IronClaw Companion Data Layer

Three questions, answered

Everything you need to trust your agents

IronClaw runs your agents. The companion data layer tells you exactly what happened.

💸

01 / SPEND

What did it cost?

Estimated token costs are confirmed against NEAR AI's billing API within 3%. Every inference ID is reconciled — nothing is silently billed as $0 and nothing is fabricated.

📋

02 / ACTIVITY

What did my agents do?

At least 98% of agent turns, tool calls, and model calls are logged as redacted OpenTelemetry spans. Query run timelines, failure rates, and memory health from pre-built dbt models — no raw SQL required.

🔐

03 / VERIFICATION

Was the answer verified?

Responses from NEAR AI's confidential inference API carry attestation quotes checked against Intel Trust Authority tokens. The companion tracks the percentage of verified responses per model.

How it works

From agent runtime to your dashboard

A clean, OTel-native pipeline with no proprietary lock-in.

IronClaw runtime

hooks + LiteLLM proxy

Secret sanitizer, event log, OTel GenAI exporter, hash-chained audit log

OTLP Collector

redaction pass-2

Redacted spans forwarded; metadata-only by default

Ingest workers

cost reconciler + verifier

Confirms costs vs NEAR AI billing API; verifies TEE attestation quotes

Raw tables

Postgres or Parquet

Immutable event store; nightly Parquet/Delta export available

dbt models

usage, audit, failures

usage_ledger, run_timeline, audit_verification, memory_health

Outputs

dashboard · alerts · CLI

Org dashboard, budget alerts via n8n, CLI audit verify, compliance PDFs

Features

Built for trust and observability

Every guarantee is backed by deterministic code with tests — no LLM decides what is billed or verified.

💰

Cost reconciliation within 3%

Batches up to 10,000 inference IDs per call to NEAR AI's billing API. Confirmed, incomplete, and mismatched rows are tracked separately — nothing collapses to $0.

📡

≥98% call logging

OTel GenAI spans capture every agent turn, model call, and tool execution. The LiteLLM proxy provides full coverage while upstream hooks are in review.

🔗

Hash-chained audit log

Privileged actions are recorded in an append-only chain; head hashes are signed periodically by the TEE key. ironclaw audit verify exits non-zero on any tamper.

🛡️

Attestation verification (TEE)

Gateway quotes are checked against Intel Trust Authority tokens for every NEAR AI TEE model. Models that don't support attestation are explicitly labelled — never silently skipped.

🔒

Privacy-first by design

Metadata-only spans by default. A two-pass redaction pipeline strips secrets before data leaves the runtime. Any runtime LLM calls run on NEAR AI TEE or local models only — agent data never reaches a public API.

⚡

Self-host in under 10 minutes

One Docker Compose stack, a single install.sh, and a .env template. Services run non-root with read-only filesystems and all Linux capabilities dropped.

Open & vendor-neutral

Deploy where you want

Built on OpenTelemetry. Runs anywhere Docker Compose runs. Designed for Tailscale-private deployments.

Open standards

OpenTelemetry GenAI semantic conventions — no proprietary span format
OTLP over gRPC/HTTP — works with Grafana, Langfuse, Jaeger, any OTel backend
dbt models are plain SQL — inspectable and forkable
Postgres or Delta/Parquet raw store — no vendor lock-in

Deploy targets

ZBOX homelab (self-host, Tailscale-only network)
Hetzner VPS — same Compose stack, works out of the box
Inside a customer TEE for hosted agents
GitHub Codespace for development and gate testing

curl -fsSL https://raw.githubusercontent.com/\
  EdwardPlata/ironclaw-datalayer/main/\
  deploy/install.sh | bash

GitHub repo Docs OTel-native Zero SQL required

Know what your agents did, what it cost, and whether it was verified.