Scopeon – AI 관찰 가능성 – 토큰 분석, 캐시 ROI, 비용 추적, CI 게이트

◈ ╔═╗╔═╗╔═╗╔═╗╔═╗╔═╗╔╗╔ ╚═╗║ ║ ║╠═╝║╣ ║ ║║║║ ╚═╝╚═╝╚═╝╩ ╚═╝╚═╝╝╚╝ AI Context Observability for Claude Code & friends The AI context observatory — for every coding agent, every token, every dollar. Install · Quick Start · Docs · Contributing You fire up Claude Code and start building. An hour later: "Context window full." You have no idea what burned it — was it the MCP tools? The thinking budget? Yesterday's file edits? You're flying blind on a meter that costs real money. Scopeon gives you the instrument panel. | Without Scopeon | With Scopeon | |---|---| | "Why is this session so expensive?" | Turn-by-turn cost breakdown with waste signals | | "Is the prompt cache actually working?" | Hit-rate gauge, USD saved, optimization suggestions | | "How close am I to the context limit?" | Real-time fill bar + "~12 turns remaining" prediction | | "Which project costs the most?" | Per-project / per-branch cost breakdown | | "Did my optimization actually help?" | compare_sessions before/after diff | | "Can I gate AI cost in CI?" | scopeon ci report --fail-on-cost-delta 50 | | "How is my whole team using AI?" | scopeon team — per-author cost from git history, no server needed | | "Can teammates see my live metrics?" | scopeon serve — privacy-filtered API + SSE stream for IDEs | | "Can I see AI cost in Grafana / Datadog?" | Prometheus bridge or OTLP push — docs/opentelemetry.md | 🔬 X-ray vision into every token — see exactly what burned your context: input, cache reads/writes, thinking budget, output, MCP calls — broken down turn by turn. No more guessing. 💸 Know your real cost before the bill arrives — live USD per turn, per session, per project, per day. Set budgets with actual alerts, not surprises. ⚡ Prompt cache that actually tells you if it's working — hit-rate gauge, dollars saved vs. uncached baseline. Know in seconds whether your cache setup is doing anything. ⏳ "You have ~12 turns left" — Scopeon tracks context fill rate over time and tells you how many turns remain before the wall. Stop being blindsided mid-task. 🔔 Context advisory before the crisis — compaction advisory fires at 55–79% fill when context is accelerating, so you compact at the optimal moment — not after it's too late. 📡 Zero-token ambient awareness — every 30 s the MCP server pushes a free status heartbeat to the agent. No polling, no token spend, just continuous situational awareness. 🧭 See what the agent actually did — provenance-aware history shows which skills, MCPs, hooks, tasks, and subagents were involved, with exact vs. estimated support called out per provider. 🤖 Your AI agent monitors itself — 17 MCP tools let Claude Code query its own token stats, provenance history, trigger alerts, and compare sessions — without you doing anything. 🚦 Fail PRs on AI cost spikes — one command in CI, zero config. scopeon ci report --fail-on-cost-delta 50 catches runaway cost before it merges. 🌐 Live browser dashboard + IDE stream — scopeon serve → WebSocket-powered charts at localhost:7771 and a GET /sse/v1/status SSE feed for IDE extensions. No npm, no Node, just Rust. 👥 Team cost from git history — scopeon team reads AI-Cost: trailers from git log and prints a per-author cost table. No server, no cloud, works from any machine. 🐚 Cost follows you everywhere — in your shell prompt, in every git commit as an AI-Cost: trailer, in Slack via digest webhooks. 🔒 Fully local, forever — no cloud backend, no account, no telemetry. Your prompts never leave the machine. Ever. Sessions — full session list with per-turn cost, cache %, MCP call count Insights — health score, waste signals, cache ROI & optimization tips Budget — daily/weekly/monthly spend, by model, by project, 14-day chart Providers — auto-detected agents with data path & session stats cargo install cargo-binstall # one-time cargo binstall scopeon curl -fsSL https://raw.githubusercontent.com/sorunokoe/Scopeon/main/install.sh | sh cargo install --git https://github.com/sorunokoe/Scopeon Requirements: Rust 1.88+ · macOS 12+ or Linux (glibc 2.31+) · Windows 10+ Download from GitHub Releases: | Platform | Asset | |---|---| | macOS Apple Silicon | scopeon-vX.Y.Z-aarch64-apple-darwin.tar.gz | | macOS Intel | scopeon-vX.Y.Z-x86_64-apple-darwin.tar.gz | | Linux x86_64 | scopeon-vX.Y.Z-x86_64-unknown-linux-gnu.tar.gz | | Linux ARM64 | scopeon-vX.Y.Z-aarch64-unknown-linux-gnu.tar.gz | | Windows x86_64 | scopeon-vX.Y.Z-x86_64-pc-windows-msvc.zip | scopeon onboard # auto-detect AI tools, configure MCP + shell integration scopeon # open the TUI dashboard scopeon serve # browser dashboard → http://localhost:7771 scopeon status # quick inline stats, no TUI scopeon doctor # health diagnostics scopeon init # → writes MCP server config to ~/.claude/settings.json # Claude Code now has 17 Scopeon tools + proactive push alerts scopeon [start] # daemon + TUI (default) scopeon tui # TUI only (no file watching) scopeon mcp # MCP server over stdio scopeon status # inline stats scopeon serve [--port N] [--tier 0-3] scopeon tag set feature # tag sessions for cost attribution scopeon export --format csv --days 30 scopeon reprice # recalculate costs after a price change scopeon digest [--days N] [--post-to-slack ] scopeon badge [--format markdown|url|html] scopeon ci snapshot --output baseline.json scopeon ci report --baseline baseline.json [--fail-on-cost-delta 50] scopeon shell-hook # emit shell prompt hook (bash/zsh/fish) scopeon git-hook install # add AI-Cost trailer to commits scopeon onboard # interactive setup wizard scopeon doctor # health diagnostics | Topic | Link | |---|---| | Full feature list | docs/features.md | | TUI guide (tabs, shortcuts, Zen, Replay, filter) | docs/tui.md | | MCP tools & push notifications | docs/mcp.md | | Webhook escalation | docs/webhooks.md | | CI cost gate | docs/ci.md | | Shell & git integration | docs/shell-git.md | | Team mode & REST API | docs/team.md | | OpenTelemetry integration | docs/opentelemetry.md | | Supported providers | docs/providers.md | | Configuration reference | docs/configuration.md | | Architecture & codebase map | ARCHITECTURE.md | Claude Code · GitHub Copilot CLI · Aider · Cursor · Gemini CLI · Ollama · Generic OpenAI Scopeon discovers log files automatically — no config needed for standard install paths. Adding a new provider takes ~50 lines of Rust. See docs/providers.md. - Local-first — no cloud backend, no accounts, no API keys required - Reads token counts and costs only — never prompt text or code scopeon serve is read-only and localhost-bound by default- Webhooks are opt-in; payloads contain only metric data Contributions are warmly welcomed — bug fixes, new providers, dashboard features. git clone https://github.com/sorunokoe/Scopeon cd Scopeon make # fmt-check + clippy + test (same as CI) make install # install to ~/.cargo/bin - CONTRIBUTING.md — dev workflow, PR process, adding providers - ARCHITECTURE.md — codebase map: crate roles, data flow, schema Every PR must pass the full CI suite (fmt · clippy · tests on Linux/macOS/Windows · MSRV · docs). See CHANGELOG.md for the full version history. Dual-licensed under MIT OR Apache-2.0 — use it however you like.