A Atenea orchestration core
GitHub

Local orchestration for coding agents

Decide what should happen. Delegate who does it.

Atenea is the orchestration core that sits outside the CLIs it serves. It keeps capabilities stable while implementations, providers and health change underneath.

Get started → Read the architecture
contract 3.1.0 · Atenea 1.0.1 · local-first
goal→capability→implementation constraints · reach · health · cost

Atenea#

An orchestration core for agents and MCP tooling that lives outside the CLIs it serves. Every client — omp, Claude Code, Codex — connects to the same core. OpenCode is an opt-in model backend rather than a client adapter.

The one sentence that explains the rest: Atenea decides and delegates, it does not execute. The flow is always goal -> capability -> implementation.

Why it exists#

Tooling changes constantly. A workflow written against ripgrep breaks the day ripgrep is replaced. A workflow written against code.search does not: the capability is stable, and which tool answers it is a decision Atenea makes with what it knows about the repository in front of it.

The pieces#

PieceWhat it is
CapabilityA stable, tool-agnostic action. code.search, symbol.definition.
ImplementationA concrete provider of a capability, in four blocks: capability, constraints, cost, health.
RepositoryThe unit of work. Not the project — the repository.
SelectorThe funnel that picks an implementation: constraints, reach, health, then cost.
AdapterA dumb translator between the core and one far side — a CLI, or an MCP server. All the intelligence stays in the core.

Where the project stands#

These pages describe Atenea 1.1.0, speaking contract 3.6.0. The product is stable; Claude Code, Codex, OpenCode and the Kivgraph web viewer remain optional external provider surfaces. The changelog has what landed.

The core, the Capability Registry and the funnel run on all four stages. The shipped adapters include OMP, Claude Code and Codex, with graph providers available through MCP. It installs as a background service that keeps its own history in shape, proven now against a live language server on this repository: symbol.definition and symbol.references resolve across files, not only within the one a caller happened to be looking at, symbol.implementations answers for real too, not only the empty result an earlier bug let through, and symbol.overview lists what a file declares before anything in it is known by name — the question the other three all assume has already been answered.

Cost ranks the survivors rather than filtering them, and it says which figure it used. Until an implementation has been measured a couple of times, that figure is the estimate the catalog declared — the trace prints estimated so nobody reads a guess as an observation. This is the design’s break-in period: the numbers get better as real work runs through, and nothing has to be rewritten when they do.