Which tiers support tool calling?

Bee Comb, Buzz, and Hive serve structured tool calling, along with JSON mode and vision/video input, over the OpenAI-compatible endpoint.

Does JSON mode guarantee valid output?

Output is produced under schema-guided decoding, so it conforms to your schema at generation time rather than being validated after the fact.

Can I use existing agent frameworks?

Yes — anything that targets the OpenAI Chat Completions protocol (including LangChain-style stacks and custom loops) works by switching the base URL and key.

How should agents handle irreversible actions?

Put a human-review gate before them. Bee's Acceptable Use Policy requires at least one out-of-band human checkpoint before an agent takes physical-world actions or makes irreversible legal or financial commitments.

What does an agent workload cost?

Workspace plans pool tokens (Buzz: 25M/month across 6 seats), and the API meters per token at published rates — e.g. $0.50/1M input and $1.50/1M output on the Build plan's pay-as-you-go rates.

Agents need structure. Bee returns it.

Structured tool calls, schema-valid JSON, and streaming on an OpenAI-compatible endpoint — the primitives agent frameworks actually depend on, served on Comb, Buzz, and Hive.

The primitives, done properly

→Tool / function calling: the model returns structured tool_calls with arguments your dispatcher executes — the same shape OpenAI-compatible frameworks already parse.
→JSON mode: schema-guided decoding produces valid JSON at generation time, so agent state transitions never hinge on regex-repairing malformed output.
→Streaming: server-sent events for responsive multi-step loops.
→256K native context — and 1M long-context on Comb and Buzz — so long-running agent transcripts don't fall off a cliff.

Pick the tier by workload, not by guesswork

Bee Buzz is sized for agent workloads — multi-tool orchestration and code-writing pipelines — with 6 seats and 25M pooled tokens at $149/mo. Comb covers lighter automations; Hive adds the 99.5% uptime SLA and opt-in quantum reasoning for teams running agents in production.

All three serve tool calling, JSON mode, and vision/video input on the same wire format, so moving a workload between tiers is a model-alias change.

MCP as the integration bus

Bee also plays the tool side of the equation: the hosted MCP server exposes Bee's chat, code, security, research, document, and memory tools to any MCP client. Agents built in Claude Desktop, Claude Code, Cursor, or your own MCP runtime can call Bee as a capability rather than a raw endpoint.

The safety line we actually publish

Bee's Acceptable Use Policy requires autonomous agents to include at least one out-of-band human-review gate before physical-world actions or irreversible legal and financial commitments. That's not a limitation to hide — it's the design constraint that makes production agents auditable.

Model aliases resolve to dated immutable snapshots, so agent behaviour is attributable to a specific release when you debug a run.

Frequently asked questions

Which tiers support tool calling?: Bee Comb, Buzz, and Hive serve structured tool calling, along with JSON mode and vision/video input, over the OpenAI-compatible endpoint.
Does JSON mode guarantee valid output?: Output is produced under schema-guided decoding, so it conforms to your schema at generation time rather than being validated after the fact.
Can I use existing agent frameworks?: Yes — anything that targets the OpenAI Chat Completions protocol (including LangChain-style stacks and custom loops) works by switching the base URL and key.
How should agents handle irreversible actions?: Put a human-review gate before them. Bee's Acceptable Use Policy requires at least one out-of-band human checkpoint before an agent takes physical-world actions or makes irreversible legal or financial commitments.
What does an agent workload cost?: Workspace plans pool tokens (Buzz: 25M/month across 6 seats), and the API meters per token at published rates — e.g. $0.50/1M input and $1.50/1M output on the Build plan's pay-as-you-go rates.

A coding assistant that plugs into the editor you already use OpenAI-compatible, down to the wire format AI infrastructure engineers don't have to fight

Start on the free tier

Bee Cell is free — no card. Scale to paid tiers, the API, or sovereign deployment when you are ready.

Get started free See pricing