API & MCP

What you can do

Run meetings where several AIs deliberate and review together through the API. It is OpenAI-compatible, so an existing client only needs a different base URL, key, and model name.

Quick start

The last user message is the question; the messages before it are background material. A system message becomes judging criteria and constraints. Send your key in the Authorization header as a Bearer token.

https://hiai-meeting.com/v1

curl https://hiai-meeting.com/v1/chat/completions \
  -H "Authorization: Bearer $HIAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "model": "mpm-decide-standard",
  "messages": [
    { "role": "system", "content": "Decide within the confirmed budget." },
    { "role": "user", "content": "Which do we verify first: pricing, demand, or channel?" }
  ],
  "extra_body": {
    "mpm": {
      "answer_shape": "choice",
      "choices": ["Pricing", "Demand", "Channel"]
    }
  }
}'

import OpenAI from 'openai';

const client = new OpenAI({
  baseURL: 'https://hiai-meeting.com/v1',
  apiKey: process.env.HIAI_API_KEY,
});

const answer = await client.chat.completions.create({
  model: 'mpm-decide-standard',
  messages: [
    { role: 'system', content: 'Decide within the confirmed budget.' },
    { role: 'user', content: 'Which do we verify first: pricing, demand, or channel?' },
  ],
  extra_body: { mpm: { answer_shape: 'choice', choices: ['Pricing', 'Demand', 'Channel'] } },
});

console.log(answer.choices[0].message.content);

Issue a key

Issue a personal key from the API keys section in settings. The raw key appears once in the issue response, and only the mpm_xxxxxxxx prefix is kept in the list.

• Issue and manage API keys in settings

Billing

Every API and MCP call is charged in credits (Cr), regardless of purpose. There is no free trial or free single use: charges draw subscription usage first, then purchased credits. Actual usage is the billing basis, and a shortfall is caught during setup.

• Credit top-up and subscription details

Set up your tool

Use a judgment meeting as an MCP tool inside your agent (Claude Code, opencode, Codex), or call it directly over the OpenAI-compatible API (opencode provider, Python SDK). Pick your tool, copy the config as-is, and replace the key placeholder with your personal key.

claude mcp add --transport http hiai https://hiai-meeting.com/v1/mcp \
  --header "Authorization: Bearer <HIAI_API_KEY>"

Note: this is a judgment model — one request runs one meeting and takes minutes. To drive a coding agent, choose an agent-type preset (e.g. mpm-agent-standard); for judgments, connect it as the MCP deliberate tool or a subagent model.

Tool calls during a meeting (gathering material)

If the meeting needs more evidence it asks for tool calls, and it resumes toward a verdict once your client runs them and sends the results back. The server never runs the tools. What you send becomes evidence for every participant, so assume your material reaches external providers.

Judgment skill package

A skill that helps agents call the multi-AI judgment tool at the right time with a good question. It covers when to open a meeting and when not to, how to write the question, background and decision criteria, how to pick a preset, mode and cost cap, and the async polling and result delivery. Download a bundle below and install it for your tool.

• Claude Code plugin (zip)
• Codex skill folder (zip)
• opencode commands and agent (zip)
• Claude.ai skill (zip)

The MCP connection uses the environment variables HIAI_MCP_URL (endpoint) and HIAI_API_KEY (personal key). Presets and modes are not baked into the skill; it reads them from the server at run time. The public repository is set at release time.

extra_body.mpm options

Values specific to MPM go inside extra_body.mpm rather than in OpenAI fields. All of them are optional, and omitting one applies the server default.

answer_shape

Answer shape — choice, score, yes_no, or free. Defaults to free.

choices

The options to weigh. They are the subject when answer_shape is choice.

cost_cap_usd

Cost ceiling for this call in USD. The run stops before exceeding it. Omit to use the server default.

visibility

private (default) or public. Non-public is the default.

round_cap

Upper bound on discussion rounds. Omit to use the model preset value.

max_utterance_length

Character limit for a single utterance. Omit to use the model preset value.

purpose

discuss, decide, inspect, explore, commentary, or explain. When set, it changes only the preset's purpose and keeps the depth from model. Omit it to use the purpose of model; with no model either, the server derives it from answer_shape.

web_search

Web-search mode. true turns it on. Turning it on makes the meeting a coin meeting that spends credits. Omit to use the preset value.

reasoning

Reasoning mode. true turns it on. Turning it on makes the meeting a coin meeting that spends credits. Omit to use the preset value.

evidence

Evidence mode. It is enabled only when your plan includes it; otherwise the run proceeds with it off and records the swap in mpm.preset. Omit to use the preset value.

presets

Pick the whole, model, persona, and progress presets, the same ones the web new-meeting dialog uses. Picking only one leaves the rest at the purpose's Standard.

Details

Preset selection — the whole, model, persona, and progress slots. Pick the whole/model/persona/progress presets, the same ones the web new-meeting dialog uses (there is no per-seat model or persona override). whole is a whole-preset id (mpm-…), model is a model placement, persona is a persona placement, and progress is a progress preset (fast, balanced, deep, independent). Each slot also accepts a preset id you saved in the web app (my-<scope>-…@v<n>, listed by GET /v1/models); another person's preset is rejected. Picking only one leaves the rest at the purpose's Standard, and changing any value marks the result as custom.

idempotency_key

A key that keeps a repeated request from creating a second meeting. The same key returns the same job.

Model list

GET /v1/models returns recommended presets by purpose and depth. They match the Light, Standard, and Deep presets in the web new-meeting dialog, and the name is the model id, so pass it to model as is. The preset sets the participant tier, headcount, persona placement, rounds, and deliverable; omit model to use the standard preset for the purpose.

Judgment

• mpm-decide-fast
• mpm-decide-standard
• mpm-decide-deep
• mpm-inspect-fast
• mpm-inspect-standard
• mpm-inspect-deep
• mpm-discuss-fast
• mpm-discuss-standard
• mpm-discuss-deep
• mpm-explore-fast
• mpm-explore-standard
• mpm-explore-deep
• mpm-commentary-fast
• mpm-commentary-standard
• mpm-commentary-deep
• mpm-explain-fast
• mpm-explain-standard
• mpm-explain-deep

Judgment form: one request = one meeting, takes minutes (up to 10), billed per meeting. Use it as a judgment tool (MCP deliberate) or a subagent model, not as a coding agent's main model.

Agent

• mpm-agent-fast
• mpm-agent-standard
• mpm-agent-deep

Agent form: a coding agent's main model — seconds per step, a meeting only when needed. Each step bills one moderator call; a step that convenes a meeting adds one meeting.

• mpm-agent-fast — summon rule: at most 1 meeting(s) per conversation · always: planning · cap $0.10 per meeting
• mpm-agent-standard — summon rule: at most 2 meeting(s) per conversation · always: planning · final review · cap $0.30 per meeting
• mpm-agent-deep — summon rule: at most 4 meeting(s) per conversation · always: planning · final review · cap $0.60 per meeting

Status

Read progress from mpm.status in the response, and keep watching the meeting in the same web room. Streaming (stream: true) is not token-by-token output: it keeps the connection alive with a heartbeat and streams round progress. For long meetings, prefer stream: true.

created

The job was accepted.

running

The deliberation is running.

awaiting_tools

The meeting is waiting for client tool results — run the tool_calls and return them as tool messages with the same tool_call_id.

completed

It reached a conclusion.

partial

It ended partially without a conclusion.

blocked

A safety rule stopped it — the response carries finish_reason content_filter.

failed

The run failed.

Limits and safety

Meetings are non-public by default. Debt brokering and gambling requests are refused; legal, investment, and medical topics proceed only after you confirm, with output limited to general information. Inputs with personal data, trade secrets, or credentials are warned about or blocked.

Policies

• Terms of service
• Privacy policy
• Refund policy