API and MCP

Use ChatGPT or Claude? Build your team there.

Tell it what takes up your time and it suggests agents. It tests them with cases from your business and starts them with your permission. Then ask your team for help in the same chat. Your developers can connect through the API too.

Start your assessment See what it can do

Illustrative product scenes. Sample businesses and data. Connecting to ChatGPT or Claude depends on each provider's acceptance.

How it works

Ask in the chat. You decide when it starts.

Scroll slowly. At each stop, see what Kharyo does and where you make the call.

01 · Connect

Choose which business it can see and what it can do.

Start in ChatGPT or Claude, then choose permissions in Kharyo: read about your business, suggest agents, test them, make them ready and start them.

ChatGPTClaudeAccess authorized by Gabriela Tovar

09:58

Business: Repuestos Tovar
Can: Read, suggest, test and start agents
Cannot: Connect accounts or approve payments
Expires: November 14

02 · Tell it

Tell it what takes up your time.

ChatGPT or Claude read how your business works and put together a proposal using the permissions you gave them. The rest stays out of reach.

ChatGPT

I need someone to send overdue invoice reminders and tell me when customers pay.

Kharyo2 steps

Repuestos Tovar: WhatsApp sales, Pago Móvil collectionsLearn about your business
Collections proposalPrepared the proposal

I suggest a Collections agent. Before making it ready, I'll test it with cases from your business.

Collections team

GI need someone to send overdue invoice reminders and tell me when customers pay.

KharyoLearn about your businessRepuestos Tovar: WhatsApp sales, Pago Móvil collections
KharyoPrepared the proposalCollections proposal

I suggest a Collections agent. Before making it ready, I'll test it with cases from your business.

03 · Kharyo tests

Collections gets tested before it's ready.

Kharyo tests it with cases from your business. The version that passes is the one that becomes ready.

Collections testRepuestos Tovar · 10:01

Invoice overdue by five days
A customer requests a payment arrangement
A payment appears in the bank
The invoice was already paid

Test: 4 of 4 cases passed

04 · Activate

With your permission, Collections starts working.

Making it ready and starting it are separate permissions. You connect Gmail yourself in Kharyo, never in the chat.

CollectionsVersion 1 · Repuestos Tovar

Active

Send overdue invoice remindersOn its own
Payment arrangementsAsks you
GmailConnect it in Kharyo

05 · Request work

Request work from the same chat.

Collections follows its usual rules and checks with you where you asked it to. Receiving a request is the first step. Check the result separately.

Work No. 4821Requested from ChatGPTClaude

Completed

08:30Sent reminders for 3 overdue invoicesFrom your email
08:41Andrés Pérez paid invoice 1043Confirmed by the bank
08:45Payment arrangement with Motos CaraboboWaiting for your approval

Illustrative product scenes and recreations of ChatGPT and Claude. Sample businesses and data. Connecting to ChatGPT or Claude depends on each provider's acceptance.

What can the app access?

Choose what it can do, one permission at a time.

Reading about your business, requesting work, suggesting agents, testing them and starting them need separate permissions. The app does what you selected for the business you chose. Try it here.

ChatGPT or Claude · Estudio NébulaLearn about your businessRead your profile, activity, and connected accounts, as in Kharyo's assessment. Your customers' phone numbers and payment references are hidden, and every read is recorded. Only the owner or an administrator can grant this.See my agents and their resultsSee the agents you have ready and how your requests are going.Request work and stop itSend new requests to your agents and stop pending ones.Suggest and adjust agentsPut together proposals and changes, with costs you approve.Test proposals and see the resultsTest a version with cases from your business and save the result.Make a tested version readyMake the version that passed ready. It hasn't started working yet.Start and pause agentsDecide when an agent receives work, according to its rules and budget.

Your rules stay the same

Whether you ask through the chat, the API or Kharyo, your agent follows the same rules and checks with you in the same cases.

Each connection sees one business

It stays linked to the business you chose, even if you switch businesses in Kharyo later.

Remove access whenever you like

The app loses access. Work already done stays saved.

Illustrative product scenes. Sample businesses and data.

Each version gets tested before it's ready and starts when you decide.

What if I change something?

Each change is prepared, tested and made ready. It starts when you say.

Each step has its own permission. ChatGPT or Claude can help you prepare an agent without having permission to start it.

01
Proposal
Prepare changes around your goal, accounts and business context. The agent doesn't talk to anyone yet.
Design and edit draftsPrepared the proposal
02
Test
Test the version that will become ready and save the result.
Test draftsTested: 4 out of 4 passed
03
Ready
The version you tested is now ready. It hasn't started working yet.
Publish versionsReady to use
04
Start working
Decide when it starts receiving work and pause it whenever you want.
Activate and pause agentsStarted working

A test with sample data shows how it would respond. See what happened with a real customer in the Room.

For your technical team

Have your own system? Connect it through API or MCP.

Your system asks an agent for work with a key for each request to prevent duplicates. Kharyo confirms it received the request. Check the result and its supporting details afterward.

Each access belongs to one business and one personAlways
Repeat a request with the same keyDoesn't duplicate work
Cancel workPreserves evidence

Request work from an agent

POST /v1/agents/work
Authorization: Bearer <workspace-token>
Content-Type: application/json
 
{
  "agent_reference": "<agent-uuid>",
  "request_key": "week-42-summary",
  "input": { "request": "This week's project summary" }
}

Example response

HTTP/1.1 202 Accepted
 
{ "data": {
    "accepted": true,
    "work_id": 4821,
    "status": "<status>",
    "next_action": "<next-step>",
    "completion_verified": false
} }

Technical documentation: routes, MCP tools, and connection

API routes

GET /v1/agents: Lists the agents in your workspace.
POST /v1/agents/work: Request work with agent_reference (UUID), request_key and input (object). Requires agents:execute. Returns 202 with accepted, work_id, status, next_action and completion_verified: false inside data.
GET /v1/agents/work/{workId}: Look up a numeric ID with agents:read. Returns the work's status, result, evidence, and timestamps.
POST /v1/agents/work/{workId}/cancel: Request work cancellation with agents:execute. Preserves case evidence.

MCP tools

list_agents
No arguments. Query published agents with mcp:read.
create_agent_draft
With mcp:design: launch_ref, request_key, proposal_json. Creates a disabled draft.
get_agent_draft edit_agent_draft
With mcp:design. Editing requires draft_reference, reviewed_digest and proposal_json.
test_agent_draft get_agent_test
With mcp:test. Testing requires draft_reference, reviewed_digest, request_key, mode (admission or synthetic) and input_json. A synthetic test does not prove a real operation.
publish_agent_draft
With mcp:publish: draft_reference, reviewed_digest, test_reference and acknowledge: true. Does not activate the agent or approve its actions.
activate_agent_draft
With mcp:activate: reference, digest, enabled and acknowledge: true. Enabling also requires current input_json and test_reference values. Does not create work.
create_agent_team publish_agent_team · activate_agent_team
Build, publish, and activate an agent team.
create_work get_work · cancel_work
Creating requires mcp:execute, agent_reference, request_key and input_json. Acceptance returns a work ID, with completion_verified: false. Check the result separately.

Connection and revocation

Consent: GET/POST /oauth/agents/authorize
Token exchange: POST /oauth/agents/token
Code authorization with PKCE S256; renewal using a refresh token.
Operational MCP: POST /mcp/agents
Apps: GET /settings/agent-oauth
DELETE /settings/agent-oauth/{reference}

Revoking access disconnects the app. Work already done stays in place. The connection requires a registered client.

Illustrative product scenes. Sample businesses and data. Connecting to external assistants such as ChatGPT or Claude depends on each provider's acceptance.

Stop doing
it all yourself.

Tell Kharyo what takes up your day. It suggests a team you can try for free.

Start your assessment View plans

Operational AI for Modern Business.