Observer MCPサーバー

Try it in one command

npx @superbased/observer

No install, no config, no network. It reads the session logs your coding agents already keep on disk, prints a 30-day cost table by tool and model, and deletes its temporary database on exit.

On Windows PowerShell, type npx.cmd and npm.cmd instead of npx and npm (or use Command Prompt): PowerShell's default execution policy blocks the npx.ps1 and npm.ps1 scripts.

Why SuperBased

One local binary that tracks Claude Code token usage, Codex cost and Cursor spend side by side, and helps you control them.

• Provider-exact cost, not estimates. An optional local proxy records the tokens your provider returned: net input, 5-minute vs 1-hour cache writes, reasoning tokens and long-context repricing. Log-only trackers have to estimate those.
• Every coding agent in one place. 50 adapters normalize Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, Cline, Aider, Kiro, Devin, OpenCode, Junie and more into one local database.
• Local-first, no account. No sign-in. Everything lives in one SQLite file under ~/.observer/, and your session content is never uploaded without your explicit consent.
• Know the next bill before it lands. Prompt-cache tracking with expiry warnings, a low / typical / high cost band for your next message, and a 5-hour and weekly limit gauge for proxied sessions.
• Guardrails on your own agents. The guard checks each agent action against built-in rules (destructive commands, secrets egress, MCP pinning, budgets) and your own. It starts in observe mode; observer guard simulate replays your history before you turn on enforcement.
• Your agents as terminals, from anywhere. Launch agent CLIs from the dashboard as real terminals, join a running session from a paired device over your Tailscale tailnet, and hand a session from one tool to another.

How it compares

	SuperBased	ccusage	CodeBurn	Langfuse	Helicone	LiteLLM
Built for	Coding-agent cost & control	CC/Codex cost meter	Local CC spend + budget toolkit	LLM-app observability	LLM-app observability + gateway	AI gateway
Coding tools covered	50 adapters	Claude Code, Codex, + companion pkgs	31-36 per their own docs	Instrument-per-app	Instrument-per-app	Anything routed through it
Provider-reported (not estimated) tokens	✓ via local proxy	Partial (from logs)	Partial (from logs)	✓ (SDK)	✓ (proxy)	✓ (proxy)
Cache 5-min vs 1-hour write split	✓	✗	✗	✗	✗	✗
Reasoning + long-context surcharge accounting	✓	Partial	Partial	Partial	Partial	Partial
Runs 100% local, no cloud required	✓	✓	✓ (opt-in team-sync preview)	Self-host a stack	Self-host or SaaS	Self-host
Agent-queryable memory over your history (MCP)	✓	✓ (usage MCP)	✓	✓ (trace MCP)	✗	✗
Cross-tool session handoff	✓	✗	✗	✗	✗	✗
License	Apache-2.0	MIT	MIT	MIT (core)	Apache-2.0	MIT

Checked against each project's public repo (July 2026). Each of these tools is good at what it was built for; the full comparison covers more tools and when to pick which.

Install

Channel	Command
npm	npm install -g @superbased/observer
PyPI	uv tool install superbased-observer (or pipx install / pip install)
VS Code	code --install-extension superbased.superbased-observer (Cursor, VSCodium and Windsurf: Open VSX)
Installer	curl -fsSLO https://github.com/superbasedapp/observer/releases/latest/download/install.sh && sh install.sh

Every channel ships the same prebuilt, pure-Go binary for Linux (x64, arm64), macOS (Intel, Apple Silicon) and Windows x64. The installer (Windows: install.ps1) also offers smaller editions (minimal, developer, managed), checks each archive against SHA256SUMS and its Ed25519 signature, never runs sudo, and supports offline installs with --from-dir. Run sh install.sh --help for every option. On Windows PowerShell, type npx.cmd and npm.cmd instead of npx and npm (or use Command Prompt): PowerShell's default execution policy blocks the npx.ps1 and npm.ps1 scripts.

macOS needs version 13 Ventura or later. Observer is built with Go 1.27, which no longer runs on macOS 12 Monterey. On macOS 12, stay on v1.36.0 or upgrade macOS first. v1.36.0 was meant to carry an advance notice of this and did not. Linux and Windows are not affected.

Get started

observer start    # proxy + watcher + dashboard; opens http://localhost:8081
observer scan     # backfill history from the session logs already on disk

1. Route your agent through the proxy for provider-exact tokens. In the dashboard, open Compression, click your tool's status pill and choose Route through the observer proxy: it previews the exact config change and writes only when you confirm. One-off alternatives: observer claude / observer codex, or ANTHROPIC_BASE_URL=http://localhost:8820.
2. Use your agent as normal. Sessions, cost and cache numbers appear within minutes.
3. Optional: observer init also registers the MCP server so your agents can query their own history (it adds tool schemas to every turn, so it stays opt-in).

observer doctor checks the install; observer status shows what has been captured.

What you get

• Dashboard (localhost:8081): Overview, Live, Sessions, Cost, Plans & limits, Cache, Security, Settings and more, filterable by tool and project, and readable on a phone.
• Cost you can reconcile: per session, model, project and tool, with cache read/write splits and a session drill-down down to each API turn.
• Cross-tool memory: an opt-in MCP server so an agent can ask what another agent already read, ran or broke in the same repo, plus observer <tool> --continue-from <session> handoffs.
• Waste detection: stale re-reads, repeated commands and files several agents fight over.
• Terminals: daemon-owned PTY terminals with multi-seat join, a live cost cockpit and native resume after a restart.
• Your editor and shell: a VS Code extension (status bar, sidebar, dashboard tab), a one-line observer statusline, and a CLI for everything (observer usage, cost, predict, tail).

Supported agents include Claude Code, Codex, Cursor, Copilot and Copilot CLI, Gemini CLI, Cline and Cline CLI, Antigravity, OpenCode, OpenClaw, Pi, Hermes, Kilo Code, Qwen Code, Kiro CLI, Kiro Crew, Crush, Kimi Code, Grok, Devin, Aider, Qoder, goose, droid, Open Interpreter, Command Code, Junie, Poolside, Zed, Mistral Code, Muse, Prime Agent, DeepSeek, zcode, Freebuff, Cowork, Grok Bot, Autohand Code, IBM Bob, Cortex Code, CodeBuddy Code, Deep Agents Code, DimCode, Dirac, Kimchi, Stakpak and VT Code, plus ChatGPT, Claude, Gemini, Copilot and Perplexity on the web through an optional browser extension. Per-tool setup: superbased.app/docs/track.

Inside a session

Click any session to open its detail view. Four of its tabs:

Privacy

SuperBased needs no account. Today the default install sends us no usage statistics, analytics or crash reports, and it never uploads your session content without your explicit consent. It stores paths, commands and short excerpts, scrubs secrets before anything is written, and binds the proxy and dashboard to 127.0.0.1 by default. The proxy forwards your own requests, unchanged, to the provider you already use. By default it also makes a few small requests of its own that carry none of your data, such as a daily download of the public model price table from superbased.app (turn it off with [pricing.feed] enabled = false). Remove everything with observer uninstall --all --purge. Details: PRIVACY.md.

Documentation

• Quickstart and install options
• Connect your agents and track costs by tool
• Guides: dashboard, token and cost tracking, cache tracking, cost predictor, terminals, security guard, MCP server
• Reference: CLI, configuration, measurement honesty

For organizations

Running coding agents across a whole engineering org? SuperBased also offers central visibility across developers, budgets and policy, and SSO sign-in. See SuperBased for organizations or write to contact@superbased.app.

Community

Questions go to Discussions, bugs and adapter requests to Issues, and security reports through SECURITY.md. Pull requests are welcome: see CONTRIBUTING.md (building needs Go 1.27+, no CGO). If SuperBased saves you a surprise bill, a star helps others find it.

License

Apache 2.0