Manifold orchestrates AI specialists, stores memory across time, and runs your agents on schedules — all through a single, composable engine.
Index — What Manifold Does
Register any AI model — cloud or local — as a named specialist. Compose them into Teams that collaborate on objectives. The orchestrator routes tasks automatically, delegates subtasks, and synthesises results across specialists.
Specialists exchange intent, negotiate capability, and hand off work with full context — a team of models behaving like one sharper mind.
Summarization compresses context on the fly and works across different providers. Transit shares state across agents in real time. Evolving Memory distils lessons across runs. MAGMA builds a multi-dimensional knowledge graph with semantic, temporal, causal, and entity edges — all composing before every LLM call.
Every interaction settles into layers — fast context on top, durable graph and vector recall beneath — so agents remember what mattered.
Build complex agent pipelines visually. Connect nodes for data retrieval, LLM enrichment, transformations, and outputs. Chains execute in dependency order — parallel where possible, sequential where required — with real-time flow visualisation.
Branches align, split, and merge as each dependency resolves — every run leaves a readable, inspectable execution path behind.
Pulse lets you schedule agents to run on intervals, daily times, or one-off timestamps. Agents wake up, execute their prompts, write results to Transit, and go quiet — no manual intervention required. Fully integrated with the room and project system.
Temporal triggers wake the right specialist at exactly the right moment, write their results to shared memory, and fade back into standby.
Iterate on prompts with live model output. Load datasets and run batch evaluations. Compare experiment runs side-by-side. Everything stays versioned and reproducible — turn any playground session into a reusable workflow in one click.
Prompt variants flow through controlled, side-by-side runs — turning rough ideas into measured, repeatable agent behavior you can ship.
Manifold implements the Model Context Protocol, letting any MCP-compatible tool or server plug directly into the agent engine. Expose file systems, APIs, databases, and custom tools — agents discover and invoke them automatically with zero boilerplate.
Connectors fan outward from the core, binding file systems, APIs, and databases into a single mesh that agents discover and operate on their own.
Manifold is open source and free. Bring your own models.
The whole engine lights up at once — every model, memory, and tool moving as one coordinated system. Open source, local-first, and yours to run.