A vendor-neutral control plane for AI agents.
Every agent runs a different stack. Corellia governs them as one fleet.
A vendor-neutral control plane for AI agents — spawn, deploy, govern, and manage them across any model, any provider, any framework, from one place. Deployment is a commodity; the unsolved problem is governing a fleet of heterogeneous agents and controlling what each one can reach. Corellia is the single layer that holds all of them to the same guardrails.
Pick a harness, name the agent, choose any model or provider, and deploy. Agents are vehicles you select and drive — each spun from a pinned, bit-identical template.
IAM for a fleet of agents. Define each agent's capabilities and tool permissions once; it then operates autonomously inside the guardrails you set — never beyond them.
One gallery for the whole fleet: live status, start, stop, and decommission, bulk actions with dry-run preview, and an inspector to edit any agent's config in place.
One contract for any framework. IAM for the whole fleet. Every agent spun from a pinned template — and held to the guardrails you set.
Spawning an agent is RPG character creation: choose a harness from a curated library of pre-integrated runtimes, give it a name, pick any provider and model, set its deployment config, and deploy. The harness is the vehicle; the model is the engine you drop into it. Every instance is digest-pinned — bit-identical to its template, with no mutable tags — so an upgrade is always an explicit, audited action, never a silent drift.
Corellia talks to every agent framework through a single Harness Interface Contract — "LSP, but for agents." LangGraph, CrewAI, the Hermes harness, or something you wrote yourself all present the same surface to the control plane. That is what vendor-neutral means in practice: heterogeneous agents, built on whatever stack their authors preferred, governed and operated as one fleet rather than five disconnected point tools.
Governance here is not user access control — it is capability and tool-permission boundaries for the agents themselves. A policy-setter defines, once, what each agent is allowed to reach: which skills it can load, which tools it can call, what context and memory it carries. The agent then runs autonomously inside that boundary. Deployment is a commodity; governing what a fleet of agents can actually do is the unsolved problem Corellia exists to solve.
A typed contract end to end: a Go backend over Connect-go RPCs and Protobuf, a Next.js console, Supabase Postgres with offline-validated JWTs. The control plane is the product — the layer that unifies orchestration, deployment, tool permissions, provider routing, and observability that teams otherwise stitch together by hand. Spawn anything; govern everything; from one place, across any model, any provider, any framework.
Corellia is a vendor-neutral control plane for AI agents — a single layer to spawn, deploy, govern, and manage agents across any model, any provider, and any framework, from one place. It unifies orchestration, deployment, tool permissions, provider routing, and observability that teams would otherwise stitch together by hand.
Corellia solves the problem of governing a fleet of heterogeneous agents. Its premise is that deployment has become a commodity while the unsolved problem is controlling what each agent can actually reach, so Corellia holds every agent to the same guardrails — defining once which skills, tools, context, and memory each one is allowed, then letting it run autonomously inside that boundary.
Corellia is open source, released under the Apache 2.0 license, with its source public at github.com/kaminocorp/corellia.
Corellia is in development and not yet shippable to run as a product — its page carries a Coming Soon call to action and lists status as In Development — though its source is already public under Apache 2.0.
Corellia differs from agent frameworks such as LangGraph or CrewAI in that it is not a runtime for building a single agent; it is the control plane above them. It talks to every framework through a single Harness Interface Contract — 'LSP, but for agents' — so LangGraph, CrewAI, the Hermes harness, or a custom framework all present the same surface and are governed as one fleet.
Corellia is built on a Go 1.26 backend (using Chi) exposing Connect-go RPCs over Protobuf, a Next.js 15 and React 19 console, and Supabase Postgres with sqlc and offline-validated ES256 JWTs. It spans TypeScript, Go, and Python, and integrates the Hermes harness with a Corellia sidecar.