Find out what your coding agents cost you ACE Fleet released ACE Sidecar, a local developer observability tool that records the cost of AI coding agent sessions, supporting Claude Code and Google Antigravity, with metrics stored in a local SQLite file and no data leaving the machine. The tool requires Python 3.12+ and can be installed via uv or pipx, then run with 'ace up' to start a proxy and dashboard at http://127.0.0.1:8787/dashboard. Local developer observability for AI coding agents — see what your Claude Code and Antigravity sessions actually cost. What it does · · Requirements requirements · Install install · Quickstart quickstart · Features features · Who builds this who-builds-this · Configuration configuration · Endpoints endpoints · Development development License license ACE Sidecar runs a proxy on your machine in front of the model provider, recording what each turn costs — tokens in and out, what came from cache, how long you waited. It also reads your existing transcripts ~/.claude/projects , ~/.gemini/antigravity/brain , so there is history to show from the first run. Nothing leaves your machine: no account, no upload. Metrics live in a local SQLite file you can delete. Built by ACE Fleet https://acefleet.dev — see Who builds this who-builds-this . Python 3.12+ — the one hard requirement. Check with python3 --version .- macOS, Linux, or Windows. No admin rights needed. A coding agent you already use. Claude Code and Google Antigravity are supported today. However you pay for it — subscription or API key — is how it stays paid; the sidecar adds no account of its own. Need a newer Python? brew install python@3.12 macOS , sudo apt install python3.12 Debian/Ubuntu , sudo dnf install python3.12 Fedora , python.org https://www.python.org/downloads/ Windows , or uv python install 3.12 anywhere . uv tool install ace-sidecar or: pipx install ace-sidecar Both give ace its own isolated environment and put it on your PATH. With plain pip, use a virtual environment: python3.12 -m venv ~/.venvs/ace && source ~/.venvs/ace/bin/activate pip install ace-sidecar If install fails with Could not find a version that satisfies the requirement ace-sidecar , your Python is older than 3.12. The message blames the package, but the package is fine — the interpreter is too old. If ace: command not found after installing , run uv tool update-shell or pipx ensurepath , then open a new terminal. ace up Then point your agent at it and open the dashboard: eval "$ ace env " exports ANTHROPIC BASE URL open http://127.0.0.1:8787/dashboard Use your coding agent as normal — turns appear live, with your transcript history already loaded. On a Claude subscription, start with ace up --no-key Claude Code sends its own credential and the sidecar relays it , or put {"no key": true} in ~/.ace/config.json . ace up --help lists every flag. Antigravity needs no setup and no base URL: it is read from its transcripts on disk, so its sessions appear in the dashboard whether or not the sidecar was running at the time. Only Claude Code routes through the proxy. Unified view across agents. Claude Code and Antigravity in one place, with per-agent cost, sessions, turns and models. Pick one agent and the page scopes to it. Real spend against published prices. Per-turn cost from a versioned rate catalog — input, output, cache-read, and derived cache-write rates — with the source and the date it was checked. Cache savings shown as a counterfactual. Recommendations off a measured threshold. Each one fires on a number from your own transcripts and carries its saving, its cost, and its risk. Optimisation levers, ranked by what they are worth to you. The rail orders every lever by the money it would recover on your transcripts, with the share of your bill and the risk beside it. Each is scored alone, so the figures overlap and do not sum — and none are wired yet: this release measures. Workflow skill miner. Repeated command sequences become reusable SKILL.md rules, installable into .agents/skills/