HiveGate — run your Agno agents as a production API

Everything is an HTTP call

Your app talks to agents, teams, approvals and tokens through one REST API, with plain JSON or streamed responses.

Each user gets their own session. HiveGate stores the history, so the next message picks up where this one left off.

Request

curl -X POST $HIVEGATE/v2/agents/customer-support/chat \
  -H "X-API-Key: agk_…" \
  -d '{
    "message": "I need help resetting my password",
    "user_id": "user-42",
    "session_id": "sess-abc123",
    "stream": false
  }'

Response 200

{
  "content": "I can help you reset your password…",
  "agent_id": "customer-support",
  "session_id": "sess-abc123",
  "model": "gemini-3-flash-preview",
  "status": "completed",
  "run_id": "run-f47ac10b"
}

Agents API reference

A team gets one request. The supervisor splits the work between its agents and returns one answer, with token usage for the whole run.

Request

curl -X POST $HIVEGATE/v2/teams/support-team/runs \
  -H "X-API-Key: agk_…" \
  -d '{
    "message": "A customer can't log in after changing their email",
    "user_id": "user-42",
    "session_id": "team-sess-001",
    "stream": false
  }'

Response 200

{
  "content": "The customer needs to verify their new email…",
  "team_id": "support-team",
  "status": "completed",
  "token_usage": {
    "input_tokens": 1250,
    "output_tokens": 340
  },
  "run_id": "run-team-abc"
}

Teams API reference

When an agent wants to do something that needs sign-off, like sending an email, the run pauses. It continues once a person approves.

The run pauses

{
  "status": "paused",
  "run_id": "run-abc",
  "tools": [{
    "tool_call_id": "tc-001",
    "tool_name": "send_email",
    "requires_confirmation": true,
    "tool_args": { "to": "jane@acme.com" }
  }]
}

A person approves it

curl -X POST $HIVEGATE/v2/approvals/run-abc:tc-001/approve \
  -H "X-API-Key: agk_…"

{
  "job_id": "run-abc:tc-001",
  "status": "approved",
  "message": "Tool call approved"
}

Approvals API reference

Store a user's Google or Microsoft token once. Their agents can then use Gmail, Calendar, Contacts and Drive, and HiveGate refreshes the token when it expires.

Request

curl -X POST $HIVEGATE/v2/users/user-42/tokens \
  -H "X-API-Key: agk_…" \
  -d '{
    "integration_key": "google",
    "provider": "google",
    "token_type": "oauth2",
    "token_data": {
      "access_token": "ya29…",
      "refresh_token": "1//0dx…"
    },
    "scopes": ["calendar", "email"]
  }'

Response 201

{
  "integration_key": "google",
  "provider": "google",
  "token_type": "oauth2",
  "message": "Token stored successfully for google"
}

Tokens API reference

What you'd otherwise build yourself

Running an agent in a notebook is easy. Serving it to many users means building all of this around it. HiveGate ships it.

Serving

Agents over HTTP: Create, version and chat with agents through a REST API, with JSON or streamed replies.
Per-user sessions: Conversation history is kept per user and per session, and separated by tenant.
Teams of agents: Group agents into a team that coordinates, or one where a supervisor hands out the work.

Identity and integrations

API keys: Per-client API keys, plus a separate admin secret for management endpoints.
OAuth token store: Each user's Google and Microsoft tokens are stored and refreshed for you.
Workspace toolkits: Ready-made Gmail, Calendar, Contacts and Drive tools for Google and Microsoft accounts.

Knowledge

Knowledge base: Per-tenant documents your agents can search, stored in Qdrant.
Versioned prompts: Edit an agent's prompt through the API without redeploying. Every change gets a new version number.
Skills: Reusable step-by-step instructions, reference material and scripts that agents can follow.

Operations

Approvals: Risky tool calls pause until a person approves or denies them.
Tracing and errors: OpenTelemetry, Sentry and Logtail, switched on with environment variables.
Deploy anywhere: A prebuilt image for Render, Railway, Koyeb, Kubernetes, AWS, Azure and Google Cloud.

How a team run works

The hive at the top of this page, step by step.

A request comes in

Your app sends a message to a team, along with who the user is and which tenant they belong to.
The supervisor splits the work

A supervisor agent decides which agents in the team should handle each part of the request.
Workers act for the user

Each agent calls its tools, such as Gmail, Calendar or the knowledge base, using that user's own tokens.
Risky actions wait for a person

Anything marked as needing approval pauses the run until someone approves or denies it.
One answer comes back

The team replies as JSON or a stream. Every step is traced, so you can see what happened afterwards.

Run it locally in five minutes

You need Docker, Python 3.11+ and an API key for a model provider. The demo agent uses Gemini.

Clone it and add your key
```
git clone https://github.com/hivegate-ai/hivegate
cd hivegate
cp .env.example .env
```
In .env, set GOOGLE_API_KEY and uncomment AUTH_DISABLED=true for local use. The server reads .env when it starts, so do this first.

Start the databases and the API

docker compose up -d
./scripts/dev_setup.sh
source .venv/bin/activate
./scripts/start_server.sh

This starts Postgres and Qdrant and loads some demo agents.

Chat with the demo agent

curl -X POST \
  http://localhost:8000/v2/agents/demo-assistant/chat \
  -H 'Content-Type: application/json' \
  -d '{"message": "Hi", "user_id": "u1",
       "session_id": "s1", "stream": false}'

Interactive API docs are at http://localhost:8000/docs.

Where it runs

Every option uses the same image, ghcr.io/hivegate-ai/hivegate, built for both x86 and ARM.

Platform	Good for	Config
Render	The easiest first deploy	`render.yaml`
Railway	The smoothest Postgres setup	`railway.toml`
Koyeb	Trying it on a free tier	`koyeb.yaml`
Kubernetes	Self-hosting and scaling	Kustomize manifests
AWS, Azure, Google Cloud	Your own cloud account	ECS, Container Apps, Cloud Run

Questions

Is it production-ready?

Yes, with caveats. The API, supervisor, job queue and toolkits are covered by an integration test suite that runs against real Postgres and Qdrant on every change. Choose your observability backend and review authentication before exposing it to end users.

How does it relate to Agno?

HiveGate runs on top of Agno. Agno provides the agent itself: model calls, tools and memory. HiveGate adds the HTTP API, multi-tenancy, storage, supervisor teams, OAuth tokens and the operational side.

Is it an API gateway like Kong?

No. Despite the name, HiveGate doesn't route traffic between services. It runs AI agents and exposes them over a REST API.

Which models can I use?

Any model Agno supports, including OpenAI, Anthropic (Claude) and Gemini. Set the model per agent, or per request with the model field.

Do I need a vector database?

Only for the knowledge base. Qdrant comes with the Docker Compose setup, but chat works without it.

Can I run it without Docker?

Yes. You need Postgres 14 or later, and Qdrant if you use the knowledge base, reachable over the network. See the deploy guide.

Run your Agno agents as a production API.

Everything is an HTTP call

What you'd otherwise build yourself

Serving

Identity and integrations

Knowledge

Operations

How a team run works

A request comes in

The supervisor splits the work

Workers act for the user

Risky actions wait for a person

One answer comes back

Run it locally in five minutes

Clone it and add your key

Start the databases and the API

Chat with the demo agent

Where it runs

Questions