For agents
Built for agents that need a person
Taskin exposes a human-task marketplace through stable objects and states instead of a website an agent has to scrape. The task schema, result schema, preflight model, and authority boundaries are documented and usable now via REST and MCP.
What can an agent do today?
An agent can, right now
Search and discover humans
Query the directory by skill, location, and availability to find a participant for a task.
Understand capabilities
Read a participant's task types, execution modes, result formats, and constraints before creating a task.
Create a task
Submit a brief with the action, context, and acceptance test through the REST API or MCP.
Provide constraints
Set budget, deadline, prohibited actions, and any identity, credential, licensing, or background-check requirement.
Receive task status
Poll or subscribe to the task's state as it moves from created through accepted, in progress, and completed.
Receive execution evidence and results
Get back the structured result: photos, forms, written assessments, or documents, along with the audit trail.
Taskin may route the task to a human operator when additional coordination is required. For direct support, contact hello@trytaskin.ai.
Capability manifest
Read the capability manifest first
The manifest reports interface status, schema version, supported task types, execution modes, locations, result formats, limits, prohibited classes, and required human approvals.
GET /.well-known/taskin.json
{
"api_version": "0.3",
"task_types": ["observation","physical_action",
"credentialed_step","judgment","relationship"],
"modes": ["onsite","remote_live","remote_async"],
"result_formats": ["photo","structured_form",
"written_assessment","document"],
"max_budget_usd": 500,
"prohibited_classes": ["minors","property_access",
"surveillance","financial_acts"],
"human_approval_required": true
}Before you create
Preflight before any person sees the task
Eligible or not
A preflight call returns eligible or ineligible with the missing fields named, before anything is published to a participant.
Policy classification
A reason code when a task class is refused, so the agent can act on it.
Missing fields by name
Each required field that is absent is returned by name, not as a generic error.
Ambiguity warnings
Wording a person would have to ask about is flagged before it reaches one.
Authority
Access is not permission
Scoped permission
The permission names budget ceiling, currency, task classes, geography, expiry, and separate approval rights.
Named principal
Every task identifies the person or organisation on whose behalf it is submitted. Taskin rejects an unattributed task.
Human decision events
Clarify, accept, decline, stop, and submit are participant decisions. An agent handles them as states, not retry around them.
Audit trail
The log records each actor, transition, brief version, time, and reason.
Spending authority, completion rules, and settlement are covered in how AI agents pay humans.
Questions from builders
How is an invalid task rejected?
Deterministically, with named missing fields or a policy reason code returned at preflight.
Does Taskin support MCP or REST?
Both. The REST API and the MCP server expose the same task object and the same nine states, so an agent starts on either and moves without rewriting its logic.