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

Shipped (L0–L2): carbon-md passport signs the ledger summary as a verifiable credential (Ed25519, did:key, canonicalized) with retirement anchors; carbon-md verify re-derives the trust level from evidence — checking the signature, the uncertainty ranges, on-chain transactions on Base, and whether counted anchors really are removal. Works offline for everything except the chain lookup, and exits non-zero as a CI gate.

Still to come: the hosted attestation API (/v1/verify, live badge, signed registry), the public passport page, and L3 certification — the human-audited, revocable tier that is the paid surface.

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.