HiveGate 0.3.1 is the first release on PyPI: `uv tool install hivegate-ai`, then `hivegate init` and `hivegate serve`.

It also ships per-call cost metering, Claude prompt caching and approvable team runs.

https://github.com/hivegate-ai/hivegate/releases/tag/v0.3.1

If cache_write_tokens keeps pace with cache_read_tokens, your Claude prompt prefix changes per request. HiveGate records both for every call.

https://hivegate.dev/blog/claude-prompt-caching-agents/?utm_source=bluesky&utm_medium=social&utm_campaign=2026-W43

The log needs the run_id, the tool, the arguments as approved (after edits) and the person. HiveGate currently records the API key that approved, not the person, and a key shared by a team makes that log far less useful.

Worth specifying early: where paused state lives and what resume is keyed on. In HiveGate we keep paused runs in process memory today, so a restart or another replica answers the resume with 410 Gone. A durable run id plus an explicit expiry on the pause avoids that.

An approval gate should fail closed. HiveGate stops the run before a write tool executes. If the paused run is lost, commit returns 410 Gone and nothing is sent.

https://hivegate.dev/blog/human-approval-agent-tool-calls/?utm_source=bluesky&utm_medium=social&utm_campaign=2026-W43

Agreed. The gates that hold are the ones the model can't argue with: code outside the LLM checks the real tool name and arguments, and a write pauses until a person approves. A rule in the system prompt can be undone by text the agent reads. That's how approvals work in HiveGate.

Failed agent runs are still billed. HiveGate records each one with its real tokens and USD cost, so /usage/summary?group_by=status shows what failures cost you.

https://hivegate.dev/blog/llm-token-usage-cost-tracking/?utm_source=bluesky&utm_medium=social&utm_campaign=2026-W42

Before #110, 0% of input tokens on our deployment came from cache. Claude agents now cache the tools and system prompt, plus the conversation between round trips. Cache reads bill at a fraction of input.

https://github.com/hivegate-ai/hivegate/pull/110

If your Agno agents are defined in code and AgentOS covers what you need, use AgentOS. HiveGate is for agents stored as data and served to many users.

https://hivegate.dev/blog/deploy-agno-agents-production-rest-api/?utm_source=bluesky&utm_medium=social&utm_campaign=2026-W42

Kyle Atkinson explains what he saw when he videoed bees using #HiveGate. a way to protect #honeybees from #wasps. It was quite a revelation about #beehive insulation.
www.vita-europe.com/beehealth/bl...

There is some confusion at the beehives today. Someone changed the entrances into the hivegate system. They'll figure it out. #beekeeping

Wafflegate, Linkgate, Hivegate, and probably some other gates from last week+ that I forgot

The Vita Bee Health HiveGate against wasps was developed in NZ. www.vita-europe.com/beehealth/pr...

Are your #bees being troubled by #wasps? Look here at HiveGate: youtu.be/qkusguQ7k-0

A #honeybee swarm has evicted a #wasp in a stack of stored supers. Obviously a desres. Will the wasps try to wreak revenge at the end of the season? - see Vita's HiveGate protection: www.vita-europe.com/beehealth/pr...