The app.humantaste/taste-mcp MCP server provides expert, human-reviewed evaluation for AI agents. It enables pay-per-call human work delivered over MCP plus x402, with vetted humans reviewing AI agent outputs across multiple media types. The project is described as a public manifest for discovery, with no install steps required.

🛠️ Key Features

Human experts perform review of AI agent outputs
Pay-per-call human work over MCP + x402
Reviews supported across text, images, video, audio, and social posts
Public discovery via a server.json manifest; server source maintained privately
Manifest repository is hosted (nothing to install)

🚀 Use Cases

Human-in-the-loop quality or safety review for agent outputs
Content moderation-style review across mixed media
Evidence tracking for “on-chain proof of human review”

⚡ Developer Benefits

MCP registry discovery using server.json
Consistent review workflow for AI agent results via MCP
Topics include human-in-the-loop and human-judgment concepts

⚠️ Limitations

Server source is maintained privately; only the public manifest repository is described
Reviews are described as pay-per-call human work, with no deployment details beyond “hosted” and “nothing to install”

Taste

Expert review for AI agents. On-chain proof of human review.

This is the public manifest repository for the Taste MCP server. The server is hosted — there is nothing to install. server.json is the discovery manifest consumed by MCP registries; the server source is maintained privately.

what

Pay-per-call human work for AI agents over MCP + x402. Vetted humans review an AI agent's output across any media (text, images, video, audio, social posts), audit reasoning chains, run prepublish safety checks, gate high-stakes actions for human approval, and convene curated expert teams for working sessions on ideas, problems, or decisions. Approved outputs receive a Taste content certificate recorded on Base — any downstream agent can verify the review.

connect

code

MCP endpoint:  https://api.humantaste.app/mcp/
Transport:     Streamable HTTP (MCP 2025-06-18)
Payment:       x402, USDC on Base mainnet, $1.00–$15.00 per paid call
Free tools:    list_offerings, list_expert_profiles, get_result, verify_certificate
Typical wait:  5–30 min for evaluation calls; up to 24h for a think tank session

No authentication, no API key. Paid tools answer an unauthenticated call with HTTP 402; sign an EIP-3009 USDC transfer and retry.

tools

Free (no payment, no session):

Tool	Purpose
`list_offerings`	list offerings, prices, certificate flag, operating-hours state
`list_expert_profiles`	aggregate availability + estimated response time, no PII
`get_result`	poll session status by sessionId or order reference code; returns the full deliverable when complete
`verify_certificate`	read-only Taste content certificate lookup on Base, rate-limited 60/min/IP

Paid (x402 — USDC on Base):

Tool	Price	Purpose
`verify_external_source`	$1.00	verify whether an external source, project, or claim is trustworthy
`review_content`	$1.00	review AI-generated content for hallucinations, factual errors, domain mistakes — issues a certificate
`verify_reasoning_chain`	$1.00	audit a step-by-step reasoning chain for mid-chain errors — issues a certificate
`prepublish_review`	$1.00	review content for cultural sensitivity, brand safety, derivative risk before publishing — issues a certificate
`predict_audience_reaction`	$1.00	a human matching a target demographic rates and reacts to content
`arbitrate_dispute`	$1.00	an impartial human arbiter decides whether a deliverable meets a contract
`request_human_approval`	$1.00	pause a workflow for explicit human approval before an irreversible action
`order_think_tank_session_30`	$5.00	convene a curated team of human experts for a 30-min working session on an idea, problem, or decision; returns synthesized directions (1 free revision turn)
`order_think_tank_session_60`	$15.00	as above, 60-min session (2 free revision turns)

Each think tank order returns a TASTE-XXXX-XXXX-XXXX reference code. Two contextual helper tools — request_think_tank_revision and request_illustration_revision — spend a free revision turn against that code.

The live tool list is authoritative — server.json sets tools: "dynamic", so MCP clients call tools/list against the endpoint for the current surface.

example

bash

curl -X POST https://api.humantaste.app/mcp/ \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call",
       "params":{"name":"list_offerings","arguments":{}}}'

Paid calls return HTTP 402 with an X-Payment-Requirements header (base64 JSON). Sign EIP-3009 transferWithAuthorization for USDC on Base, then retry the same request with an X-Payment header.

compliance

EU AI Act Article 50(2) exempts AI-generated content from mandatory AI-disclosure labels when it has undergone human review or editorial control. Taste content certificates are the on-chain audit trail for that exemption.

sessionId	string	optional	The session ID returned by a paid tool call.
referenceCode	string	optional	An order reference code (TASTE-...) from a think tank or illustration order. Use this instead of sessionId to retrieve the result — it always returns the latest revision.

question	string	required	The specific question to ask the expert.
domain	string	required	Area of expertise needed (e.g. Farmer, Protestant priest, UX designer, Stockholm local, art curator, museum staff, culture journalist, food critic).
context	string	optional	Optional. Why you are asking and what you are working on.
referenceUrls	array	optional	Optional. Publicly accessible URLs the expert should review.
parentSessionId	string	optional	Optional. If present, the URL submits a follow-up to an existing consultation (capped at 3 per parent, same expert).

referenceCode	string	required	The TASTE-... reference code returned when you placed the think tank order.
adjustments	string	required	What you want the team to adjust, expand, or rethink.

referenceCode	string	required	The TASTE-... reference code returned when you placed the illustration order.
adjustments	string	required	What you want the illustrator to adjust, expand, or rethink.

subject	string	required	The URL, project name, or claim to verify. Be specific (e.g. a URL, a project name + token address, or a verbatim claim).
context	string	optional	Optional context. Use to clarify intent, constraints, audience, or anything that helps the expert evaluate.

Taste

🛠️ Key Features

🚀 Use Cases

⚡ Developer Benefits

⚠️ Limitations

Topics

list_offerings

list_expert_profiles

get_result

prepare_consult_domain_expert

request_think_tank_revision

request_illustration_revision

verify_external_source

review_content

verify_reasoning_chain

prepublish_review

predict_audience_reaction

arbitrate_dispute

request_human_approval

verify_certificate

order_think_tank_session_30

order_think_tank_session_60

consult_domain_expert

ask_a_human

review_plan

review_code

Taste

what

connect

tools

example

compliance

links

Install

Remote endpoint

Related servers

content	string	required	The content to review. Text directly, or a publicly accessible URL for non-text media (images, video, audio, posts).
context	string	optional	Optional context. Use to clarify intent, constraints, audience, or anything that helps the expert evaluate.

chain	array	required	Ordered array of reasoning steps. Each step: { step (number), reasoning (string), output (string), toolUsed? (string) }.
context	string	optional	Optional context. Use to clarify intent, constraints, audience, or anything that helps the expert evaluate.

contractDescription	string	required	The agreed contract or deliverable specification.
deliverable	string	required	The deliverable to assess. Text or a publicly accessible URL.
partyAPosition	string	required	One side of the dispute. The expert evaluates both impartially — order is arbitrary.
partyBPosition	string	required	The other side of the dispute.
context	string	optional	Optional context. Use to clarify intent, constraints, audience, or anything that helps the expert evaluate.
termsHash	string	optional	Terms anchor this dispute is bound to: keccak256 of the att/v1 terms manifest (as echoed from the x402 payment), or an opaque 32-byte document hash such as an LCP atrHash ("sha256:0x…" accepted). Persisted on the ruling for the receipt + precedent lookups.
termsUri	string	optional	Where the att/v1 manifest is hosted (https/ipfs). Convenience only — the hash is the identity.
termsJson	object	optional	The full att/v1 terms manifest object. If provided it must be a valid manifest whose keccak256 equals termsHash.

actionDescription	string	required	Describe the action awaiting approval. Include numbers (amounts, recipients, URLs) the human needs to see.
stakes	string	required	What kind of risk the action carries. Reversible: undoable. Irreversible: cannot be undone. Financial: moves value. Public: visible externally.
timeoutMinutes	integer	optional	Maximum minutes to wait for a human response before returning denied_timeout. Default 30, max 120.
context	string	optional	Optional context. Use to clarify intent, constraints, audience, or anything that helps the expert evaluate.

contentHash	string	optional	keccak256 hash of the content, as 0x-prefixed hex. Provide this OR certificateId.
certificateId	any	optional	Numeric certificate id. Provide this OR contentHash.

description	string	required	The idea, concept, or problem you want the think tank to work on. Be specific about the goal and any constraints.
mediaUrls	array	optional	Optional. Publicly accessible URLs to reference material (images, decks, docs, links) the team should review.
requestedDeliverables	array	optional	Optional. What you want out of the session — e.g. "5 campaign concepts", "a positioning statement", "product name ideas".
buyerWalletAddress	string	optional	Optional. An EVM wallet address. If provided, you can later retrieve the deliverable on humantaste.app by connecting this wallet.

plan	string	required	The implementation or project plan to review. Paste the full plan — steps, approach, architecture, sequencing.
goal	string	required	What the plan is meant to achieve — the objective or task it addresses.
context	string	optional	Optional context. Use to clarify intent, constraints, audience, or anything that helps the expert evaluate.