Ottto - Observabilidade Local para Agentes de IA
ottto-ai/ottto
Veja custos e performance de seus agentes de IA em tempo real, local-first
Ver no GitHubFerramenta de observabilidade local-first que monitora uso, custo e performance de agentes de codificação IA como Claude Code, Codex e Pi, diretamente no seu Mac. Oferece dashboard com análise de gastos, limites de plano e recomendações de otimização sem enviar transcrições ou código para a nuvem.
Instalação
npx convez-ai ottto-ai/ottto1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Monitorar uso e custos de múltiplos agentes de codificação simultâneos
- Identificar gargalos de setup que prejudicam performance dos agentes
- Rastrear consumo de API tokens e planos de limite entre diferentes logins
- Auditar e otimizar configuração de MCP servers para agentes locais
Veredito

Ottto preenche uma lacuna real: visibilidade completa sobre agentes IA que rodam localmente com privacidade garantida. Implementação sólida em Rust com atenção genuína a segurança dos dados do usuário, ainda que limitado ao ecossistema macOS.
README
A tradução para o português deste README ainda está na fila.
Ottto
Ottto is one tool to see what is right and what is wrong with your coding agents. It works with several coding agents on your Mac, such as Claude Code, Codex and Pi.
- Everyday work: usage, plan limits and cost across your agents, logins and Macs, at a glance.
- What should you optimize: where your cost goes and which setup problems may hold your agents back, based on your own sessions and setup, so you fix what matters first.
What Ottto does on your Mac
- Reads the session files that Claude Code, Codex and Pi write (
~/.claude/projects,~/.codex/sessions,~/.pi/agent/sessions) and a few named settings from their config. - Uploads nothing about your usage until you connect this Mac to an Ottto account. Then it sends usage metadata, never your transcripts, code, tool output or sign-in tokens (what is sent).
- Sends session titles, repository names and MCP tool details by default. Every 6 hours it starts your local stdio MCP servers for Claude Code and Codex to list their tools, and sends each tool's name, description and input schema. In the web dashboard, Readable labels turns all of these off, and MCP inventory harvesting only the MCP part.
- Reads your plan limits from Anthropic with Claude Code's sign-in, and through the Codex CLI for Codex. Each can be turned off.
- Runs short
claude,codexandpicommands to check your setup (including test prompts), read limits and keep Claude sign-ins fresh. - Edits agent settings on setup and repair: Claude Code's
~/.claude/settings.json(telemetry and status line) and a marked line in~/.zshrc. Later checks put its status line back. At start it removes only its own old entries from~/.codex/config.toml. - Receives Claude Code's telemetry on 127.0.0.1 and keeps only content-free counters, which feed the usage it uploads.
Details: docs/privacy.md.
Get started
-
Install Ottto.app. Download Ottto for Mac (Apple silicon, macOS 14 or later; signed and notarized).
Prefer the command line? Homebrew installs the
otttoCLI and the background service (not the app). Use one or the other:brew install ottto-ai/tap/ottto brew services start otttoMore options, updates and uninstall: docs/install.md.
-
Sign in to your Ottto account in the app, or run
ottto setup. This connects your Mac and unlocks everything below: the full dashboard, the findings, and Ottto for Claude and ChatGPT.
With and without an account
Ottto for Claude and ChatGPT: ask about your usage, cost, data sources and recommendations from the chat you already use. Ottto's connector is read-only and cannot change anything in your account.
| Without an account | With an Ottto account | |
|---|---|---|
| Find the coding agents on your Mac | ✓ | ✓ |
| Check that Ottto's service is healthy | ✓ | ✓ |
| Collect diagnostics for support | ✓ | ✓ |
| Update and uninstall | ✓ | ✓ |
| Dashboard: usage, cost and sessions | ✓ | |
| Plans and limits for each login | ✓ | |
| Findings on cost and setup | ✓ | |
| Recommendations | ✓ | |
| Every Mac and login in one place | ✓ | |
| Ottto for Claude and ChatGPT | ✓ |
Everyday work
| What you get | |
|---|---|
| Every agent in one view | Claude Code, Codex and Pi, side by side · subscription or API key · Bedrock and Vertex AI billing (org admins) |
| Plans and limits | Several subscriptions and accounts: up to 10 Claude Code and 10 Codex logins per Mac · 5-hour and weekly windows |
| Cost breakdowns | By agent, model, repository, machine, subscription and token type · day, week or month · CSV export |
| Sessions and sub-agents | Timeline, tokens and models per run · which agent spawned which sub-agent, and what each one cost |
| Many Macs, one account | Every Mac you connect adds up in one account · logins stay separate |
- Agents are counted from what they record on your Mac, whether they run on a subscription or an API key. Org admins can connect Bedrock and Vertex AI billing, read-only.
- Limits include per-model limits, credits and reset times where the provider shows them. They appear only when Ottto can match a reading to the right login. For automatic Codex limit updates, turn on Keep limits up to date on each Mac.
- Cost shows estimated actual cost next to the pay-as-you-go value of the same usage, so you can see what a plan is worth to you. Subscription fees are prorated; unknown payers and unpriced usage can make the estimate partial.
- Sessions show context measurements and compactions where the agent records them. The Families view gives each agent's own cost where priced, and a total for the family.
- Logins are never pooled or rotated to get around limits.
What should you optimize
Sign in, and Ottto imports recent sessions from your Mac in the background, so findings start from history you already have. It points you to the biggest costs and setup problems first; the fix is yours to choose.
| What you get | |
|---|---|
| Where to save | Findings from your own sessions and setup, with evidence · a prompt you can paste into your agent |
| Recommendations | A Recommendations page with prioritized suggestions for your account |
| Setup that holds agents back | Nested AGENTS.md Claude Code does not load on its own · a CLAUDE.md too large to load · a MEMORY.md near its limit |
| What fills the context window | Instruction and memory files, skills and MCP tools at start · files read, search and tool output while working |
| Setting changes and their cost | Cost per request before and after you change the default model, reasoning effort or fast mode |
| Repeated and scheduled runs | What repeated work and automations cost |
- Findings include MCP tools whose definitions load up front in your latest setup, Claude Code sessions with very large contexts, the extra cost of Claude Code fast mode, and sessions on an older Claude Code version without idle compaction. Each finding is labeled, for example Measured or Estimate. Dollar figures on findings cover only the stated period, appear only when pricing is complete, and are not projected. The prompt asks your agent to investigate first and change nothing yet.
- Context start sizes are measured where the agent records them; what enters while working is estimated.
- Setting changes use the saved settings Ottto sees, not every temporary switch. They compare the same agent and Mac before and after; results say when the evidence is thin, and a before-and-after comparison is not proof of cause.
- Repeated runs are grouped by scheduled task, repeated prompt template or repeated title.
How it works
Ottto is not a proxy or an AI gateway. It does not reroute your agents' model requests and does not change their API key or base URL. It reads the session files your agents already write.
The ottto CLI
Use the app for the graphical interface. Use the CLI when you want to:
- script Ottto or read your data as JSON: every documented command takes
--jsonand returns documented exit codes - run Ottto on a Mac you reach over SSH
- let Claude Code or Codex check their own Ottto setup, with the skills
Install it with Homebrew, as in Get started. Ottto.app does not add ottto to
your PATH.
ottto setup # connect this Mac to your Ottto account, in your browser
ottto apps detect --json # find supported agents on this Mac
ottto doctor # show current service and agent health
ottto fix --app claude-code # repair Ottto's setup for an agent (or codex, pi)
ottto costs --json --days 7 # the last 7 days of cost across your account, as JSON
Privacy
Ottto never uploads your transcripts or full conversations.
What never leaves your Mac
- Your transcripts and full conversations, including your agents' responses
- Code, file contents, and tool and command output
- Your agents' sign-in tokens, API keys, cookies and passwords
- Instruction, memory and skill text, and absolute file paths (except text an MCP server itself puts in a tool description or input schema)
To show your Claude plan limits, Ottto uses your Claude Code sign-in only to ask Anthropic for your usage, and never sends it to Ottto. Turn this off in Ottto.app settings: Read subscription usage.
What is sent, only once you connect this Mac to an Ottto account (setup)
Usage metadata, so Ottto can price your usage and find savings: token counts, session timing and activity counts, plan limit usage, metadata about your accounts and workspaces (such as login emails, plans and ids), agent settings, one-way hashes of paths and of repeated prompts to group sessions, and basic details about the Mac. Setup sends the Mac's basic details when it starts. Diagnostics are a separate upload, sent only when you approve it.
Two settings in the web dashboard (app.ottto.net → Settings → Data & privacy) add more:
- Readable labels (on by default; a workspace admin can turn it off for everyone). With it on, Ottto also sends each session's title, or the first 80 characters of its first prompt when the session has no title, and, for prompts you repeat, a short start of that prompt (96 bytes at most, with paths removed), so you can recognize sessions. It also sends project folder and repository names, the names of skills, custom agents, MCP servers and MCP tools, and the repository paths of files in a session's context. With the MCP inventory harvesting setting also on (the default), it sends each local MCP server's configured name, transport kind, whether it answered, whether it is disabled and the character length of its instructions (never their text), and each tool's name, description and full input schema as the server reported them. Server commands, arguments, environment values, working directories and URLs are not sent (details). Turning Readable labels off stops all of those. Built-in tool names, counts and account metadata are still sent.
- Session links (off by default; a workspace admin setting). With it on, Ottto also sends pull request, issue and commit links found in a session. They name their repository even with Readable labels off.
If you add Ottto to Claude or ChatGPT, the answers you ask for go into that chat.
More: privacy policy for the hosted service and docs/privacy.md for this runtime.
Open source
This repository holds Ottto's local runtime, written in Rust: the ottto CLI, the ottto-service
background service, the connector SDK and the readers for Claude Code, Codex and Pi. Ottto.app and
the hosted dashboard are not in this repository. Build and test: CONTRIBUTING.md.
To add an agent, start with connectors/README.md.
Docs · Install · Troubleshooting · Contributing · Security · Code of conduct · Apache-2.0 license