# Decentralized markets for AI agents: where to enter and what is actually free

Research date: **26 October 2026 (UTC)**. This is a desk-research report, not a transaction audit or a tested integration.

## Answer and recommended starting points

**Assessment:** Start with **x402 Bazaar** to discover tools an agent can buy, **Virtuals ACP** to sell or hire an agent for a defined job, and **Olas Mech Marketplace** for an existing ecosystem of agents purchasing specialist services. **Fetch.ai Agentverse** is useful for free discovery and limited hosting. **Nevermined** offers a managed payment/catalog route with a free starting allowance; **Akash** supplies decentralized compute rather than agent jobs. These choices serve different needs; none establishes a universally free market with guaranteed buyers.

The strongest concrete spending example found is Olas’s account of Pearl agents purchasing token-price data and language-model responses. Marketplace catalogs and payment protocols show what agents *can* buy; they do not by themselves show how much independent demand exists. Evidence and entry links follow below.

## Scope and evidence labels

Here, an agentic market means software can discover, purchase, or deliver a service under delegated authority. A market **for agents** may sell agent services to humans; a market **used by agents** enables machine purchasing. Both are covered, with priority on the latter.

- **Documented fact:** a capability or term stated in official documentation. This establishes what the provider documents, not successful execution here.
- **Provider-reported observation:** an operator’s account of actual usage or a published metric, not independently reproduced.
- **Inference / recommendation:** this report’s interpretation or proposed action.
- **Uncertain / unanswered:** conflicting material, missing evidence, or something requiring an account or live test.

“Decentralized” is assessed separately for discovery, settlement, and service execution. Open software, a token, or blockchain settlement alone does not demonstrate decentralized hosting, governance, or access.

## Platform comparison

All capabilities and prices below are documented claims as accessed on the research date. Classification of decentralization and practical fit is this report’s assessment.

| Platform and direct entry | What the market provides | What is free; what costs money | Decentralization and limits |
| --- | --- | --- | --- |
| **[x402 + CDP Bazaar](https://docs.cdp.coinbase.com/x402/buyer/discover-services)** | Machine-readable discovery of payment-gated services; protocol-based API purchases. | Public discovery needs no CDP key. x402 has zero protocol fees; services, networks, and facilitators can charge separately. | Open payment standard; CDP’s index is an operated catalog. An API behind a listing may be entirely centralized. [Protocol](https://x402.org/), [discovery docs](https://docs.cdp.coinbase.com/x402/buyer/discover-services). |
| **[Virtuals ACP / EconomyOS](https://os.virtuals.io/quickstart)** | Agent offerings, negotiated jobs, USDC escrow and delivery evaluation. | Paid jobs. Documented split: provider 95%, protocol 5%; with evaluator, provider 90%, evaluator 5%, protocol 5%. | On-chain job settlement; service execution and the CLI’s OAuth onboarding still have operational dependencies. [Job and fee concepts](https://os.virtuals.io/acp/concepts), [onboarding](https://os.virtuals.io/quickstart). |
| **[Olas Mech Marketplace](https://olas.network/mech-marketplace)** | Request and supply AI tools, predictions and other specialist services. | Providers charge for tasks; operating costs remain. Olas currently states a 15% marketplace fee is enabled. | Blockchain-mediated requests/payments with work performed off-chain. Fee activation is governance-controlled. [Market](https://olas.network/mech-marketplace), [fee disclosure](https://olas.network/). |
| **[Fetch.ai Agentverse](https://docs.agentverse.ai/documentation/getting-started/overview)** | Agent registration, search visibility, hosted or external agents. | Basic plan documents 8 hosted agents, 60,000 seconds of compute and 5 MB storage; quotas apply. Free mailbox access requires login at least every 30 days. | A hosted discovery/growth service connected to an agent ecosystem; evidence does not establish a fully decentralized storefront or completed autonomous payments. [Current v2 quotas](https://docs.agentverse.ai/documentation/advanced-usages/agentverse-subscriptions-and-quotas). |
| **[Nevermined](https://nevermined.ai/)** | Managed agent payments, spending mandates and an AI services catalog. | Free personal account and 2,000 starter catalog credits. Merchant settlement fees: 1% stablecoin, 2% card, plus processing costs. | Managed infrastructure supporting crypto and cards; should not be equated with a permissionless decentralized marketplace. [Current pricing](https://nevermined.ai/pricing/). |
| **[Akash Network](https://akash.network/)** | Marketplace for CPU, GPU, memory and storage to run workloads. | Paid compute leases; open access does not mean free hardware. | Independent infrastructure providers sell compute. This is an input market for agent builders, not proof of agents autonomously buying compute. [Provider explanation](https://akash.network/docs/providers/getting-started/should-i-run-a-provider/). |

### What “free” actually buys

**Documented fact:** x402’s protocol fee is zero, but Coinbase’s CDP Facilitator separately lists 1,000 free monthly on-chain transactions and $0.001 per transaction above that allowance. This is a facilitator-specific schedule, not universal x402 pricing. It does not include the seller’s service price. [CDP Facilitator pricing](https://docs.cdp.coinbase.com/x402/seller/facilitator).

**Documented fact:** Coinbase’s buyer quickstart uses test USDC on Base testnet and provides a faucet route. It requires CDP credentials; the TypeScript path specifies Node.js 22+. This is a way to learn payment mechanics without purchasing production services, not a free production wallet balance. [Buyer quickstart](https://docs.cdp.coinbase.com/x402/buyer/quickstart).

**Uncertainty resolved partly:** an older search-index version of Nevermined’s pricing emphasized a free account and settlement fees. The fetched page now also lists $20/month catalog credits or $10 packs, organization plans from $50/month, and a 2% buyer fee for managed external-service access. Buying from Nevermined merchants instead has no added buyer fee. Do not treat starter credits, personal payment plans, and paid catalog distribution as equivalent products. [Fetched pricing](https://nevermined.ai/pricing/). Its homepage says merchant catalog listing starts at $50/month. [Merchant entry and FAQ](https://nevermined.ai/).

**Inference:** For a zero-cash exploration, use public discovery, Agentverse’s documented basic allowance, and testnet exercises. No reviewed evidence supports unlimited free production inference, hosting, or outsourced labor. Provider allowances may change or require accounts.

## Where agents spend money: evidence versus possibility

| Spending category | Attributable evidence | What it establishes—and does not |
| --- | --- | --- |
| **Inference and current market data** | Olas’s 8 October 2026 post says Optimus, Basius, Omenstrat and Polystrat now obtain price data and model responses through Mechs; previously they used external x402 services. [Olas account](https://olas.network/blog/pearl-mech). | **Provider-reported usage**, stronger than a hypothetical integration. No independent receipts sampled here. A route change can increase marketplace activity without increasing underlying demand. |
| **Specialist predictions and delegated tasks** | Olas describes Mechs selling intelligence/data and publishes task-payment accounting. [Mech economy](https://olas.network/agent-economies/mech). | **Provider-reported paid activity.** Does not establish how many buyers are independent of the ecosystem or whether incentives cover their spending. |
| **Creative work and analysis** | ACP documentation describes purchasable offerings, requirements, deliverables and escrow; image generation and dataset analysis are examples. [ACP concepts](https://os.virtuals.io/acp/concepts). | **Documented capability and examples**, not independently verified sales of those examples. |
| **Search, research and translation** | Nevermined’s homepage illustrates Exa search, You.com research and DeepL translation; it explicitly labels its spending dashboard an example. [Nevermined](https://nevermined.ai/). | **Catalog/product evidence**, not transaction evidence. Example prices are not reliable live quotes. |
| **Runtime and GPU capacity** | Akash documents container deployment through configuration, competitive provider bids and lease acceptance. [Akash deployment overview](https://akash.network/). | **Documented compute purchasing workflow.** Whether an autonomous agent rather than a human initiated a given lease is unanswered. |
| **Trading capital and financial execution** | Virtuals documents swaps and Hyperliquid trading as well as commerce. [EconomyOS quickstart](https://os.virtuals.io/quickstart). | **Documented capability.** Capital deposited or traded is not marketplace service revenue; trading turnover is not evidence of useful agent labor. |

### How much spending is demonstrated?

**Provider-reported observation:** the Olas Mech page labels its 26 October 2026, 06:40 UTC figures as **$110,300.85 total task payments**, **$103,138.43 claimed payments / realized Mech earnings**, and **$907.77 marketplace fees collected**. Claimed payments and earnings are the same amount, not additive. Fees are a component of payments, not additional turnover; fee coverage excludes some token/chain combinations and begins 15 June 2026. These are cumulative figures with stated coverage, not current monthly revenue or a market-wide estimate. [Olas accounting disclosure](https://olas.network/agent-economies/mech).

**Research finding, with limits:** Ling, Zhou, Wu and Wang’s July 2026 arXiv preprint measures x402 settlements and cautions against interpreting count as adoption. Its abstract reports a 280-day Base window with about 136.7 million settlements and $44.1 million in value, classifying 21.20% of settlements as fictitious and 63.78% as internal to linked clusters. These are the authors’ classifications and historical sample, not this report’s audit or a current ecosystem fraud rate. The abstract does not establish peer review. [Original study and abstract](https://arxiv.org/abs/2607.12575).

**Inference:** There is evidence of paid agent-service activity, but insufficient comparable evidence to rank platforms by independent autonomous customer spend. Never add Olas requests, x402 settlements, token trades and compute leases into one market-size number: definitions differ and workflows may overlap.

## How to get there

These are proposed onboarding routes derived from documentation. No accounts, deployments or payments were created for this report.

### 1. Give an existing agent tools it can purchase

**Recommendation:** Start with [CDP Bazaar discovery](https://docs.cdp.coinbase.com/x402/buyer/discover-services). Search by the needed outcome, then constrain network, asset and maximum price. Inspect schemas and the reported recent calls/unique payers. Public discovery is available through REST, SDK or MCP. Treat usage fields as screening signals, not proof of buyer independence.

Follow the [testnet buyer guide](https://docs.cdp.coinbase.com/x402/buyer/quickstart): provision the documented wallet, obtain test USDC and complete a protected request. Before a production call, require a known recipient, expected chain/asset, per-call ceiling and total budget. Check both settlement evidence and output usefulness. These controls are recommendations, not a claim that all clients enforce them automatically.

### 2. Sell a small API or tool to agents

**Documented route:** deploy an HTTPS endpoint, add x402 payment handling and Bazaar metadata, supply input/output schemas and examples, validate it, then complete a settled call through CDP to trigger indexing. Discovery is not simply submitting a name to a website. CDP screens metadata; curation is separate from indexing. The documentation also supports testing discovery on Base Sepolia or Solana Devnet. [Seller discovery guide](https://docs.cdp.coinbase.com/x402/seller/get-discovered).

**Recommendation:** Start with one measurable output—such as a sourced data lookup—with a clear price, response schema, freshness timestamp and error behavior. Measure paid completions, repeat independent buyers and margin after upstream API/compute costs. Listing alone is not distribution or demand validation.

### 3. Hire or sell work requiring a job lifecycle

**Documented route:** use [EconomyOS onboarding](https://os.virtuals.io/quickstart) for sign-in, agent creation, signer approval and wallet funding. Browse agent offerings before creating a funded job. Tokenization is optional; there is no need to launch a tradable agent token to use jobs. [Token prerequisites and optionality](https://os.virtuals.io/agent-identity/token/overview).

For sellers, [the provider workflow](https://os.virtuals.io/acp/cli/provider-workflow) offers ACP Serve or an agent-driven lifecycle. Register an offering with price, SLA, requirements and deliverable; consume job events, propose a budget, wait for funding, perform the work and submit. Approval releases escrow; rejection returns it to the client. **Unanswered:** this review did not test disputes, rejection incentives, timeouts or fee behavior for failed jobs. Escrow is not a guarantee of output quality.

### 4. Supply or consume services in Olas

**Documented buyer route:** [Mech Client](https://stack.olas.network/mech-client/) supports listing Mechs, inspecting tools and schemas, and submitting requests. Its current supported client configurations are Gnosis, Base, Polygon, Optimism and Robinhood; payment types vary by Mech and chain. Choose a provider and inspect its rate/payment type before funding or requesting. Do not assume every chain accepts the same stablecoin.

**Documented seller route:** [Mech Server](https://stack.olas.network/mech-server/) describes installing the server, configuring a chosen chain/RPC, setting up a Mech, adding a custom tool, publishing metadata and running the service. It recommends initial testing on a virtual network. **Recommendation:** budget for RPC, upstream models/data, hosting, transaction costs and the disclosed marketplace fee; validate margins with a tiny bounded workload before maintaining a continuous service.

### 5. Explore with a free account or use managed payments

**Agentverse route:** [current documentation](https://docs.agentverse.ai/documentation/getting-started/overview) supports hosted agents or registering an existing agent for discovery through ASI:One. Publish accurate capabilities and inspect analytics. **Uncertainty:** the same overview describes monetization options but later labels monetization tools “Coming soon.” Treat it as a discovery/hosting starting point; confirm actual checkout and settlement availability before relying on revenue. The site itself returned HTTP 403 through the research tool, so account flows were not inspected.

**Nevermined route:** start with a personal account, connect the documented MCP endpoint or SDK, set a spending mandate and inspect catalog services. Its homepage says general card payment to any web merchant remains a roadmap item; do not infer universal purchasing from card support. [Connection routes and FAQ](https://nevermined.ai/). Confirm whether the intended seller needs a personal payment plan or paid catalog listing before budgeting.

**Akash route:** use [Akash Console via the official homepage](https://akash.network/) to specify a container and resources, compare provider bids and accept a lease. This is appropriate when compute is the actual need; it adds deployment/operations work compared with buying a single inference API call.

## Remaining questions and decision gates

1. **Buyer or seller?** For purchasing tools, begin with Bazaar; for selling complex jobs, compare ACP and Mech. For discovery without spending, begin with Agentverse/public catalogs. This is a recommendation based on documented workflow fit, not a revenue ranking.
2. **Which services have repeat independent customers?** Request dated receipts or anonymized buyer cohorts, and separate grants, internal transfers and incentive-funded use. This review did not reproduce such cohorts.
3. **What does the completed flow cost?** Obtain current quotes for service, facilitator, network, model/data, hosting and distribution fees. Public pricing cannot settle account-specific processing or funding costs.
4. **What degree of decentralization is required?** Verify catalog control, wallet/signing authority, facilitator alternatives, service hosting and evaluator control separately. No reviewed platform was demonstrated here to decentralize every layer.
5. **What can this requester access?** Geography, account eligibility, wallet preference, deployment skills, budget and service requirements were unspecified. No universal eligibility or guaranteed onboarding is claimed.

## Method and limits

Sources were searched and read on 26 October 2026. Technical and pricing claims rely on official project documentation/pages; the original x402 measurement preprint supplies a separate research perspective. Links are attached to the claims they support. Operator dashboards and blogs remain operator claims. Search snippets were used for discovery; where fetched pricing differed, the fetched page was preferred and the discrepancy noted.

Coverage is a practical shortlist, not an exhaustive census. No live paid calls, wallet attribution analysis, source-code audit, uptime study or revenue verification was performed. No global spend total, guaranteed earnings, or fully free production market is asserted. Product pages can change after this snapshot; actual entry requires rechecking the linked terms and the offered payment challenge.