Toloka
A self-serve platform where vetted contributors record first-person task video to your brief.
Self-serve API, key required, irlpls adapter · docs
- Mode
- Made to order · hours
- Coverage
- 100+ countries
- Price
- Price not stated
- Key
- Required
- Resale
- Not allowed
- AI training
- Allowed
Eight links
1 of 8 links claimed · 0 evidenced
Compiled from public sources, unclaimed. What they mean
Each link
- 1. Real personClaimed
Person · Claimed. Contributors must pass Toloka's identity, residency and eligibility verification (an identity document showing country of residence); labeling tasks default to experts who have passed ID verification.
Still open: Whether the verified account holder is the person who filmed each clip, how multi-account use is detected, and whether collection templates (not only labeling nodes) require ID-verified contributors. No public evidence. The eligibility terms name no identity-verification vendor ('our identity verification provider') and document no liveness check or re-verification; an account can be shared after enrolment.
- 2. Captured locallyNot stated
Camera · Not stated
Still open: No check that a video came from a real camera rather than a screen, edit or generator is described; HomER rejected re-used clips by rule, not by a documented detector.
Place · Not stated
Still open: No location capture or check is documented. The audience filter restricts contributors by country of residence, not where a clip was recorded.
- 3. Genuine deviceNot stated
App · Not stated
Still open: Contributors record with their own smartphones or cameras and upload files; no capture-app or device attestation is documented.
- 4. Sealed captureNot stated
Camera · Not stated
Still open: No check that a video came from a real camera rather than a screen, edit or generator is described; HomER rejected re-used clips by rule, not by a documented detector.
Content · Claimed. An LLM reviewer checks every submission against the requester's written quality criteria (e.g. head-mounted, both hands visible 95%+ of the time, no third-person footage), with a verdict and reason per criterion, and watches video fields; rejected work goes back for correction or reassignment. Optional human review checks a 10 to 100% sample. Toloka cites 89.1% accuracy at catching failures.
Still open: How the 89.1% figure was measured, the false-accept rate, and whether people review anything the model passes when human review is not switched on.
Counts only with a stated signing mechanism (C2PA, a manifest).
- 5. Known timestampNot stated
Time · Not stated
Still open: No trusted capture time is documented.
- 6. Tracked in transitNot stated
Custody record (optional field) · Not stated
- 7. Independently verifiedNot stated
Cross-check · Not stated
Still open: No independent cross-check of a collected recording is described. Multi-contributor agreement (Quorum nodes) is documented for labeling, not for confirming a capture.
- 8. Delivered to youNot stated
Delivery receipt (optional field) · Not stated
API
Needs a sales or approval step first.
- base
- https:/
/ platform.toloka.ai/ api/ v2-beta - auth
- API key · Authorization +2 more
- access
- Self-serve
- adapter
toloka- mcp
- https:/
/ mcp.tendem.ai/ mcp - pay
- card, prepaid balance
- GET /
projects - POST /
pipelines/ {pipelineId}/ cost-estimate - POST /
pipelines/ {pipelineId}/ runs - GET /
runs/ {runId} - GET /
runs/ {runId}/ items
Rate limit. 60 requests per minute. Per endpoint; RateLimit-* headers; 429 when exceeded. A run processes at most 10,000 items. source
Sources
22 sources
- Self-serve egocentric video collection: describe the task, set a quality bar; no sales cycle, no minimums; 200,000+ contributors, 100+ countries; 3 hrs to first batch, results in 24 hours; LLM QA validates every submission, 89.1% accuracy at catching failures; HomER: 63 videos, 17 categories, 3 hours, $50 toloka.ai
- HomER built with the self-service platform; head-mounted forward-facing camera, both hands visible 95%+ of the time, handheld footage rejected; verified crowd in 100+ countries and 40 languages; 1,800+ contributors matched; LLM QA rejects or flags failures against rubrics toloka.ai
- Projects run via API, TypeScript SDK and CLI; Pipeline API in beta; price calculated after setup, no hidden fees, minimums or contracts; no onboarding; card via Stripe, bank transfer on request; $50 credit match; egocentric video and utterance recording listed; custom data collection on request; agent bootstrap at platform.toloka.ai/agents/SKILL.md toloka.ai
- API base URL, 'Authorization: ApiKey' header, key generated in the Business Console; long-running jobs polled; 60 requests/minute per endpoint; 10,000 items per run; unverified teams are on a trial and cannot start a run platform.toloka.ai
- OpenAPI 3.1 for the beta Pipeline API: apiKeyAuth (Authorization) and scopedTokenAuth (X-Scoped-Token); endpoints for projects, cost-estimate, runs, run items, exports platform.toloka.ai
- CLI: toloka login --browser, API key alternative for scripts and agents, TOLOKA_API_KEY platform.toloka.ai
- TypeScript SDK @toloka-platform/sdk, 0.x platform.toloka.ai
- Agent skill: never ask a person to paste an API key; show the plan or maximum charge before a paid mutation platform.toloka.ai
- Run pricing: per-task price plus fixed 33% commission, LLM QA per call, human QA per review, video tool base plus per-minute; Expected and Maximum forecast; price frozen at deploy; pay-as-you-go from balance; run pauses when funds run out platform.toloka.ai
- Prepaid team balance in USD; top-up and payment methods in the Business Console; platform takes no payments platform.toloka.ai
- Data collection templates: Egocentric Activity Video Collection (50 to 1,000 videos, 13 default household categories, audience countries and languages); data collection covers video, audio, images, in-the-field capture; deploy takes 5 to 15 minutes platform.toloka.ai
- LLM review grades every item against written criteria; optional human review of a 10 to 100% sample; rejected expert work returned or reassigned; Quorum nodes for multi-expert agreement platform.toloka.ai
- Audience by language, country and specialization; labeling tasks default to ID-verified experts; export links expire after 30 minutes platform.toloka.ai
- Contributor eligibility: restricted and unsupported jurisdictions (incl. Switzerland, Taiwan, Kenya); identity, residency and eligibility verification with an identity document; project-level country limits toloka.ai
- Self-Service Agreement (22 Apr 2026): API use (§2.4); prepaid balance, purchases final (§4); KYB and sanctions screening before paid orders (§5.3); trial Data and Output licensed to Toloka perpetually, opt-out does not apply (§5); no reselling Output, no automated accounts, no competing-model training (§6.2); buyer owns paid Output subject to Seller Background IP (§7.1, §7.2); contributors assign rights (§7.3); Toloka trains on Data and Output unless opted out; opt-out written, prospective, excludes derived Toloka Data (§7.4, §7.9); project moderation (§3.8); EDD (§2.2) toloka.ai
- Contributor posting: smartphone or camera at 1080p/30fps or better, mounted at forehead or eye level, recording household activities www.builtinaustin.com
- Older field-task docs (geotagged photos in the Toloka mobile app) are no longer current; the old docs path redirects to the homepage toloka.ai
- Tendem MCP in the official registry: io.github.Toloka/tendem-mcp 1.0.1 active, remote streamable-http at mcp.tendem.ai/mcp registry.modelcontextprotocol.io
- Tendem MCP protected-resource metadata: authorization server auth.tendem.ai/realms/tendem mcp.tendem.ai
- Tendem authorization server publishes a registration endpoint and PKCE S256 auth.tendem.ai
- Tendem MCP: OAuth on first use or 'Authorization: ApiKey' for headless use github.com
- Identity document must reflect country of residence; checked by 'our identity verification provider' toloka.ai
Notes
About. Self-serve data platform: a requester describes a collection task, an AI assistant builds the pipeline (instructions, contributor audience by country and language, price, quality criteria), and ID-verified contributors record and upload the data, checked by an LLM reviewer and optionally by human reviewers. The documented collection template is egocentric (first-person, head-mounted) household-activity video; other collection (images, audio, in-the-field capture) is built from a description. Funded from a prepaid balance topped up by card; projects, runs and exports can also be driven through a beta REST API, a TypeScript SDK and a CLI. Toloka's older mobile field-task app (geotagged photos of places) is no longer documented and is not described here.
Price. from_usd 0 means no unit price is published. You set the per-task price paid to each contributor; Toloka adds a fixed 33% commission, plus LLM QA per review call, human QA per sampled review (payout scaled to review time plus 10%), and a video-review charge (base rate plus per-minute rate) where the LLM reviewer watches video. The real per-task and per-call prices appear only in the in-app cost breakdown. Deploying freezes the run's prices. No minimums or contracts. HomER (63 head-mounted household videos, 17 categories) cost about $50 in total; that is one example, not a rate. New accounts get $50 credit after the first $50 of qualifying spend.
Coverage. Toloka cites 200,000+ contributors in 100+ countries and 40 languages, but publishes no country list, only exclusions: contributors cannot be resident in sanctioned jurisdictions (e.g. Russia, China, Iran, Belarus, Cuba, North Korea, Syria) or in places its ID-verification or payment providers do not cover (including Switzerland, Taiwan and Kenya). The requester can limit the contributor audience by country and language. Recording happens wherever contributors live (mostly their homes), not at a place the requester picks.
Timing. Egocentric video page: '3 hrs' to first batch, 'production-ready results in 24 hours'.
Licence. Self-Service Agreement (22 Apr 2026). §7.1: the buyer owns Output created specifically for it once Fees are fully paid (§7.2 excludes the seller's Background IP); §7.3: contributors must assign those rights to the buyer. The terms assume buyers train AI models on Output (§2.6, §8.4.1) but contain no express training grant but forbid using it to build or train products that compete with Toloka (§6.2(vii)). §6.2(iv) forbids reselling Output, and §6.2(iii) forbids making it available for the benefit of unaffiliated third parties. §7.4/§7.9: Toloka may train its internal models on your Data and Output unless you opt out in writing; the opt-out is prospective only and does not cover derived de-identified 'Toloka Data'. §5: Data and Output from a free trial are licensed to Toloka irrevocably and perpetually for any purpose, including model training, and the opt-out does not apply.
Compiled from public sources. Is this your company? Publish your card ›
Compiled from public sources · updated 2026-10-28JSONSiteCompare