Tare

Warning

Tare's dollar figures are transparent estimates, not provider invoices, and interfaces may change between versions.

A local-first cost profiler for AI agents. Tare shows which prompt component, retry loop, uncached context, or model choice consumed the money—and what you can change—without sending your prompts to a hosted analytics service.

What Tare does

• Captures provider-reported usage through local Claude Code history, OpenTelemetry, an optional loopback proxy, or a self-hosted-model metrics agent.
• Attributes tokens and estimated cost to run → step → prompt component → cache class. Counts come from provider usage; Tare does not re-tokenize payloads.
• Explains spend through a cost-weighted flamegraph, timelines, comparisons, anomaly causes, and prompt/configuration lineages.
• Turns findings into a ranked savings plan with cache advice, model what-if scenarios, budgets, regression gates, and before/after verification.
• Recomputes receipts offline with versioned pricing, so an estimate can be checked without an account or network request.

Tare is designed for an individual developer analyzing their own local agent usage. It is not a team analytics SaaS, customer-billing meter, production tracing platform, or authoritative invoice.

Product surfaces

Every surface reads the same local SQLite store and uses the same Rust accounting engine.

Surface	Primary use
Pulse	Monitor current spend, budget pace, capture health, anomalies, and active sessions
Investigate	Explore runs, steps, sessions, time, prompt templates, comparisons, and profiles
Optimize	Rank savings opportunities, test offline scenarios, and verify applied changes
CLI	Capture commands, reports, exports, automation, and CI gates
MCP server	Let an agent query its own recorded spend

Capture, Trust & Pricing, and Settings are utility sheets in the browser and desktop workbench.

Quick start

# Install the CLI from this checkout.
cargo install --locked --path tare-cli

# Start the local capture service and UI.
tare up

# Confirm the store, pricing, receiver, capture, and agent wiring.
tare doctor

# Inspect captured spend.
tare report --today
tare report

Interactive bare tare is an alias for tare up. The service and browser UI bind to loopback by default; the UI is available at http://127.0.0.1:8788/__tare/.

Choose the capture lane

Lane	Setup	Fidelity
Claude Code JSONL	None; enabled by default	Counts, model, timing, sessions; no prompt text
OpenTelemetry	tare connect, codex-connect, or gemini-connect	Live out-of-band usage without entering the request path
Loopback proxy	Run a command with tare run -- …	Deep prompt-component and cache attribution
Self-hosted agent	tare agent --scrape … --hub …	Token deltas from vLLM/llama.cpp-style metrics

The connection commands are additive and reversible:

tare connect                 # Claude Code
tare codex-connect           # Codex CLI
tare gemini-connect          # Gemini CLI
tare disconnect
tare codex-disconnect
tare gemini-disconnect

For deep attribution on one command, route only that process through Tare's local proxy:

tare run -- python agent.py

See Running the personal observability mesh for service mode, manual OpenTelemetry configuration, self-hosted models, and local pricing overlays.

Common workflows

# Find and prioritize changes.
tare report
tare plan
tare savings
tare advise

# Compare or model alternatives without another provider call.
tare diff --run before-run --run after-run
tare whatif --recommend --cross-provider
tare trend --anomalies --why

# Export or independently verify a run.
tare export --run RUN_ID
tare attest --run RUN_ID --out receipt.json
tare verify receipt.json

# Enforce a deterministic spend-and-shape gate in CI.
tare gate --max-spend 0.50 --baseline-run BASELINE --fail-on-regression

tare --help lists the full command surface, including rollups, sessions, cache economics, lineage, work units, quality signals, reconciliation, heatmaps, receipts, and capture services. The repository also contains a composite GitHub Action in action.yml.

How capture fits together

Claude JSONL ─┐
OTLP events ──┼──▶ local capture service ──▶ SQLite ──▶ CLI / web / desktop / MCP
proxy wire ───┤              │
local agent ──┘              └──▶ attribution + effective-dated pricing

The OpenTelemetry and history lanes stay out of the model request path. The optional proxy is plain HTTP on loopback, not a TLS-intercepting MITM, and is used when request-body structure is required for component-level attribution.

Proxy parsing covers Anthropic Messages, OpenAI-compatible chat completions, Azure OpenAI, Bedrock Converse, Gemini, and local-model formats. Anthropic and OpenAI proxy fixtures contain real captured wire bytes; the other proxy parsers are schema-tested where real capture was unavailable. See fixture provenance for the exact boundary.

Privacy and estimate semantics

• Tare is local-first, account-free, and counts-only by default.
• Authentication headers are forwarded unchanged and never persisted.
• Prompt and response bodies are not stored under the default privacy profiles.
• The opt-in max_inspect profile stores scrubbed bodies in a separate local store. Scrubbing is best-effort, so treat that store as sensitive.
• Unpriced usage is reported as unpriced rather than silently presented as $0.
• Cost uses integer micro-USD and an effective-dated pricing edition; it remains an estimate rather than a provider invoice.

Read SECURITY.md before enabling payload capture or exposing any local service beyond loopback.

Build from source

Prerequisites:

• Rust 1.80 or newer
• Node.js 20 or newer for the web workbench
• Platform WebView libraries only when building the Tauri GUI

cargo build --release -p tare-cli

cd web
npm ci
npm run typecheck
npm test
cd ..

The CLI binary is written to target/release/tare. Build and run the desktop shell with:

cargo run -p tare-tauri --features gui --profile dev-fast

The complete repository gate checks formatting, strict Clippy, Rust tests, frontend tests, Playwright journeys, contrast, generated assets, and privacy invariants:

cargo fetch
bash scripts/ci.sh

scripts/ci.sh deliberately runs dependency use offline after caches are populated. Use TARE_SKIP_ASSET_GIT_DRIFT=1 bash scripts/ci.sh in a dirty worktree to compare generated UI assets byte-for-byte without requiring them to be staged. Regenerate the two tracked embedded UI trees with bash scripts/build-ui.sh.

Linux desktop builds additionally require WebKitGTK 4.1, GTK 3, libsoup 3, and libayatana-appindicator. macOS and Windows provide their platform WebViews.

Maintained documentation

Document	Purpose
Mesh usage	Capture services, agent wiring, and self-hosted models
Design guide	Visual, responsive, accessibility, and component rules
Route contract	Canonical browser/desktop deep links
Desktop testing	Browser twin, IPC seam, and manual native checks
Build performance	Fast desktop iteration and build-profile rationale
Release guide	Desktop packaging, signing, and release verification
App imagery	Reproducing screenshots, the hero, and the optional GIF

Contribution and conduct guidance lives in CONTRIBUTING.md and CODE_OF_CONDUCT.md.

Screenshots

These are the real fixture-backed workbench, rendered offline.

Pulse. Current spend, budget pace, forecast, and attention queue.

Investigate. Why a run cost what it did, down to steps and cache classes.

Optimize. Evidence-backed changes and before/after verification.

Regenerate screenshots with bash scripts/gen-ui-shots.sh; see App imagery for details.

License

MIT — see LICENSE.