Local developer observability for AI coding agents — see what your Claude Code and Antigravity sessions actually cost.
What it does · Requirements · Install · Quickstart · Features · Who builds this · Configuration · Endpoints · Development · 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 — see 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 (Windows), or uv python install 3.12 (anywhere).
uv tool install ace-sidecar # or: pipx install ace-sidecarBoth 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-sidecarIf 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 upThen point your agent at it and open the dashboard:
eval "$(ace env)" # exports ANTHROPIC_BASE_URL
open http://127.0.0.1:8787/dashboardUse 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/<id>/ in one click.
Where the time goes. Wall clock split across model generating, tool execution, human composing, and idle — including time parked on approval prompts.
Prometheus exporter. 15 metrics in standard text exposition format at GET /metrics, for Prometheus, Grafana Alloy, OpenTelemetry Collector, VictoriaMetrics, or Datadog. See docs/PROMETHEUS_METRICS.md.
ACE Sidecar is built by ACE Fleet.
ACE Fleet is a middleware proxy for companies scaling AI applications. It sits between their services and the model providers and reduces what they spend on inference as that usage grows — across every workload in the business, not one team's tooling. That is the product.
This sidecar is one vertical of it, open-sourced on its own: the same accounting, pointed at a single developer's coding agents.
| ACE Sidecar (this repo) | ACE Fleet | |
|---|---|---|
| Scope | One developer's machine | An organisation's whole inference bill |
| Workload | Coding agents — Claude Code, Antigravity | Any AI application in production |
| What it does | Measures. Records and explains the spend | Acts. Reduces the spend in the request path |
| Where it runs | Loopback on your machine; nothing leaves it | Managed middleware between your services and the providers |
| License | Open source, AGPL-3.0 | Commercial |
The two answer different questions. The sidecar answers where is my money going on the machine in front of you, at a scale small enough to check by hand. Fleet answers what do we do about it once that question is being asked of an entire company's traffic.
Open-sourcing the coding-agent slice is deliberate: it is the part a developer can run in one command, on their own data, without talking to anyone — and the clearest way to show how the larger system reasons about cost. If it is useful at your desk, we would like to hear about it.
Settings resolve in order: CLI flags → ~/.ace/config.json → environment variables → defaults.
{ "no_key": true, "port": 8787, "log_level": "warning" }| Path | Holds |
|---|---|
~/.ace/telemetry.db |
Turn telemetry — local SQLite, never uploaded |
~/.ace/config.json |
Your settings |
~/.claude/projects, ~/.gemini/antigravity/brain |
Agent transcripts — read only |
Delete ~/.ace/ to remove everything recorded.
| Endpoint | Purpose |
|---|---|
POST /v1/messages |
The relay your agent talks to |
GET /dashboard |
The dashboard above |
GET /healthz |
Liveness and config state, without leaking your key |
GET /api/stats |
The dashboard's numbers as JSON |
GET /metrics |
Prometheus exposition |
Binds loopback and refuses non-local callers; a public bind needs --allow-remote.
git clone https://github.com/ACE-Engineering/ace-sidecar.git && cd ace-sidecar
python3.12 -m venv .venv && source .venv/bin/activate
pip install -e ".[test]"
pytest # 40 unit tests
python scripts/e2e_test.py # live route verification