MCP server
Give your agents the Platform as tools: the models, Voices, Datasets and Fieldwork, the catalogue, and reads of your account and wallets. Agents use the services exactly as you do through the console and the API.
| |
|---|
| URL | https://api.tenzro.xyz/mcp |
| Transport | Streamable HTTP (stateless); a tool call that asks for progress is answered as an event stream |
| Auth | Your Platform API key: Authorization: Bearer tzl_…. Without it only catalogue tools work |
| Tools | 108 |
| Limits | 120 calls a minute per key; start tools for long jobs count as more than one call |
Provision an agent
You are responsible for what your agent can reach. It needs no account of its own: it uses a key you create and, if it pays or gets paid, a wallet you approve with your passkey.
- 1. Create a key with only the scopes it needs on the API keys page:
inference for Text, Embeddings and Forecasts; voices:speak, voices:batch, voices:train, voices:admin for Voices; datasets:read/datasets:write; fieldwork:read/fieldwork:write; read for account reads; wallets:read for wallet reads; manage to prepare orders and propose agents. A tool the key cannot use answers key_lacks_scope with the scope it needs; add it to the key and the next call works. Revoking the key stops the agent at once. - 2. Optionally, give it a wallet and limits. On the Agents page, create an agent: you approve it with your passkey, which gives it its own identity and wallet under yours, with per-payment, daily and total limits you set. Its key is issued from the agent’s page, and usage made with it counts against those limits. Agents always act under your passkey delegation; Tenzro Labs never provisions one on its own. See Wallets.
- 3. Connect the MCP server at
https://api.tenzro.xyz/mcp with the key, using one of the clients below.
Calls are charged like the same calls through the API, to the account the key belongs to: from the included allowance of your plan first, then your credits (Tenzro Labs credits, not a currency). Usage is attributed to the key and, for an agent key, to the agent. Every call is recorded per key (tool, outcome, time taken; never arguments or results); see it with list_mcp_calls.
Long jobs
Batches, custom voice training and exports start with a start_ tool that returns a job id at once. Its status tool (for example get_voice_batch) takes wait_seconds (up to 25) to wait for the job before answering, and sends progress notifications while it waits when your client asks for progress.
Connect a client
Claude Code
Cursor
~/.cursor/mcp.json or .cursor/mcp.json in a project:
VS Code
.vscode/mcp.json — prompts for the key instead of storing it in the file:
Claude Desktop and other stdio-only clients
Bridge to the HTTP server with mcp-remote in claude_desktop_config.json:
Test it
Try the models
The chat, embed and forecast tools call the same models as the API. Try them here as your account.
No model is serving right now, so there is nothing to try yet.
Tools
Text
Generate text and chat with the language models.
chatinference
Send messages to a Tenzro Labs language model and get its reply. Use list_models for model ids (modality text or text+image).
Charged per 1M input and output tokens, as on the API.
- model: string — Language model id from list_models
- messages: object[] — The conversation, oldest first
- max_tokens?: number — Longest reply, up to the model's limit
- machine?: string — Where to run the model; defaults to tenzro-partner-cluster
generateinference
Generate text from a prompt with a Tenzro Labs language model: one prompt in, the model's text out. Use chat for a conversation. Use list_models for model ids.
Charged per 1M input and output tokens, as on the API.
- model: string — Language model id from list_models
- prompt: string — What to write about or answer
- system?: string — Instructions that steer the model
- max_tokens?: number — Longest reply, up to the model's limit
- machine?: string — Where to run the model; defaults to tenzro-partner-cluster
Embeddings
Vectors for search and similarity, from text and, with multimodal models, images.
embedinference
Turn texts, and images where the model takes them, into unit-length vectors with a Tenzro Labs embedding model, for search and similarity. Pick the model from list_models (embedding models; multimodal ones accept images).
Charged per 1M tokens, as on the API.
- model: string — Embedding model id from list_models
- input: string | object | object[] — What to embed: strings, or parts { type: "text", text } and { type: "image_url", image_url: { url } } for multimodal models
- dimensions?: number — A shorter vector, where the model supports it
- input_type?: document | query — query for search queries; document otherwise
- machine?: string — Where to run the model; defaults to tenzro-partner-cluster
Voices
Speech, batches, detection and custom voices.
list_voice_modelsno key
Speech models with languages, licence and prices in Tenzro Labs credits.
list_voicesvoices:speak
Stock voices and your custom voices, with language and model.
- language?: string — BCP 47 tag, e.g. en or de-DE
- include_custom?: boolean — Include your custom voices (default true)
speakvoices:speak
Turn up to 5,000 characters into speech. Returns a short-lived audio link. Every clip is watermarked and carries C2PA content credentials.
Charged per 1,000 characters; custom voices at a premium.
- text: string — Up to 5,000 characters
- voice: string — Voice id from list_voices: a stock voice or one of your custom voices
- model?: string — Speech model id from list_voice_models; default the voice's model
- language?: string
- format?: wav | mp3 | flac | opus — Audio format (default wav)
Datasets
Create, fill, review, version and download datasets.
list_datasetsdatasets:read
Datasets on the account with kind, source, item counts and latest version.
- source?: fieldwork | voices | upload — Only datasets from this source
- limit?: number — Most results to return (default 50)
- cursor?: string — Continue after the `next` value of a previous page
get_datasetdatasets:read
One dataset with counts by item state and its versions.
- dataset_id: string — Dataset id
create_datasetdatasets:write
Create an empty dataset to add items to.
- name: string — Dataset name
- kind: audio | image | video | text — What the items are
- description?: string
list_dataset_itemsdatasets:read
Fieldwork
Consented data collection with adults.
list_fieldwork_projectsfieldwork:read
Collection projects with state, counts and the public page link.
- status?: draft | open | paused | closed
- limit?: number — Most results to return (default 50)
- cursor?: string — Continue after the `next` value of a previous page
get_fieldwork_projectfieldwork:read
One project: settings, notice, consent options, counts by state, and its dataset.
- project_id: string — Project id
create_fieldwork_projectfieldwork:write
Create a draft collection project. Participants must be 18 or over; purposes on the acceptable use policy's prohibited list are refused.
- name: string
- purpose: string — The purpose stated in the participant notice
- kind: audio | image | video | text
- consent_scopes
Catalogue
Models, machines, operators and prices. No key needed.
list_modelsno key
All models in the Tenzro Labs catalogue with pricing in Tenzro Labs credits, context length, availability and hosting machines. Charging starts when a payment rail is configured; until then usage is recorded at list price and charged nothing.
get_modelno key
One model's details plus its current route order (which machines requests go to, best first).
- model_id: string — Model id
list_machinesno key
The machines Tenzro Labs runs, with hardware, live health, serverless models, and rent terms and monthly plans where a machine offers them. Filter to narrow the list.
- service?: serverless | rent | subscribe — Only machines offering this
- region?: string — Region, e.g. Europe
- gpu?: boolean — true: GPU machines only; false: CPU-only
- operator?: string — Operator id
- model?: string — Only machines serving this model id
Account
Balance, activity, orders, plans and rentals.
get_accountread
The account this key belongs to, and the key's permissions and scope.
get_balanceread
Total credits, split into credit usable anywhere and credit pinned to specific machines.
list_activityread
Recent credit additions, plan credit and usage charges, newest first.
list_ticketsread
Credit and gift tickets this account bought, with codes and redemption status.
list_ordersread
Orders with status and payment method, newest first.
API keys
Read-only: keys are created and revoked by their owner, never by an agent.
list_api_keysread
API keys on the account with permissions, scope and last use. Secrets are never returned.
list_mcp_callsread
Recent MCP tool calls on the account, newest first: tool, key, agent, outcome and time taken. Arguments and results are not kept.
- key_id?: string — Only this key's calls
- limit?: number
Wallets
Read-only, with wallets:read. Your own wallet and agent wallets, and your Wallets org's agents, approvals and headless wallets. Wallets are non-custodial; nothing here signs or sends.
list_walletswallets:read
Your Tenzro wallet and the agent wallets you created under your passkey: address, state and limits.
get_walletwallets:read
One wallet: address, state and, for an agent wallet, the limits you approved and what it has spent today and in total.
- wallet_id: string — Wallet id from list_wallets: "account" or an agent id
list_wallet_agentswallets:read
Agents in your Wallets org: each one's DID, the person it is delegated from, state and approved limits.
- org?: string — The Wallets org (org_...), when your account administers more than one
get_wallet_agentwallets:read
One agent in your Wallets org: state, limits, template and wallet address.
- did: string — The agent's DID, from list_wallet_agents
- org
Agents
Every change returns a link the owner approves with their passkey.
get_identityread
The account's did:tenzro:human identity and wallet (rooted in the user's passkeys): state, DID, address, devices, and whether sending is enabled.
list_agentsread
Agents the user delegated with their passkey: status, DID, wallet, the limits the owner approved, and any approval still waiting. Collects outstanding approvals first.
get_agentread
One agent, including what it has spent today and in total. Collects an outstanding approval first, so call it after the user says they approved.
- agent_id: string — Agent id
create_agentmanage
Propose an agent under the user's identity. Returns { approve_url, agent }. The USER must open approve_url and approve with their passkey on the account site, where they may also adjust the settings; you cannot approve for them, so relay the link. Once they have, get_agent shows it active with the settings they approved. Refused with no_serving_nodes while the network has none.
- name: string — 1 to 64 lowercase letters, digits, dots, dashes or underscores
Orders and plans
Prepare orders for a person to pay. No tool moves money.
create_ordermanage
Create an unpaid order for credits, a package, a rental or a plan. Returns the order and a pay_url a person can open. Confirm the amount with the user first (see quote).
- type: credits | package | rental | subscription — What to buy
- amount_usd?: number — credits: the fiat amount to pay, in whole US dollars. Refused while Tenzro Labs has not set the credit price
- machine_id?: string — credits: pin to a machine. rental/subscription: the machine
- gift?: boolean — credits/package: issue a redeemable ticket code instead of crediting the account
- package_id?: string — package: package id from list_packages
- gpu?: number — rental: GPUs
- cpu_cores?: number — rental: vCPUs
- memory_gb?: number — rental: memory in GB
- storage_gb?: number — rental: storage in GB
- hours?: number — rental: term in hours
- plan_id?: string — subscription: plan id from the machine's plans
- renew_subscription_id?: string — subscription: renew this stablecoin plan for 30 more days
Try asking
- “Which machines are online right now, and which models is each one serving?”
- “Which models take image input, and how long is their context?”
- “How much credit do I have left, and what did I spend it on this week?”
- “Embed these three passages with Qwen3-Embedding and tell me which two are closest.”
- “Read these three lines aloud in a calm voice and give me the audio link.”
- “How many submissions are waiting for review in my fieldwork project?”
Safety
No tool moves money: pay_order returns a Stripe Checkout link or a deposit address for a person to act on, and wallets are only read. Agent changes wait for your passkey. Keys are read-only over MCP. Tools that change things are marked non-read-only so clients ask before running them, and deletions are marked destructive. Give an agent the fewest scopes it needs.
speed?: number — 0.5 to 2 (default 1)detect_voicevoices:speak
Check audio for the Voices watermark and C2PA credentials. Give audio_url or audio_base64.
Charged free or nominal.
- audio_url?: string — A link to the audio
- audio_base64?: string — The audio itself, up to 10 MB
start_voice_batchvoices:batchstarts a job
Speak many texts at once, at a lower rate. Returns a job; follow it with get_voice_batch.
Charged per 1,000 characters at the batch rate.
- voice: string — Voice id from list_voices: a stock voice or one of your custom voices
- model?: string — Speech model id from list_voice_models; default the voice's model
- format?: wav | mp3 | flac | opus — Audio format (default wav)
- items?: object[] — Texts to speak, up to 1,000. Or give dataset_id
- dataset_id?: string — A text dataset to speak instead of items
- dataset_version?: number
- output_dataset_name?: string — Save the clips as a new dataset with this name
get_voice_batchvoices:batch
Where a batch stands: progress, credits used, and links or the output dataset when done.
- batch_id: string — Batch id
list_voice_batchesvoices:batch
Batches, newest first.
- limit?: number — Most results to return (default 50)
cancel_voice_batchvoices:batch
Stop a running batch. Clips already made are kept and charged.
- batch_id: string — Batch id
list_custom_voicesvoices:speak
Your private voices and languages: base model, dataset version, state.
- limit?: number — Most results to return (default 50)
quote_custom_voicevoices:train
How many dataset items may be used (consent includes training), the estimated training time and credits. Charges nothing.
- dataset_id: string — Dataset id
- dataset_version?: number
- base_model: string — Base speech model id; only models licensed for fine-tuning and commercial use are offered
- language: string — BCP 47 tag of the voice or language to train
start_custom_voice_trainingvoices:trainstarts a job
Fine-tune a base model on a dataset version into a private voice or language. Only items whose consent includes training are used; withdrawn items never are. Returns a job; follow it with get_custom_voice_training.
Charged per GPU-hour of training.
- name: string
- dataset_id: string — Dataset id
- dataset_version?: number — Default the latest version
- base_model: string
- language: string
- settings?: object — Training settings, as on the console
get_custom_voice_trainingvoices:train
Where a training job stands: progress, credits used, and the voice id once ready.
- job_id: string — Training job id
cancel_custom_voice_trainingvoices:train
Stop a training job; it can be resumed. Time already used is charged.
- job_id: string — Training job id
resume_custom_voice_trainingvoices:train
Resume a stopped training job from its last checkpoint.
- job_id: string — Training job id
delete_custom_voicevoices:admin
Delete a custom voice and its model. Cannot be undone.
- voice_id: string — Custom voice id
Items with media type, size, hash, state, consent reference and provenance. Paged.
- dataset_id: string — Dataset id
- status?: pending | accepted | rejected | withdrawn — Only items in this state
- limit?: number — Most results to return (default 50)
- cursor?: string — Continue after the `next` value of a previous page
add_dataset_itemdatasets:write
Add one item. Send small content inline as content_base64, or leave it out to get an upload_url to PUT the bytes to before it expires. New items start pending.
Charged dataset storage per GB-month.
- dataset_id: string — Dataset id
- mime: string — Media type, e.g. audio/wav or text/plain
- bytes: number — Size of the item in bytes
- sha256?: string — Hex SHA-256 of the content, checked on upload
- filename?: string
- meta?: object — Free-form metadata stored with the item (labels, transcript, ...)
- content_base64?: string — The content itself, for items up to 5 MB. Leave out to get an upload_url instead
set_dataset_item_statusdatasets:write
Accept or reject a pending item. Only accepted items go into versions.
- dataset_id: string — Dataset id
- item_id: string — Item id
- status: accepted | rejected
- reason?: string
create_dataset_versiondatasets:write
Freeze the accepted items into a numbered version with a manifest (JSONL plus metadata, consent and provenance). Withdrawn items are never included.
- dataset_id: string — Dataset id
- note?: string — What changed in this version
list_dataset_versionsdatasets:read
Versions of a dataset, newest first, with item counts.
- dataset_id: string — Dataset id
get_dataset_downloaddatasets:read
Short-lived links to a version's manifest and items, or to one item.
Charged dataset egress per GB.
- dataset_id: string — Dataset id
- version?: number — A version's manifest (JSONL + metadata incl. consent and provenance); default the latest
- item_id?: string — One item's content instead of a version manifest
delete_datasetdatasets:write
Delete a dataset, its items and versions. Cannot be undone.
- dataset_id: string — Dataset id
: purpose | training | cloning | public_release | third_party_sharing[] — What participants are asked to agree to; each is a separate opt-in. purpose is always asked
description?: string — What the project is for, shown to participantsprompts?: string[] — What participants record or write, one per tasktarget_items?: number — Stop accepting once this many items are acceptedlanguages?: string[] — BCP 47 language tagscountries?: string[] — ISO country codes participants may take part frommin_age?: number — Minimum age; never below 18retention_months?: numberdataset_id?: string — Collect into this dataset; default a new oneupdate_fieldwork_projectfieldwork:write
Change a project's settings; fields left out keep their values. Consent texts of a published project are versioned.
- project_id: string — Project id
- name?: string
- description?: string — What the project is for, shown to participants
- prompts?: string[] — What participants record or write, one per task
- target_items?: number — Stop accepting once this many items are accepted
- languages?: string[] — BCP 47 language tags
- countries?: string[] — ISO country codes participants may take part from
- min_age?: number — Minimum age; never below 18
- retention_months?: number
- dataset_id?: string — Collect into this dataset; default a new one
publish_fieldwork_projectfieldwork:write
Open the project's public participant page. Returns its link.
Charged per published project per month.
- project_id: string — Project id
set_fieldwork_project_statusfieldwork:write
Pause or resume taking part, or close the project for good.
- project_id: string — Project id
- status: open | paused | closed — open resumes a paused project
list_fieldwork_submissionsfieldwork:read
Submissions with state, reviews and a short-lived preview link. Paged.
- project_id: string — Project id
- status?: pending | accepted | rejected | withdrawn
- limit?: number — Most results to return (default 50)
- cursor?: string — Continue after the `next` value of a previous page
review_fieldwork_submissionfieldwork:write
Vote to accept or reject a submission; it is decided once enough reviews agree. Accepted items go into the project's dataset.
Charged per accepted item.
- project_id: string — Project id
- submission_id: string — Submission id
- decision: accept | reject
- reason?: string — Reject reason id or short text
list_fieldwork_participantsfieldwork:read
Pseudonymous participants with consent version, scopes granted and counts. No personal data.
- project_id: string — Project id
- limit?: number — Most results to return (default 50)
- cursor?: string — Continue after the `next` value of a previous page
start_fieldwork_exportfieldwork:writestarts a job
Export accepted items as a new dataset version. Returns a job; follow it with get_fieldwork_export.
- project_id: string — Project id
- dest_country?: string — Where the data goes, when outside the EEA
- transfer_mechanism?: string — Transfer mechanism for a non-EEA destination
get_fieldwork_exportfieldwork:read
Where an export stands, and the dataset version once it is done.
- export_id: string — Export id
list_fieldwork_ingest_keysfieldwork:read
The project's ingest keys (browser-safe keys for its own page): name, prefix, allowed origins, last use, revoked. Never the secret.
- project_id: string — Project id
create_fieldwork_ingest_keyfieldwork:write
Create a key for the project's own web page or app to send recordings through the ingest API. It works only for this project and the listed origins, and cannot read data. The secret is in this result only; it is not shown again.
- project_id: string — Project id
- name?: string — A label, such as the page or team that uses the key
- allowed_origins: string[] — Page origins the key works from, as https://host[:port] (http only on localhost); no wildcards or paths
update_fieldwork_ingest_key_originsfieldwork:write
Replace the page origins an ingest key works from.
- project_id: string — Project id
- key_id: string — Ingest key id
- allowed_origins: string[] — Page origins the key works from, as https://host[:port] (http only on localhost); no wildcards or paths
revoke_fieldwork_ingest_keyfieldwork:write
Stop an ingest key; the page using it is refused on its next request.
- project_id: string — Project id
- key_id: string — Ingest key id
get_machine
no key
One machine's hardware, health, rent rates and availability, and plans.
- machine_id: string — Machine id
list_operatorsno key
Who runs the machines, and which machines they run: operators on the Tenzro Network that partner with Tenzro Labs.
get_operatorno key
One operator's profile and machines.
- operator_id: string — Operator id
route_modelno key
Which machines a request for a model would be routed to right now, in priority order.
- model_id: string — Model id
- region?: string — Preferred region
list_packagesno key
Credit packages and offers, if any: price paid and credit received. Empty when none are offered.
list_payment_methodsno key
Whether card (Stripe) and stablecoin (Bridge) payments are available, and accepted chains and tokens.
quoteno key
Price credits, a package, a rental or a plan without creating an order. Use before create_order to confirm cost with the user.
- type: credits | package | rental | subscription — What to buy
- amount_usd?: number — credits: the fiat amount to pay, in whole US dollars. Refused while Tenzro Labs has not set the credit price
- machine_id?: string — credits: pin to a machine. rental/subscription: the machine
- gift?: boolean — credits/package: issue a redeemable ticket code instead of crediting the account
- package_id?: string — package: package id from list_packages
- gpu?: number — rental: GPUs
- cpu_cores?: number — rental: vCPUs
- memory_gb?: number — rental: memory in GB
- storage_gb?: number — rental: storage in GB
- hours?: number — rental: term in hours
- plan_id?: string — subscription: plan id from the machine's plans
- renew_subscription_id?: string — subscription: renew this stablecoin plan for 30 more days
An order's current status, checking with the payment processor if it is still settling.
- order_id: string — Order id
list_subscriptionsread
Monthly plans on the account with status, renewal and period end.
list_rentalsread
Rentals on the account with resources and term.
?: string — The Wallets org (org_...), when your account administers more than one
list_wallet_approvalswallets:read
Changes and held actions in your Wallets org that wait for, or were decided by, a person's passkey.
- status?: pending | approved | rejected | expired | failed — Only approvals in this state
- org?: string — The Wallets org (org_...), when your account administers more than one
get_wallet_approvalwallets:read
One approval: its kind, state and, for an approved held action, the step-up to send.
- approval_id: string — Approval id, from list_wallet_approvals
- org?: string — The Wallets org (org_...), when your account administers more than one
list_headless_walletswallets:read
Headless wallets in your Wallets org: keys held in each machine's TPM or Secure Enclave, each controlled by a person's passkey.
- org?: string — The Wallets org (org_...), when your account administers more than one
get_headless_walletwallets:read
One headless wallet: state, address, controlling identity and the machine's hardware backend.
- headless_id: string — Headless wallet id, from list_headless_wallets
- org?: string — The Wallets org (org_...), when your account administers more than one
capabilities
?: inference | payments[] — Default: [inference]
max_transaction_tnzo?: string — Network cap on any one payment, TNZO, e.g. "1.5"max_daily_tnzo?: string — Network cap per day, TNZOdaily_limit_credits?: number — Tenzro Labs credits this agent may spend per UTC daytotal_limit_credits?: number — Tenzro Labs credits this agent may spend in totalexpires_in_days?: number — The delegation ends after this many daysupdate_agentmanage
Propose new limits or settings for an active agent; fields left out keep the values the owner approved last. Returns { approve_url, agent }. Nothing changes until the USER approves at approve_url with their passkey; relay the link.
- agent_id: string — Agent id
- capabilities?: inference | payments[] — Default: [inference]
- max_transaction_tnzo?: string — Network cap on any one payment, TNZO, e.g. "1.5"
- max_daily_tnzo?: string — Network cap per day, TNZO
- daily_limit_credits?: number — Tenzro Labs credits this agent may spend per UTC day
- total_limit_credits?: number — Tenzro Labs credits this agent may spend in total
- expires_in_days?: number — The delegation ends after this many days
revoke_agentmanage
Propose revoking an agent. Returns { approve_url, agent }. Once the USER approves at approve_url with their passkey, its identity is revoked on the network and its API keys stop working; relay the link.
- agent_id: string — Agent id
pay_ordermanage
Start payment. card: returns a Stripe Checkout URL for the user to open. stablecoin: returns the chain, token, exact amount and deposit address for the user to send from their wallet. Never pays by itself.
- order_id: string — Order id
- method: card | stablecoin
- chain?: string — stablecoin: e.g. base, ethereum, solana
- token?: string — stablecoin: usdc or usdt
redeem_ticketmanage
Redeem a ticket code (TZ-XXXX-XXXX-XXXX-XXXX) into this account's balance.
cancel_subscriptionmanage
Stop a card plan from renewing. It stays active until the end of the current period.
- subscription_id: string — Subscription id
resume_subscriptionmanage
Undo a pending cancellation on a card plan.
- subscription_id: string — Subscription id
claude mcp add --transport http tenzro-labs https://api.tenzro.xyz/mcp \
--header "Authorization: Bearer tzl_..."
{
"mcpServers": {
"tenzro-labs": {
"url": "https://api.tenzro.xyz/mcp",
"headers": {
"Authorization": "Bearer tzl_..."
}
}
}
}
{
"inputs": [
{
"type": "promptString",
"id": "tenzro-key",
"description": "Tenzro Labs API key",
"password": true
}
],
"servers": {
"tenzro-labs": {
"type": "http",
"url": "https://api.tenzro.xyz/mcp",
"headers": {
"Authorization": "Bearer ${input:tenzro-key}"
}
}
}
}
{
"mcpServers": {
"tenzro-labs": {
"command": "npx",
"args": [
"-y",
"mcp-remote",
"https://api.tenzro.xyz/mcp",
"--header",
"Authorization:${TENZRO_AUTH}"
],
"env": {
"TENZRO_AUTH": "Bearer tzl_..."
}
}
}
}
npx @modelcontextprotocol/inspector # then connect to https://api.tenzro.xyz/mcp with the header above