Scenario
A YAML manifest plus persona files: places, population, time model, budget, questions. Versionable, diffable, generatable.
The Platform
Autonomous organisations as a runtime
A single agent scales to the edge of its context window. An organisation scales through roles: own context per agent, communication over events, knowledge in shared memory. HiveWeaver is the substrate for that.
The three-layer model
Three layers, clear contracts: you describe a world, plugins provide its behaviour, the engine carries both — and never changes per scenario.
Plugin
A .NET library with places, tools, decision strategies and metrics. Domain knowledge lives here — persona- and role-neutral.
Engine
An Orleans cluster of virtual actors: agents, places, simulations. Memory, audit, approvals, budget and replay are engine features.
Anatomy of a decision
Every agent decision is a thought — an auditable tree instead of a black box:
Trigger
An event in the place, a heartbeat, a directed message — or nothing, in which case the agent keeps sleeping.
Perception & recall
The relevant place state plus recall from episodic and semantic memory — bounded and logged.
Decision
Deterministic strategy, LLM choice or hybrid — the model provides content, the control flow stays yours.
Action & audit
A tool call with a declared side effect: mocked or real, approval-gated where needed — always in the log.
Where this leads
The human is the apex
The chain runs autonomously upwards until it hits something it cannot decide itself — then it escalates with a decision-ready summary. Approvals, budgets and audit make that dependable.
Guardians with real independence
The strongest reviewers aren't LLMs at all: test runs, builds, deterministic fact checks. Where LLMs review, fresh context and adversarial framing keep the gate from being theatre.
One substrate, many organisations
Dev department, radio station, research team, underwriting desk — the same primitives: roles, places, events, memory, governance. The domain is a scenario, not a fork.
Where the project stands
Dependable today
- Simulation, live and hybrid mode with approvals, a hard budget cap and audit
- Verdict with mandatory evidence, replay, thought log, agent interviews
- Scenario scaffolder: documents in, runnable scenario out
- Studio cockpit, a 91-command CLI, REST/SignalR, MCP and A2A
- Tenant isolation, encrypted secrets, GDPR erasure with key destruction
- Two public proofs: hiveweaver.fm and the dev-department runs
In the works
- Public release: source code, Docker images and CLI packages
- Plugin scaffolder: generating integrations from API specifications
- Webhook delivery end-to-end plus IMAP, queue and file triggers
- Container isolation for third-party plugins and curated marketplace operations
- Python and TypeScript plugin SDKs on the language-neutral contract
- A public demo access to the Studio
Your first scenario in ten minutes
Install the template
One command installs the project templates for scenario and plugin.
Scaffold a scenario
The template generates a manifest, a persona and a mini plugin with a first tool.
Start the platform
One command brings up the whole environment — database, messaging, engine, API, dashboard.
Validate & start
The validator checks the manifest against the loaded plugins; then the simulation starts.
Watch it live
Follow in the terminal or in the Studio: cockpit, pulsing world graph, timeline, social feed.
Read the result
Verdict as Markdown or JSON, inspect the thought log — and chat with the report to dig deeper.
The second path is even shorter: upload documents, state your question — the scaffolder builds the scenario.
Four ways in
CLI
91 commands, each with JSON output for CI and scripts — from starting a sim to searching thoughts.
Studio
Live cockpit with sim clock and cost per hour, world graph, prediction report, agent interview — declaration-driven, no per-domain UI code.
REST & SignalR
API-first: the API is the only way into the engine; CLI and Studio are just clients. Live streams via SignalR and SSE.
MCP & A2A
Your AI assistant drives HiveWeaver via MCP; agents talk to external agents over A2A.
The Weaver family
HiveWeaver FM
The publicly audible showcase — an independent product that consumes HiveWeaver like a customer, through the API.
To the showcaseTreeWeaver
Deterministic, hierarchical retrieval over your document corpus — the agents' shared knowledge.
Stratara & LoomWeaver
The CQRS/identity framework beneath the API, and the frontend platform the Studio runs on.
Licence & availability
HiveWeaver is released as open source under the Apache License 2.0: free to use — commercially too — with an explicit patent grant. Self-hosting is an explicit part of the design — all the way to fully local operation without a cloud LLM.
.NET 10 · Orleans · Aspire · PostgreSQL + pgvector · Redis · RabbitMQ · OpenTelemetry — a production stack, not a research script. 4,000+ tests keep it honest.
Let's talk
HiveWeaver is in pre-launch: source code, Docker images and a CLI will be published — preparation is under way, and a SaaS offering is deliberately not planned. If you have a use case — simulation, automation or both — we'd love to talk about it: the contact form is the shortest route.