Easy AI coding observability. No guessing who did what.

rubberai records what your AI agents actually do — every prompt, the tokens it cost, the files it changed and the diff — across any IDE, agent or model. One Go binary and a Postgres. No LLM proxy, no Kafka, no cloud account, and nothing leaves your machine unless you say so.

The demo is a read-only view of the real project used to build rubberai — no signup, nothing to install. Or docker compose up --build for your own. MIT licensed. Metadata-only collection by default.

One row per prompt you typed

A single prompt fans out into dozens of model calls and tool calls. rubberai collapses that back into the turn you actually asked for, with everything it cost and everything it touched.

The activity overview — real data from building rubberai with Claude Code

Click a file count to see every file with its git state and diff. Click a prompt to read the agent's own account of what it did.

Click a file count — every file the turn touched, with its git state and the diff

@@ -13,13 +13,30 @@ let group=$state('prompt'),userFilter=$state(''),- let filePop:{x:number;y:number;rows:FileChange[]}|null=$state(null);- function showFiles(event:MouseEvent,rows:FileChange[]|null){-  if(!rows||!rows.length){filePop=null;return}+ let changeSet:{label:string;rows:FileChange[]}|null=$state(null);+ function mergeChanges(rows:FileChange[]):FileChange[]{+  // One event is recorded per change, so a file edited twice+  // in a turn arrives twice. Merge by path.+  const byPath=new Map<string,FileChange>();  function closeChanges(){changeSet=null;changeFile=null}

Changes are read from git after every tool call, so an edit made through a shell heredoc or a build is caught the same as one made through an edit tool. Staged, unstaged and untracked are reported separately, never flattened.

Hover a username — who, from which machine

Identity comes from git — the email a developer already signs commits with — plus the host and its LAN address, read from the routing table with no packet sent. All three travel as metadata, which a metadata-only project discards on arrival. Shown here with placeholder values.

What the demo button opens — the real dashboard, read-only

How it stays safe: the button opens /demo on the server, which issues a short-lived session that can only read one designated project. There is no password in the page, no shared login, and the session cannot create keys, change settings or write events. Visitors who want their own data create a workspace of their own.

What it records

• Prompts you typed — and only those. The agentic loop that follows is attributed to the prompt, never mistaken for one.
• Tokens, split honestly — new input, output, cache reads and cache writes, each priced at its own rate per model. Cache reads were 80% of the bill while building this; a blended figure hides that.
• Every file changed, however it was changed — read from git after each tool call, so a sed, a heredoc or a build is caught the same as an edit tool. With the diff.
• Who, from where — git identity, host and address, opt-in twice: once in the integration, once in the project's privacy mode.
• Idempotent ingestion — retry any request after a timeout; the same event is never counted twice.
• Any tool — the contract is HTTP JSON. Claude Code and Codex adapters are included; OpenTelemetry logs are accepted directly.

Privacy first, by default

A new project collects metadata only: counts, timings, identifiers. Prompt text, diffs and personal data are discarded on arrival unless you switch the project to full collection. Nothing is inferred — an AI change without evidence is recorded as unknown, not claimed. Usage is not a developer performance score, and the dashboard says so.

Integrate in one request

Any language, no SDK required. This is the whole contract:

# one event; batches of up to 100 are atomic
curl https://rubberai.example/api/v1/events \
  -H "Authorization: Bearer $RUBBERAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"event_id":"evt_1","event_type":"llm.request.completed",
       "project_id":"prj_…","timestamp":"2026-09-12T10:00:00Z",
       "model":{"provider":"anthropic","name":"claude-sonnet-5"},
       "usage":{"input_tokens":1200,"output_tokens":300,"cached_tokens":900}}'
→ {"accepted":1,"duplicates":0}

Examples in curl, Python, JavaScript and Java. A durable local collector buffers events to disk and retries, so a server restart never loses a turn.

Read more

Pricing