cache.
In PowerShell, run:
irm https://raw.githubusercontent.com/HE11032006/EveryCli/main/install.ps1 | iexThe installer asks for the language, downloads the release, verifies its integrity and configures EveryCli in your user profile. The first path can be somewhat slow, especially during the download and first model load. Open a new terminal if needed for the everycli command to become available.
To inspect files before installation, download the matching archive from GitHub Releases, extract it and run the installer without arguments from the extracted folder.
On Ubuntu/Linux:
mkdir everycli-linux-x86_64
tar -xzf everycli-linux-x86_64.tar.gz -C everycli-linux-x86_64
cd everycli-linux-x86_64
./install.sh
source ~/.profileOn Windows PowerShell:
Expand-Archive .\everycli-windows-x86_64.zip .\everycli-windows-x86_64
cd .\everycli-windows-x86_64
.\install.ps1The complete archive contains the client, daemon, model.onnx, tokenizer.json, the native ONNX Runtime library and the built-in corpus. To uninstall, use ./uninstall.sh on Linux or ./uninstall.ps1 on Windows; personal data is preserved by default.
macOS is compiled and tested by CI, but no installable macOS archive is published yet — the native runtime and installer still need end-to-end validation.
📖 Full guide: prerequisites, one-command installation, archive installation, checksum verification, uninstall and troubleshooting → docs/tutorial_installation.md.
The most useful everyday workflow is the interactive path:
everycli search "how to undo my last commit" --top 2 -i⭐ Recommended workflow.
--top 2limits the display to the two most relevant candidates and-iopens the interactive picker. After choosing, EveryCli shows the selected command instead of making you scan a longer result list. On Windows, the selected command is copied automatically to the clipboard, so you can paste it immediately withCtrl+V. On Linux, automatic copying from the interactive workflow is not implemented yet; it is planned for a future update.
The simple form remains available:
everycli search "describe your intent"Common options:
everycli search "query" --top 3 # maximum number of candidates
everycli search "query" -i # choose with the keyboard
everycli search "query" --copy # copy the selected result
everycli search "query" --run # run after confirmation
everycli search "query" --json # machine-readable output
everycli search "query" --no-daemon # force the local lexical fallbackInteractive mode lets you review candidates before choosing. EveryCli never runs a command without explicit confirmation.
Add and maintain your own commands:
everycli add
everycli list
everycli removeThis turns a command found once into a reusable local shortcut. Personal commands live in ~/.everycli/commands (Linux) or %USERPROFILE%\.everycli\commands (Windows), separate from the built-in corpus and preserved through a normal update or uninstall.
search is the main local path — corpus + Rust daemon, no API key. Two optional companions exist:
- everycli askcalls an OpenAI-compatible API to propose a command, explanation, warning and tags, then offers to save it to your personal corpus. It is not the local ranking of- search.
- Sentinel (everycli plan) is a separate Python planner that reviews an already-retrieved command. It never executes a shell command for you.
Configure a key (stored privately, 0600 on Unix):
everycli config set api_key "your-api-key"
everycli config show # never prints the full key
everycli ask "compress the current directory"The key may also come from EVERYCLI_API_KEY. Known prefixes are auto-detected, or set a provider explicitly (everycli config set provider openai, plus api_url / api_model for OpenAI-compatible endpoints). Without a key, ask reports a config error while search keeps working offline.
EveryCli is driven by a handful of environment variables and a ~/.everycli/config.toml file. The essentials:
Installers replace the dev defaults with absolute paths. A corpus entry is a YAML record with id, description, tags, per-platform commands, explanation and an optional warning.
📖 Full reference: every variable, data paths, corpus schema and the JSON daemon protocol → docs/reference_config.md.
EveryCli separates the human interface from the wrapper protocol so a command is never executed by surprise. The --shell (-s) mode prints only the resolved command to stdout (no trailing newline), sends diagnostics to stderr, and does not confirm or run anything — ideal for Bash/Zsh/PowerShell wrappers that place the command in an editable buffer.
everycli search "undo my last commit" --shellTo stay deterministic, --shell does not combine with -i, --run, --copy, --error or --top > 1.
📖 Full guide: interactive mode, Bash/Zsh/PowerShell wrappers, execution-safety rules → docs/shell_integration.md.
For a prebuilt release you don't need this — it targets developers building a local bundle.
# Rust binaries
cd rust
cargo build --release -p everycli-rs -p everycli-daemon
# Targeted tests
cargo test -p everycli-rs
cargo test -p everycli-core
cargo test -p everycli-daemonThe production ONNX artifact lives in Michelhe/everycli-minilm-ft-boosted-onnx; CI pins a revision and verifies SHA-256 checksums. The ort 2.0.0-rc.13 crate requires a native ONNX Runtime 1.27.x+ library (the workflow pins 1.27.0).
📖 Full guide: ONNX assets, native runtime, local staging, CI/release flow → docs/how_to_build.md.
Measurements are development observations, not universal guarantees. On the bilingual eval/confusion_set.yaml set the recorded ranking result was 58/66 queries (87.9%). An indicative Windows latency baseline: ~383 ms for a full daemon search vs ~33 ms for the local lexical fallback (a cold first load is much slower).
python tools/evaluate_confusion.py # hybrid path
python tools/evaluate_confusion.py --matcher lexical # lexical only📖 Full guide: measurement rules, bundle validation, CI checks, regression reports → docs/benchmarking.md.
See CHANGELOG.md for factual change history and measurements.
EveryCli/
├── everycli/ # Sentinel Python component and historical data
│ └── data/commands/ # Built-in YAML corpus
├── rust/
│ ├── everycli-core/ # Corpus loading and lexical search
│ ├── everycli-inference/ # Semantic encoder and ONNX Runtime
│ ├── everycli-daemon/ # Local TCP server
│ └── everycli-rs/ # Rust CLI client
├── docs/ # Standalone technical guides
├── scripts/ # Platform staging scripts
├── install.sh / install.ps1 # Linux / Windows installers
├── uninstall.sh / uninstall.ps1
└── .github/workflows/build.yml # Tests, builds and CI bundles
Code is released under the MIT License; documentation under a Creative Commons license. See LICENSE.md.