Turn engine
Own lifecycle, cancellation, context assembly, provider streaming, and end-of-turn semantics in one stable loop.
Harnessinfrastructurefor
Roder gives serious agent teams the machinery every production harness needs: real tools, policy, provider abstraction, durable state, replayable events, and one control plane for every client.
Open sourceProvider neutralBuilt for evals and products
Rusttyped runtime core
JSON-RPCshared app-server control plane
Replayevents, sessions, and checkpoints
Extensionsproviders, tools, memory, runners
Open foundation
The shared base for harness engineering.
Every serious agent project grows the same machinery: tools, policy, provider abstraction, durable state, replay, eval traces, and client control. Roder makes that base open, reusable, and compounding.
Linux for harness engineeringBuild once. Compound together.
Communityextensions, fixes, shared lessons
Companiesproduct velocity on a common base
Startupsmove faster without rebuilding the harness
Labsrepeatable evals and training loops
roder basetools · policy · providers · state · replay
Open interfacesShared extensionsReusable eval tracesBetter defaults
Runtime architecture
One core. Replaceable edges.
Tool policy
Route shell, filesystem, browser, image, edit, and domain tools through explicit specs, approvals, scopes, and audit events.
State and replay
Persist sessions, checkpoints, compaction boundaries, forks, knowledge, media artifacts, and traces so every surface reads the same record.
Extension kernel
Replace providers, memory, storage, policy, event sinks, runners, process-hosted services, and UI surfaces without taking over the runtime.
Who it is for
Built for teams whose agents have to survive contact with real workflows.
Roder is intentionally lower-level than a polished chat app and higher-level than a pile of crates. It is the runtime foundation you build on when the agent has to be inspectable, controllable, and repeatable.
Agent product teams
Ship a differentiated product surface on top of a runtime that already understands tools, state, permissions, and replay.
AI labs and eval teams
Run repeatable trajectories, compare providers, export structured traces, and plug the harness into RL infrastructure.
Platform engineers
Embed coding agents behind internal tools, SDK hosts, CI jobs, desktop clients, Chrome control, remote nodes, or hosted tenant gateways.
Alpha, useful now
Build on the harness instead of rebuilding it.
Start with the docs, inspect the eval trajectory, or talk to the Roder team about embedded harnesses, RL infrastructure, product integrations, and OSS sponsorship.