What's coming

What is designed but not yet shipped. Anything marked planned in these docs appears here. We document a feature when it lands — see Docs conventions.

Shipped today

Carbon Passport + attestation API

The ledger.json you already produce, signed and anchored to on-chain retirement receipts — so a third party can verify it programmatically instead of trusting a badge image.

• carbon-md passport — signs a canonical credential (Ed25519, did:key locally) including retirement anchors: transaction hash, registry serial, vintage, certificate URL.
• carbon-md verify <url> — re-derives the trust level from evidence. Works offline for signature and ledger checks.
• Trust ladder — L0 declared → L1 measured → L2 contribution-verified → L3 certified. L0–L2 are free and machine-verifiable forever.
• Live badge — re-verifies on request, so it can't drift from reality.

Why it matters: it turns "trust the badge" into "check the receipt", and it's the first layer someone can reasonably pay for (L3 certification), while everything below stays free.

Unified token accounting

Capture adapters don't yet treat reasoning and cache_write tokens identically — sync hermes records them without counting them, which under-reports on reasoning-heavy models. The raw counts are preserved in every ledger event's meta, so a future factors version can recompute historical footprints rather than losing them. See Methodology.

Python SDK + framework callbacks

A thin carbon-md Python package: an EcoLogits wrapper and a LangGraph callback exporting to the same local ledger. CrewAI and AutoGen to follow.

Local compute (CodeCarbon)

Tracking on-device inference and training. Cloud inference dominates most agent footprints, which is why this is a fast-follow rather than v0.1.

Organization rollup

organization_id is already accepted in the policy file. Rolling several agents' ledgers into one organizational view — and a CSRD-oriented export of AI emissions — is the enterprise direction.

The factor dataset

Real per-workload, per-model emission factors for agentic workloads don't exist publicly. Accumulating them — with provenance, versioning, and external review — is a long-term goal of the project and the reason ranges are published rather than hidden.

Want to influence what lands next? Open an issue ↗ — especially with a real workload we should be able to measure and can't.