What it costs

There is no paid tier.

The harness is Apache-2.0. No licence key, no seat count, no paid unlock, no watermark, no attribution nag, no "contact us for enterprise". We are not building a funnel on this page. What you pay is your own Google Cloud bill and your own model plans — and since you set the caps on both, the only person who can tell you the number is you. This page is the breakdown you need to work it out.

Three bills, one of them empty

Where money can possibly go

Every cost associated with running this system falls into exactly one of these. There is no fourth column, and we do not appear in any of them.

The harness

Free

Apache-2.0, the whole tree, including the installer, the executor, the crypto envelope and the ~55-tool MCP server. Fork it, run it, modify it, ship it inside your own product. No tier above this one exists to be upsold into.

Google Cloud

Metered

Billed by Google, to your billing account, for resources in your project. Scale-to-zero across the request path means an idle install is close to nothing; a busy one costs what a busy one costs. You set the budget and the caps.

Model access

Yours already

Your own Claude plan, and optionally your own Gemini or Vertex credentials. The three-position work-mode switch decides which is used, and one of its positions spends nothing beyond the plan you are already paying for.

We will not quote you a monthly figure, because any figure we invented would be wrong for you. Cost here is a function of how much you actually run.

Line one

What Paracoding charges: nothing, and here is why that is structural

Not a promotional price. There is no mechanism in the software by which we could charge you.

The harness installs into a Google Cloud project you own, using your credentials, creating resources under your IAM. It does not register with anything of ours. There is no licence check at boot, no telemetry beacon, no usage counter reporting home, no account of ours in the trust path. That is not restraint on our part — it is the architecture. Software that runs entirely inside your tenancy has nowhere to send a bill from.

This has consequences you should weigh in both directions, so here they are honestly.

What you get. No vendor lock. No pricing change you did not vote for. No seat maths. No service of ours that can go down and take your fleet with it. No terms update that suddenly claims rights over your prompts or your repository. If we disappeared tomorrow your install would keep running, because nothing in it was ever pointed at us.
What you do not get. No SLA from us, because we operate nothing on your behalf. No support contract, no account manager, no certification — we make no SOC 2, ISO or "audited" claims. You own the operations of the project you installed into, including the bill and the failure modes. Support is the repository, the issues, and the source.

If you want somebody to be accountable to you for uptime, this is the wrong shape of product and we would rather say so on the pricing page than after you have installed it.

Line two

What Google charges you

Component by component, with the cost driver named. Prices come from Google's public pricing and change without asking us; check the current rates for your region before you commit to anything.

Cloud Run × 2
Two services from one image: the console surface behind IAP, and the MCP surface. Both scale to zero. An install nobody is using holds no instance, so an idle install costs essentially nothing beyond storage. Cost appears when requests do, driven by request count, request duration, and the CPU and memory you allocate per instance. The serverless git design is what makes this true — no server holding a clone on a disk means nothing has to stay up to keep your repository alive.
Firestore
Refs, HEAD, config, .git/index, packed-refs, the working tree, work items, memory graph, journal entries. Billed on document reads, writes, deletes and stored bytes. Agent activity is the driver: a fleet doing a lot of small commits and journal writes does a lot of small operations. Compare-and-swap on refs runs as a Firestore transaction, which is a write like any other.
Cloud Storage
The lake and the git objects, each sealed in a PCV1 envelope. Billed on stored bytes, operations and egress. Objects accumulate — that is what an object store is for — so this is the line that grows monotonically with history unless you prune. The envelope adds a small fixed overhead per object (magic, epoch, flags, a 12-byte nonce, a 16-byte tag) plus the 1120 bytes of X-Wing encapsulation ciphertext where it applies.
Secret Manager
Billed per secret version stored and per access operation. Small in absolute terms for a normal install, and driven by how many secrets you keep and how often running code reads them.
Cloud KMS
Key versions held, plus cryptographic operations. Two distinct users here: X-Wing hybrid PQC key encapsulation for the envelope (useToDecapsulate), and the asymmetric signing the executor verifies. Rotation creates key versions; each version held is billable for as long as you keep it.
Cloud Build
Billed on build-minutes. You pay this on install, on release cuts, and every time a strain builds a container for you — which is exactly what happened in the calculator screenshot. If your fleet builds a lot, this line moves.
Vertex AI (optional)
Only if you use the keyless work-mode bus for the work-item queue. Billed per token against your Google Cloud account. Leave the switch in home and this line does not exist.
Workstation VM (optional)
Only if you run one. It idles off, so the meaningful charge is running time plus the persistent disk, which bills whether the machine is on or not. Most of what people expect to need a VM for happens in Cloud Run without one.
Google IAP
The identity layer in front of the console. Its cost is negligible next to the things it protects, and it is not optional in this design — it is how an out-of-domain account gets refused before it reaches your console.
Scale-to-zero is the single most important sentence on this page. Two Cloud Run services with no traffic hold no instances. The cost of an install you set up and did not use this month is dominated by bytes at rest and key versions held — not by anything that has to stay running to keep your state alive. Nothing has to stay running, because the state is the store.

Line three

What your model plans cost

The switch that decides this has three positions, and one of them spends nothing new.

Strains think with models you bring. The harness does not resell model access, does not proxy your calls through us, and holds no key of ours in the path. A three-position work-mode switch determines where inference comes from.

home

Your own Claude plan

Work runs through the Claude plan you already pay for, from the client you already use, over the MCP endpoint your install exposes. Spends nothing beyond that plan. If you hit a usage limit, that is a limit on your plan, not a bill from anyone new — and because agent state lives in your project, you can move to another plan, paste the bootstrap, and resume with the history intact.

dual

Both, deliberately

Interactive work on your plan, queued and background work through the model bus. You are now paying two meters on purpose: the plan you already had, plus per-token charges for whatever the bus consumes. Choose this when you want work items progressing while you are not in the chat.

work

The keyless bus

Work items run against Vertex from inside your project using workload identity — keyless, no API key to leak or rotate. Billed per token to your Google Cloud account, on the same bill as everything else in the breakdown above, with the same budget controls applying.

Gemini or Vertex credentials, if you use them, are yours and stay in your project. The keyless path exists because a long-lived API key sitting in a config file is a liability with a cost attached to it that nobody puts on a pricing page: the day it leaks.

The table

Component to cost driver

If you want to estimate, estimate from the right-hand column. That is where your usage actually enters the arithmetic.

ComponentBilled byWhat drives the costIdle cost
The harness itselfNobodyApache-2.0. Nothing to meter, nothing phones home.None
Cloud Run — consoleGoogleHuman sessions: requests, request duration, allocated CPU and memory.Scales to zero
Cloud Run — MCPGoogleTool calls from your MCP clients and agents. More agents, more calls.Scales to zero
FirestoreGoogleReads, writes and deletes: commits, ref transactions, work items, journal entries, memory-graph updates. Plus stored bytes.Stored bytes only
Cloud StorageGoogleGit objects and lake files — stored bytes, operations, egress. Grows with history.Stored bytes only
Secret ManagerGoogleSecret versions held; access operations at runtime.Versions held
Cloud KMSGoogleKey versions held; encapsulation and signature operations. Rotation adds versions.Versions held
Cloud BuildGoogleBuild-minutes: the install, release cuts, and every container a strain builds for you.None when not building
Vertex AI (optional)GoogleTokens through the work-mode bus. Only in dual or work.None in home
Workstation VM (optional)GoogleRunning time plus persistent disk. Idles off.Disk only
Your Claude planAnthropicWhatever plan you already hold. Unchanged by installing this.Your plan's terms
Support & SLANobodyThere is none to buy. The repository is the support channel.None

Control

You set the caps, and you should set them on day one

An autonomous fleet with an uncapped budget is a bad idea regardless of whose software it is running. Google gives you the controls; use them before you give a strain work to do.

Set a budget with alert thresholds on the billing account, scope it to the project you installed into, and put a quota ceiling on the services that can run away — build-minutes and, if you turned it on, Vertex tokens. Then keep the work-mode switch in home until you have watched a full week of real usage and know the shape of your own curve.

# Point every control at the one project the harness installed into.
gcloud config set project YOUR_PROJECT_ID

# What has this project actually cost so far?
#  Billing console -> Reports -> filter by project -> group by service.
#  That grouping maps 1:1 onto the table above.

# Then: Billing -> Budgets & alerts -> Create budget
#  scope: this project
#  amount: a number you are willing to lose
#  thresholds: alert well below it, not at 100%
We cannot quote you a number, and we will not pretend to. Two installs of identical software can differ by orders of magnitude: one person asking a strain a question a day is not the same workload as a fleet grinding a work queue through Vertex around the clock and building a container every few minutes. Any figure on this page would be marketing, not information. Run it for a week with a budget alert set and read your own bill — that number is real and ours would not be.

The comparison

The alternative is renting

Compare this against the honest alternative, not against zero.

Renting an agent platform

You pay a monthly figure that is pleasantly predictable, and in exchange your agent state — memory, history, files, repositories, the transcript of everything your team asked and everything the fleet did — lives in somebody else's tenancy, under somebody else's IAM, on somebody else's retention schedule.

The predictable number is real. So are the things attached to it: pricing that changes on their schedule, terms that change on their schedule, an export path that is only as good as they chose to make it, an outage that is yours to explain and theirs to fix, and a compliance conversation that starts with "where is the data" and does not end with "in our project".

You are also, quietly, renting the right to leave. Every month of accumulated state raises the cost of that decision.

Growing your own substrate

You pay Google directly for resources in a project you already own, with a bill you can read by service and a budget you set yourself. The software is Apache-2.0 and free. State is encrypted at rest in your buckets and your Firestore, under keys in your KMS.

The number is less predictable — that is the honest trade, and it is why this page has no headline figure. In exchange there is no counterparty: nobody to change terms on you, nobody holding your history, nobody whose downtime is your incident, no export request to file. Deleting the project is a complete and immediate exit.

And if you decide tomorrow that this whole approach is wrong, uninstall.sh removes both services and you are left holding your own data in your own project. That is the version of "no lock-in" that survives contact with a lawyer.

Free software with a metered cloud bill is not cheaper than a subscription in every case. It is yours in every case. Decide which of those you are actually buying.