Automation
Workflow tools model processes as diagrams: nodes, edges, special cases. HiveWeaver models the organisation that does the work — and the flow emerges from its decisions.
An always-on scenario never terminates: it is woken from the outside and fully asleep between events.
Cron triggers with timezone and DST awareness: the morning briefing at 6:30, the weekly review on Friday. Ticks missed during a restart fire once, in a controlled way.
One run can trigger the next — cascades over place events. Webhook, IMAP and queue triggers exist as source types and are being wired end-to-end right now.
A waiting agent costs nothing — even if the answer takes days. Human-in-the-loop steps across arbitrarily long chains are first-class.
Every tool declares what it does to the world — read-only, reversible, irreversible, spends money, handles sensitive data. From that single declaration the engine derives: mock or execute, approval requirement, budget booking, retention period.
approvals:
- operation: "SendInvoice"
policy:
- condition: "amount < 500"
decision: auto-approve
- condition: "amount > 25_000"
decision: { route-to: cfo }
fallback: { route-to: finance-team }
approvers:
finance-team:
provider: slack
timeout: PT4H
onTimeout: escalate
escalateTo: cfoPolicies decide first, humans decide the rest: small items auto-approve, big ones route to a senior or the CFO — with deadlines, timeout behaviour and audit.
Trigger → simulation → thought → action → approval → outcome: every real action sits append-only in the audit log, with cost in dollars, approval reference and a trace id into telemetry. Six months later you can still prove why a reply went out.
The per-run cap is enforced before every model call. When it is reached, the run ends with a classified outcome — never silently.
Hybrid means: play the same task through N times safely, pick the best path — and execute for real only after sign-off.
An incoming request starts the sim phase: five response strategies run in parallel, side effects are mocked.
A criterion picks the winner — highest confidence, lowest cost, lowest risk. Below the bar: nothing executes.
A human sees the proposal and its evidence, signs off — and exactly one variant runs for real, with audit and idempotency.
Sim cost and real spend are reported separately; promoting simulation to live is one manifest line.
Binding an HTTP API means a few lines in the manifest. Credentials come from the encrypted vault; rate limiting, idempotency and audit come for free.
HiveWeaver is an MCP server — your AI assistant starts and inspects simulations — and an MCP client: agents consume third-party MCP servers as tools.
A signable agent card per tenant; internal agents delegate to specialised external agents as if they were ordinary tools.
The same stack for individuals and enterprises — self-hosted, fully offline with a local model if you want.
| Small cloud server (~€4/month) + local Llama | ≈ €0.13/day — LLM: €0 |
|---|---|
| Raspberry Pi 5 at home + local Llama | €0/day running cost |
| Small cloud server + your own cloud LLM key | ≈ €0.13 + €0.05–0.30/day |
Structurally safer than consumer agents: a tool that reads sensitive data is declared as such — and only runs with authorisation.
We won't match integration breadth by hand: from an API specification, HiveWeaver is meant to generate the integration itself — every operation risk-classified, every approval human. Designed and decided — implementation is the next step.
Cron, manual triggers and run cascades are production; webhook delivery plus IMAP, queue and file sources exist as source types and are being wired end-to-end.