Lib.rs

Token-saving companion for Claude Code, Gemini CLI, Qwen Code, and Pi. Built on a "set it and forget it!" philosophy: one install command, zero configuration, and ecotokens works automatically from there — intercepting tool outputs before they reach the model, filtering the noise, and recording how many tokens you saved.

Features highlight

How it works

ecotokens installs hooks that intercept tool outputs before they reach the model. Two interception points are supported:

PreToolUse / BeforeTool — fires before every shell (Bash) command:

1. Runs the command and captures its output
2. Applies a family-specific filter (git, cargo, python, …)
3. Optionally summarizes large outputs via a local AI model (Ollama)
4. Returns the compressed output to the model
5. Records the before/after token counts in a local metrics store

PostToolUse / AfterTool (Claude Code, Gemini CLI, Qwen Code) — fires after native file-tool calls:

1. Intercepts the tool result before it enters the context window
2. Applies a specialized filter (outline for source files, grep result trimming, glob path denoising)
3. Returns the compressed result to the model
4. Records the savings under the native_read, grep, or fs family

Claude Code uses the PreToolUse + PostToolUse hooks (~/.claude/settings.json). Gemini CLI uses the BeforeTool + AfterTool hooks (~/.gemini/settings.json). Qwen Code uses the PreToolUse + PostToolUse hooks (~/.qwen/settings.json). Pi uses a TypeScript extension (~/.pi/agent/extensions/ecotokens.ts) that intercepts tool_call (bash pre-exec) and tool_result (read/grep/find/ls post-exec) events in-process.

For a focused view of the runtime path, see docs/hook-filter-metrics-flow.md.

The result: the model sees clean, concise output — and you keep your context window.

Quick install

cargo install --git https://github.com/hansipie/ecotokens

For exact token counting (tiktoken cl100k_base instead of the character heuristic):

cargo install --git https://github.com/hansipie/ecotokens --features exact-tokens

Build from source

git clone https://github.com/hansipie/ecotokens.git
cd ecotokens
cargo build --release
./target/release/ecotokens --help

To install the locally built binary into Cargo's bin directory:

cargo install --path .

With exact token counting enabled via tiktoken (cl100k_base encoding):

cargo install --path . --features exact-tokens

By default, token counts use a fast character heuristic (chars × 0.25, ~80-85% accuracy). This has no effect on filtering behavior — only the token counts recorded in metrics are more precise.

Installation

Claude Code

cargo install --path .
ecotokens install

In addition to hook installation, this also registers an MCP server entry in ~/.claude/settings.json:

{
  "mcpServers": {
    "ecotokens": {
      "command": "ecotokens",
      "args": ["mcp-server"]
    }
  }
}

Gemini CLI

Requires Gemini CLI ≥ 0.1.0.

cargo install --path .
ecotokens install --target gemini

This writes BeforeTool and AfterTool hook entries into ~/.gemini/settings.json. The AfterTool hook intercepts read_file, search_file_content, and list_directory results.

It also registers the ecotokens MCP server in ~/.gemini/settings.json.

Qwen Code

Requires Qwen Code.

cargo install --path .
ecotokens install --target qwen

This writes PreToolUse and PostToolUse hook entries into ~/.qwen/settings.json. The PostToolUse hook intercepts read_file, search_files, and list_dir results.

It also registers the ecotokens MCP server in ~/.qwen/settings.json.

Pi

Requires Pi (@mariozechner/pi-coding-agent ≥ 0.62.0).

cargo install --path .
ecotokens install --target pi

This writes a TypeScript extension to ~/.pi/agent/extensions/ecotokens.ts. Pi auto-discovers it on next startup (or /reload inside an active session). The extension intercepts bash commands before execution and filters native tool results (read, grep, find, ls) after execution.

All targets at once

ecotokens install --target all

--target all covers Claude Code, Gemini CLI, Qwen Code, and Pi in a single command.

With AI summarization

Enable AI-powered output compression via Ollama at install time:

ecotokens install --ai-summary                          # use default model (llama3.2:3b)
ecotokens install --ai-summary-model qwen2.5:3b         # specify model (implies --ai-summary)

This writes ai_summary_enabled and ai_summary_model to ~/.config/ecotokens/config.json. Ollama must be running and the model must be pulled (ollama pull llama3.2:3b).

Uninstall

ecotokens uninstall                    # Claude Code
ecotokens uninstall --target gemini    # Gemini CLI
ecotokens uninstall --target qwen      # Qwen Code
ecotokens uninstall --target pi        # Pi
ecotokens uninstall --target all       # all targets

Commands

Shell completions

# zsh
ecotokens completions zsh > ~/.zsh/completions/_ecotokens

# bash
ecotokens completions bash > ~/.local/share/bash-completion/completions/ecotokens

# fish
ecotokens completions fish > ~/.config/fish/completions/ecotokens.fish

# PowerShell
ecotokens completions powershell >> $PROFILE

Reload your shell (or open a new terminal) to activate completions.

Gain dashboard

ecotokens gain                                          # all time, uses default model from config
ecotokens gain --period today                           # today only
ecotokens gain --period week                            # last 7 days
ecotokens gain --period month --model claude-sonnet-4-6 # last 30 days, override model
ecotokens gain --history                                # summary table: 24h / 7d / 30d
ecotokens gain --history --json                         # same, as JSON

The model used for cost calculations defaults to the value set with ecotokens config --model (or claude-sonnet-4-6 if not configured). Pass --model to override for a single invocation.

Interactive TUI showing token savings per command family and per project, with a sparkline. The --period flag filters both the stats and the history panels.

Keybindings:

Filter command

ecotokens filter runs a command directly and returns its filtered output. Useful for testing filters or wrapping commands in scripts:

ecotokens filter -- cargo test
ecotokens filter --debug -- git log --oneline -50
ecotokens filter --cwd /path/to/project -- cargo test

The output is compressed by the same family-specific filters used by the hook, and token savings are recorded in the metrics store.

Watch command

ecotokens watch monitors a directory and automatically re-indexes files as they change.

ecotokens watch                    # foreground, TUI progress
ecotokens watch --path ./src       # watch a specific directory
ecotokens watch --background       # fork to background
ecotokens watch --status           # show status of background process
ecotokens watch --status --json    # JSON status output
ecotokens watch --stop             # stop the background process

Note: Background logs are only written if global debug is enabled (ecotokens config --debug true).

Auto-watch (Claude Code & Qwen Code)

ecotokens auto-watch integrates with Claude Code and Qwen Code's session lifecycle to start and stop the watcher automatically.

ecotokens auto-watch enable    # enable auto-watch, install SessionStart/SessionEnd hooks
ecotokens auto-watch disable   # disable (hooks remain installed but are no-ops)

When enabled, ecotokens watch --background starts automatically when a session opens, and stops when it closes. The setting is stored in ~/.config/ecotokens/config.json (auto_watch: true/false).

Note: Auto-watch relies on SessionStart / SessionEnd hooks. For Qwen Code, session hooks are installed automatically if ecotokens is already installed for Qwen (ecotokens install --target qwen). Gemini CLI does not expose session lifecycle hooks.

Word abbreviations

ecotokens abbreviations enable    # transform narrative text + inject model instruction
ecotokens abbreviations list      # show the active dictionary
ecotokens abbreviations disable   # back to default

When enabled, a post-processing pass replaces full words with shorter forms in the narrative parts of tool outputs (code blocks between triple backticks are preserved). A matching additionalContext payload is emitted at SessionStart so the model adopts the same abbreviations in its own responses.

See the full list of default abbreviations in docs/abbreviations.md.

Keep the feature flag in ~/.config/ecotokens/config.json

{
  "abbreviations_enabled": true
}

... and put custom pairs in a separate ~/.config/ecotokens/abbreviations.json file:

{
  "function": "func",
  "repository": "repo"
}

Bonus Tools

MCP server (Claude Code, Gemini CLI, Qwen Code)

ecotokens mcp-server starts a stdio MCP server backed by the ecotokens index and trace engines.

ecotokens mcp-server
ecotokens mcp-server --index-dir ~/.config/ecotokens/index

Exposed tools:

• ecotokens_search — BM25 + semantic search
• ecotokens_outline — symbol outline for file/directory
• ecotokens_symbol — fetch full symbol source by stable ID
• ecotokens_trace_callers — find callers of a symbol
• ecotokens_trace_callees — find callees (with depth)
• ecotokens_duplicates — detect near-duplicate code blocks

For Claude Code, Gemini CLI, and Qwen Code, ecotokens install registers this server automatically in each target's settings file.

Search command

ecotokens search QUERY performs BM25 (+ optional semantic) search over the indexed codebase and returns results anchored to the matching line.

ecotokens search "embed_text"                        # top 5 results, 2 lines of context
ecotokens search "embed_text" --context 4            # 4 lines above and below the match
ecotokens search "error" --include "*.rs"            # Rust files only
ecotokens search "TODO" --exclude "*.md" --exclude "*.toml"
ecotokens search "find_callers" --no-trace           # pure BM25, no trace augmentation
ecotokens search "find_callers" --json               # JSON output with callers array
ecotokens search "query" --top-k 10                  # more results

Output format:

src/search/query.rs:29 (score: 11.068)
  27:  
  28:  pub fn search_index(opts: SearchOptions) -> tantivy::Result<Vec<SearchResult>> {
  29:      let index = Index::open_in_dir(&opts.index_dir)?;
  30:      let (_, file_path_field, content_field, kind_field, line_start_field, _) = build_schema();
  31:  

When the query matches a symbol name, callers are automatically appended:

# Symbol match — call sites via trace
  src/main.rs:1301 [caller]  cmd_search

Results are automatically scoped to the current git project when using the global index — files from other indexed projects are silently filtered out.

Duplicates command

Less code is less tokens

ecotokens duplicates scans the indexed codebase for near-identical code blocks and reports them grouped by similarity.

ecotokens duplicates                          # default: threshold=70%, min_lines=5
ecotokens duplicates --threshold 80           # only report ≥ 80% similarity
ecotokens duplicates --min-lines 10           # ignore blocks shorter than 10 lines
ecotokens duplicates --json                   # JSON output

Each group shows the file paths, line ranges, similarity score, and a refactoring proposal (exact duplicate, near duplicate, or subset).

Configuration

ecotokens config           # show all settings (text)
ecotokens config --json    # show all settings (JSON)

Output includes:

hook_installed        : true
debug                 : false
debuglog              : false
default_model         : claude-sonnet-4-6
exclusions            : []
embed_provider        : candle (sentence-transformers/all-MiniLM-L6-v2)
ai_summary_enabled    : false
ai_summary_model      : llama3.2:3b (default)
ai_summary_url        : http://localhost:11434 (default)
abbreviations_enabled : false

Debug mode

Enable the global debug mode to see detailed interception logs and enable background logging for the watch command:

ecotokens config --debug true
ecotokens config --debug false

This updates the debug field in ~/.config/ecotokens/config.json.

Debug file logging

Enable structured per-hook logging to a file for deeper tracing of what ecotokens intercepts:

ecotokens config --debuglog true
ecotokens config --debuglog false

When enabled, every hook invocation appends a JSONL entry to ~/.config/ecotokens/debug.log:

{"ts":"2026-05-08T12:00:00Z","uid":"a1b2c3d4","cmd":"git status","phase":"input","data":{...}}
{"ts":"2026-05-08T12:00:00Z","uid":"a1b2c3d4","cmd":"git status","phase":"output","data":{...}}

Each entry contains a short uid to correlate the input and output phases of the same invocation. Distinct from --debug (which prints to stderr) — --debuglog writes silently to disk and survives across sessions.

Default model for cost calculations

The model selected here determines the per-token price used in gain reports:

ecotokens config --model claude-opus-4-7    # set default model
ecotokens config --model ""                 # list available models
ecotokens config --model unknown-model      # unknown model → lists available models

The model name must be present in the built-in pricing table (or overridden via model_pricing in ~/.config/ecotokens/config.json). Passing an empty value or an unrecognised name prints the full list and exits.

See the full list of built-in models and prices in docs/models.md.

Override any entry or add a new model via model_pricing in ~/.config/ecotokens/config.json:

{
  "model_pricing": {
    "my-custom-model": { "input_usd_per_1m": 0.50, "output_usd_per_1m": 2.00 }
  }
}

Word abbreviations (optional)

ecotokens abbreviations enable    # transform narrative text + inject model instruction
ecotokens abbreviations list      # show the active dictionary
ecotokens abbreviations disable   # back to default

When enabled, a post-processing pass replaces full words with shorter forms in the narrative parts of tool outputs (code blocks between triple backticks are preserved). A matching additionalContext payload is emitted at SessionStart so the model adopts the same abbreviations in its own responses.

Extend or override the built-in dictionary via a separate ~/.config/ecotokens/abbreviations.json file:

{
  "function": "func",
  "repository": "repo"
}

The feature flag stays in ~/.config/ecotokens/config.json:

{
  "abbreviations_enabled": true
}

Supported command families

Note: Family detection uses the basename of the first token, so commands invoked via absolute path (/usr/bin/git), venv (.venv/bin/pytest), version managers (~/.cargo/bin/cargo), or wrappers (poetry run) are correctly matched to their family.

Embeddings

ecotokens search uses dual BM25 + vector retrieval with score fusion (0.4 × BM25 + 0.6 × cosine). The vector index is powered by Candle — a zero-config local embedding engine. No external service required.

Provider

Candle (default) — runs sentence-transformers/all-MiniLM-L6-v2 (384 dim) locally. The model is downloaded automatically from HuggingFace Hub on first use (~90 MB, cached in ~/.cache/huggingface/).

# Candle is active by default — nothing to configure
ecotokens index --path /your/project
ecotokens search "your query"

Each result includes a retrieval_source field (bm25, vector, or both) visible in JSON output.

Disable embeddings

ecotokens config --embed-provider none    # fall back to pure BM25

Workflow

# 1. Index your project (Candle embeddings computed automatically)
ecotokens index --path /your/project

# 2. Search with hybrid scoring
ecotokens search "your query"

# 3. JSON output with retrieval_source
ecotokens search "your query" --json

Model change detection

When the configured embedding model changes, ecotokens index automatically rebuilds the vector index (hnsw_index.bin) without touching the BM25 index. Embeddings for unchanged files are reused between runs.

AI summarization (optional)

When enabled, large command outputs (> ~2500 tokens) are summarized by a local Ollama model instead of being truncated. Falls back to generic filtering if Ollama is unavailable or times out.

Enable via install:

ecotokens install --ai-summary-model llama3.2:3b

Or update the config file directly (~/.config/ecotokens/config.json):

{
  "ai_summary_enabled": true,
  "ai_summary_model": "llama3.2:3b"
}

Ollama must be running locally. The model is called with a 3-second timeout to avoid blocking the model.

Benchmarks

Measured over 13 days on a real developer workstation (4 129 hook executions):

→ Full benchmark report

Precision Guarantees

Filtering is aggressive on noise, conservative on signal:

• Short outputs are never modified — outputs under 200 lines or 50 KB pass through unchanged
• Errors are always preserved — error[, FAILED, E (pytest), --- FAIL: (Go), stack traces and panic messages are never removed
• Failure sections are fully kept — structured blocks (=== FAILURES ===, failures:, failure diffs) are always passed through in their entirety
• Conservative fallback — if a family filter doesn't improve the output (filtered ≥ original), the original is returned as-is
• Secrets are redacted before filtering — 33 patterns covering cloud keys, AI APIs, VCS tokens, payment secrets and more are detected and replaced before any content reaches the model. See docs/secret-patterns.md for the full list.
• UTF-8 safe truncation — truncation always happens at character boundaries, never mid-codepoint
• Head + tail preservation — when generic truncation applies, the first and last 20 lines are always kept (start context + end result)

Requirements

• Rust ≥ 1.75 (stable)
• One or more of: Claude Code (with hook support), Gemini CLI ≥ 0.1.0, Qwen Code, Pi ≥ 0.62.0
• Internet access on first use (Candle downloads all-MiniLM-L6-v2 ~90 MB from HuggingFace Hub; cached locally after that)
• Ollama (optional, for AI summarization only)

Contributing

Contributions are welcome! Please read the contributing guidelines before submitting a pull request.

License

MIT