↑ ↓ to move · D runs the demo
Website edition · Agent Edition · Internal track · 13 agents

Helping startups scale
sustainably, from day one.

That is why we built Climatico. Website edition: a founder sees their own EI and a task list before early habits become permanent. Agent Edition: the same product, but any AI agent can submit data to it directly — and a fleet of agents runs the calculation, evidence, suggestions, and offset settlement without a human in the loop.

EventAgent Natives · Cloudflare SF · 26–27 Aug 2026
RuntimeOne Worker · 4 DOs · SQLite ledger
Cold start/ai-agent.json · mint · write
Fleetingest → audit → settle
Agents13 — Orepath, Provider, Scheduler, 3× Fleet, 4× Analysts, Clerk, Abatement, Cortex
Surfaces/mcp · /a2a · /v1/* · CLI
Desk127.0.0.1:8787
DashboardOpen live desk
The problem

Most early-stage startups never measure their environmental impact.

Founders do care — but they are not obligated to report, and they are busy shipping products and raising capital. Cloud regions, software tools, and business travel feel like quick defaults. Then they scale with the company.

01 · Timing

Early choices become permanent as the company grows

Habits formed on day one become expensive to reverse later. Waiting until Series B means guessing under pressure.

02 · Own EI, not the customer’s

Know your own impact while still Seed

Reporting will eventually demand the company’s numbers. Climatico is so a founder can predict their own EI — before they must report, and before they scale too far.

03 · No product in the gap

ESG suites, vendor dashboards, consultant PDFs

Wrong customer, one vendor, or twice a year and unqueryable. None of them is built for this stage — or for an agent to call.

One user · logistics as the agentic interface

Rae's company tracks emissions for customer shipments. It had no data on its own.

Rae Jin · founder · Orepath · Seed · 14 people in SF

Orepath traces battery materials mine → cell for EV buyers. Green on the customer’s chain. Then a cell buyer asked: what’s your company’s impact — not the mines? No answer can block the deal.

The highest-emission activity isn't on a dashboard. It's in a purchase order, a booking, a port record. Those systems are already run by agents. Climatico lets an agent submit real data at that moment — or it flags and records why, if the claim isn't backed by evidence.

Practical tools · live this weekend
  • Ground Oakland portbrief + Tavily. Evidence on the place.
  • Watch the port — L5. Survives restart. Still modeled tonnes.
  • File the tracer’s computerun_fleet on SJC. The live class.
  • File the battery freight legfreight: mode, weight, distance → real kg CO2e. Always commits, tagged grounded or modeled.
  • Flag a green chaingreenwash is a receipt, not a marketing line.
Honest next · not stubbed
  • Logistics is 12 t ±35% modeled baseline — the freight write is shipped, but one leg doesn't retire the modeled default.
  • Supplier-only / buyer-only scoped tokens for the two-sided PO flow — one credential files freight today.
  • Hardware workflow is a different client (factory they do not own). Named in the PRD. Not this Worker.
  • Dashboard: open Rae's tools · shows the same data agents submit.
Website edition · what Climatico does

Company data in. Impact out. A plan for this sprint.

The founder product — the website edition — is how Climatico helps startups scale sustainably from day one. Track, reduce, scale.

01
Company data in
Enter the website, or stage, headcount, cloud, vendors, travel. Empty fields still run on clear modeled defaults until bills replace them.
02
Impact out
A footprint estimate and hotspots — a prediction of what they would report later. Modeled until evidenced. Own EI, not the customer’s.
03
Actionable plan
What to change this sprint, with an owner and a done-when. Move a batch job to a cleaner region. Check sustainability at vendor renewal.
04
Motivation
Cohort leaderboards make progress social and comparable. Environmental impact measures business excellence.
Track Reduce Scale
Website edition · founder dashboard

See your EI. Act while it's still easy to change.

Orepath on the website edition: ~38.4 tCO₂e / yr, cloud as the hotspot, Level 1 of 5, a Monday list. Predicted vs intervened path, intensity next to revenue.

Climatico website edition dashboard for Orepath — estimated footprint, overview, and EI vs revenue paths
Website edition · founder dashboard. Enter a site, get a first estimate, then a checkable action for this sprint.
Website edition

Human in the loop

Assessment, report, dashboard, actions, cohort leaderboard. The founder product.

Agent Edition · this weekend

The same attribution, as a write

Discovery card, scoped token, policy before the write, durable receipt. Agents file the spike while it is still cheap.

One rule on both

Modeled until evidenced

Estimated numbers are labeled as estimates. Emissions reductions aren't counted until proven. We would rather show nothing than a made-up explanation.

Agent Edition · the one rule

Every rejected submission is recorded, the same as an accepted one.

Both rows live in the same SQLite table, with the same id, subject, token and permanence. The rules are checked before anything is saved. That's what makes it safe to let an unfamiliar agent use the system without supervision.

committedPOST /v1/actions · 201
brief · Houston, TX
{
  "status": "committed",
  "evidence": [ /* 5 Tavily sources */ ],
  "subject": "judge-agent"
}
flaggedPOST /v1/actions · 422
greenwash · everywhere
{
  "status": "flagged",
  "flagCode": "forbidden_intent",
  "flagReason": "the desk does not mint
    unverified claims."
}
Writes the desk accepts

brief watch offset assess abate switch refund freight run_fleet

Flagged and stored

payout greenwash exfiltrate ungrounded · unknown intent, no location, over ceiling. The policy never runs after the write.

Agent Edition · the Worker in climatico/

The dashboard shows the same data that agents submit.

Assess / Grow / Fleet Pipeline / 3-Sided Swarm / Impact & Abate / Ledger & Receipts / Inbox / Agent Clerk. Humans read the same ledger agents write. Website edition is the founder dashboard; this is the Agent Edition desk on the Worker.

http://127.0.0.1:8787/?tab=assess&from=deck
What a judge actually scores

A stranger agent, zero setup, five calls to a permanent record.

Internal track, External-shaped discovery. The 30-point field is what an unfamiliar agent can actually trigger — /ai-agent.json, not the marketing description.

01
Read the probe
GET /ai-agent.json — Immersive Commons looks for this filename. Agent Card and MCP cards sit next to it.
02
Mint a scoped credential
POST /v1/credentials — scopes freeze at mint. Ceiling 5 000¢. HMAC token, TTL ≤ 24h. Thirty mints a minute, then 429.
03
Carry the bearer
Same token on /mcp, /a2a, /v1/*. Unauthenticated writes are 401 before any work.
04
Write
MCP complete_action / run_fleet, REST POST /v1/actions, A2A SendMessage. One policy.
05
Keep the receipt
Committed or flagged, same shape. Retry with the same idempotencyKey returns the original row.
CALLER PROBE SURFACES POLICY LEDGER Any agent never seen us /ai-agent.json event probe POST /v1/credentials scopes freeze MCP /mcp · 15 tools A2A /a2a · SendMessage REST /v1/actions · fleet Clerk · Workers AI Kimi K2.6 evaluatePolicy() + Tavily or ungrounded Ledger Durable Object SQLite receipts · credentials watches · handoffs fleet_runs
Adding a new connection method never adds a way to skip the rules. Rejections are stored the same way approvals are. Even if the connecting device disconnects, the stored data stays available.
Demo · scripts/demo.sh against the live Worker

Point an agent at it. Or press play.

This is the walkthrough a judge can run without us. It hits the desk at http://127.0.0.1:8787 — mint, 401, grounded brief, stored flag, fleet over budget.

idle
Calls
Response
Press play. If the Worker is down, the snapshot path still shows what each call returns.
Internal track · real work across a boundary

A $420 SJC spike, attributed before the invoice exists.

Ingest parses the region. Audit scores compute with a labelled heuristic of 0.45 kg per USD, cited against live Tavily sources — not GHG Protocol ICT. Settle writes an offset if the month is over budget. Every handoff is a row.

cloud spend, SJC + $420 ingest spike → location + spend flags if no region audit 189.0 kg · heuristic 0.45 kg/$ 40 + 189 vs 50 kg → 179 over flags if ungrounded settle 3 580¢ offset, token-capped 20¢ per kg over budget $12 spike settles nothing fleet.ingest → fleet.audit → fleet.settle durable handoffs · list_handoffs · Cotal-shaped, runs without a mesh
Compute is the only live class today. The other six stay tagged modeled. A $12 in-budget run is also a receipt — settlement action: none.
189kg this spike
179kg over budget
$35.80offset receipt
3handoffs per run
Honesty · the Worker wins where they disagree

What you can trigger today, and what we will not fake.

Website edition · the product

See your EI. Monday list.

  • Assessment, report, founder dashboard, actions
  • Cohort leaderboard · Track / Reduce / Scale
  • Own EI, modeled until evidenced
NOW · Agent Edition in climatico/

An unfamiliar agent can submit real data

  • Discovery: /ai-agent.json, Agent Card, MCP
  • Scoped bearer, 401/422 stored, Tavily or ungrounded
  • Fleet ingest → audit → settle on compute
NEXT · named, not stubbed

Hardware is a different client

  • Six classes remain modeled until they are writes
  • Rae’s PO / freight write — shipped: mode/weight/distance → graded kg CO2e, always commits
  • Supplier-only / buyer-only scoped tokens for the two-sided PO flow — next
  • Tenki wired — sandboxes for agent test workflows
Assessment · seven classes
ClassScopeDefault±Live today
Cloud & AI computeS3 Cat 18.5 t/yr30%live · fleet + Tavily
HardwareS3 Cat 23.240%
TravelS3 Cat 6–74.825%
Vendors & SaaSS3 Cat 12.150%
LogisticsS3 Cat 4 & 912.035%
ElectricityS25.615%
DirectS11.520%
Who's in the room · real, not booth-only

Four sponsors make this live.

Host

Cloudflare

Workers, DO, Workers AI

Evidence

Tavily

Web search on the write path

Mitosis

Cortex agent memory, wired into the scheduler

Tenki

Agent test environments

Two sponsors, two real mechanics · not badges

A sandbox the Worker can create but not run. A memory that round-trips for real.

Tenki hit a real platform limit this weekend and got built around it honestly. Mitosis is a verified real write, then a real recall — not a claim.

Tenki · two hops, because a Worker can only do one

/v1/sandbox/verify → exec outside → /v1/sandbox/complete

01
Worker creates a real sandbox
A hand-built fetch transport calls Tenki's control plane straight from workerd. A real session id comes back.
02
The Worker can't run anything in it
Executing a command is a duplex stream. workerd said so directly: "the http2.connect method is not implemented." No transport swap fixes that.
03
A real runtime execs it
Hermes, or any Node/Python caller, runs the command in that exact sandbox and reports the real output back.
04
Flagged if the session isn't ours
The reported sessionId is checked against what this ledger actually created, then re-confirmed with Tenki itself. Invented output is a 422.
Mitosis · agent memory, verified round-trip

cortex_remember → cortex_recall, real universal_id

Every scheduled fleet run writes a summary to Cortex and can recall it back — the team's own agent memory, reachable via /v1/memory and MCP, not a Climatico-internal cache.

POST /v1/memory · 201
{
  "ok": true,
  "memoryId": "universal_id..."
}

What it's for: enrichment, not the write path itself — a fleet run commits or flags on policy and Tavily evidence alone. Cortex is where that history becomes queryable later, by us or another agent.

Agent Edition · an agent army on both sides of one ledger

Client swarm ↔ Climatico service → buyer ecosystem.

The same ledger grew a real cast of investigating agents this weekend. Orepath's own compute-watcher files what it finds autonomously, Climatico's fleet enforces policy and settles receipts, third-party providers deliver services through the same surface, and downstream buyers verify with read-only credentials. One ledger underneath all of it.

Orepath + Hardware Co

compute-watcher (Durable Object alarm) spikes cloud spend every 15 min during working hours. PortLogisticsAgent grounds the Oakland port. Hardware Co’s FactoryProcurementAgent files factory POs; OceanFreightAgent watches the port. A rogue MarketingBot tries to mint a green claim on purpose.

Service army

Climatico: 10 agents in one Worker

Ingest parses the spike. Audit scores kgCO2e against live Tavily sources. Settle commits a token-capped offset. Policy flags forbidden intents — and records the rejection. Scheduler (cron) runs all three plus Tavily abatement research every 30 min. Abatement researcher + summary writer (Workers AI) find real greener alternatives. Report builder compiles /v1/report. Inbox analyst pushes hotspot alerts. Clerk (Workers AI) answers questions. Every handoff is a row.

3rd-party + buyer

Provider agent · EV OEM auditor

green-offset-co (Durable Object) fulfills offsets the fleet writes, reviews receipts, earns revenue — delivered through the same surface. The buyer mints a climatico:read-only credential and audits the ledger before accepting any supplier claim. Transact rights: none. Proof of attribution that cannot write.

The boundary · what every agent must cross

/.well-known/agent-card.json A2A discovery · /v1/credentials scopes freeze at mint (5 000¢ ceiling) · /mcp · /a2a · /v1/* one policy underneath · SQLite Ledger the settlement of record · climatico.sh CLI for human operators. In the UI this is the 3-Sided Swarm tab, deployed and live.

Privacy · signal-only, like Salesforce / Google / Workday

Your business information is very secure in Climatico.

When climatico.sh connect ~/my-startup runs, it scans the folder on the startup's machine and extracts only derived signals. The server never sees file contents. The CLI prints the exact JSON payload before sending and supports --dry-run to inspect without transmitting.

Stays on your machine
  • File contents of package.json, wrangler.*, README.md, .env, source code
  • Any other file in the folder
  • Your data, your records, your secrets
Leaves (small JSON only)
  • Vendor names from deps: "stripe", "datadog"
  • Cloud provider from config: "Cloudflare Workers"
  • README keyword hits: "mentions shipping"
  • kind, class, confidence — no values, no snippets
Why this is the same model as the SaaS tools you already trust

Salesforce reads your org's metadata, not your records.   Google reads your ad counts, not your emails.   Workday reads your headcount, not your salary data.   Climatico reads your signals, not your source. We do not store your files. We do not train on them. We do not send them to third parties.

This is a binding constraint, not a feature

The CLI source is auditable. The server's /v1/connect handler does not accept a file_contents or path_contents field — by design, by contract, by source code. Future data connections (Salesforce, Google billing, Workday) will follow the same pattern: extract rollups locally, send rollups, never records.

No stubs · two_sided_swarm.py against the live Worker

Real numbers across the boundary, end to end.

A CLI swarm (climatico/scripts/two_sided_swarm.py) minted a scoped bearer, drove a genuine fleet spike, forced a forbidden claim, and let a read-only buyer inspect the ledger — all against climatico.dalrae-jin-work.workers.dev.

189kgCO2e · SJC $420
+179kg over 50 kg budget
$35.80offset receipt 02564267
6receipts a buyer verified
Committed · grounded
  • 5 live Tavily citations scored the factor — not invented as GHG Protocol
  • Offset 02564267 committed to SQLite, token-capped
  • Role separation is a factor, not a chatbot paragraph
Flagged · stored
  • Greenwash claimed at “Orepath Global Chain” intercepted before the write
  • Flag eac21a5b persisted — same table as commits
  • Read-only buyer minted climatico:read and audited 6 receipts
Modeled impact & abatement · not measured

The footprint behind the business — and how to run the same business differently.

Orepath today, modeled: ~37.7 tCO₂e/yr across seven classes. CO₂e is the modeled core. Water, air pollution and e-waste sit upstream and downstream — labelled here but not yet broken out per class, because Climatico won’t print a number it can’t back up. Alternatives are the same business, run a different way — each keyed to the business source that drives the tonne.

37.7 ttoday · modeled
48 t+12m if abating
62 t+24m clima path
132 t+24m if nothing changes
Cascade · business source → impact → alternative → modelled abatement
Business sourceImpact (t CO₂e/yr · modeled)Where it’s felt upstreamAlternative — same business, different runAbatement (modelled)
Cloud & AI compute8.5 ±30%data-centre power & cooling watercarbon-aware scheduling; move batch to a low-carbon regionlargest share of compute
Logistics (Rae’s biggest)12.0 ±35%freight diesel → local NOx & PM airOakland → rail / intermodal; route optimisationbiggest lever · freight write shipped
Purchased electricity5.6 ±15%location-based grid mixcleaner utility or on-site solar — a bill makes it measuredmoves class to measured
Travel & commuting4.8 ±25%aviation contrailsvirtual-first; rail over air where it fitsper-trip cut
Hardware & electronics3.2 ±40%mining + e-waste end-of-liferemanufactured units; extend refresh beyond 4 yearslonger device life
Vendors & SaaS2.1 ±50%supplier electricity upstreamconsolidate vendors; right-size licenceslicence-led cut
Direct emissions1.5 ±20%combustion & fugitivenear zero for a cloud-only team
Long-run projection

−70 tCO₂e/yr at +24m (clima vs default)

At +24 months the abating path runs at 62 t vs 132 t if nothing changes — a modelled ~53% reduction from how the same revenue is produced. A projection, clearly labelled, not a measured cut.

Why it’s the business source

Each class is a PO, a booking, a port

Not a dashboard metric — a system an agent already touches. That is the moment Climatico files it: the region, the freight mode, the server, the licence.

Water & air rows are upstream, not scored

CO₂e is the modelled and filed core. Water, NOx/PM and e-waste are named as knock-ons — but the desk won’t mint a pollution or handprint number it cannot back with evidence.

Cotal · $300 best-use — a resident bot is live

A resident auditor, awake on #team.climatico.

meshaudit joins wss://hack.cotal.ai/mesh-ws as its own actor, subscribes #team.climatico, and serves each region mention through Nebius (custom:nebius → DeepSeek-V4) — a real Hermes peer, not a fake write.

Live

Join · subscribe · act

  • hermes-agent 0.16 on the pinned line
  • Subscribed to #team.climatico + #announce
  • Runs tools — inspected its own env on a live turn
What it took

Small connector surgery

  • Cotal spawn strips scratch env, so the launcher reads NEBIUS_API_KEY from .dev.vars
  • Writes providers.nebius + model.provider: custom:nebius into the isolated profile
  • Persona + launcher in climatico/scripts/run_meshaudit.sh

Named, not faked

  • Connector patch edits node_modules — a cotal update resets it
  • Live listener + Nebius routing proven; a full posted audit is next
Try it

Website edition for founders. Agent Edition for any AI agent, even one Climatico has never seen before.

Run the desk
cd climatico
cp .dev.vars.example .dev.vars
npm install && npx wrangler types
npm run dev          # 127.0.0.1:8787
npm run demo         # this slide, in a shell
./climatico.sh fleet SJC 420   # CLI: 23 commands
./climatico.sh orepath         # check Orepath agent
./climatico.sh agents-start    # start the auto-pilot

Mint, prove the 401, file a Houston brief, commit an offset, store a greenwash flag, run the fleet under and over budget. CLI runs the same flows from any terminal.

Surfaces a probe will find
/ai-agent.jsonevent probe
/.well-known/agent-card.jsonA2A card
/v1/credentialsmint, scopes freeze
/mcp15 tools, bearer required
/a2aSendMessage
/v1/actionsthe write (brief, watch, offset, assess, abate, switch, refund, freight)
/v1/fleet/runingest → audit → settle
/v1/reportplain-English progress + suggestions
/v1/agents/orepathOrepath employee agent (auto-fleet)
/v1/agents/provider3rd-party offset provider
/v1/memoryCortex memory store / recall
/v1/workspacesame feed as the desk
climatico/climatico.shCLI — 23 commands

Measure early. Act early.

Helping startups scale sustainably — from day one.