Roder
Overview

Roder docs

Roder is reusable Rust infrastructure for coding agents. Use it when you need a runtime that can run tools, enforce policy, preserve state, emit replayable events, and support multiple product or research surfaces.

It is aimed at teams building agent products, AI labs running evaluations or RL infrastructure, and platform engineers who want to embed a coding-agent runtime behind internal tools, desktop clients, CI jobs, or hosted sandboxes.

Alpha contract. Roder is actively moving. The docs describe the current Rust workspace, binary, protocol, and extension boundaries; stable API guarantees will come later.

Start by goal

  • Install & run if you want to build the workspace and start the TUI or app server.
  • Use cases if you are deciding whether Roder fits a product, eval, CI, or domain-agent workflow.
  • CLI changelog if you want release-by-release notes for the roder command.
  • Runtime architecture if you need to understand how a turn moves through the Rust core.
  • Extension model if you want to replace providers, storage, memory, policy, tools, or UI surfaces.
  • Blaxel sandbox runner if you want coding tools to run in a persistent remote sandbox that can pause, detach, and rejoin.
  • Eval scoreboard if you want current development trajectory results and run artifacts.

Repository map

The Rust implementation is organized around the roder binary androder-* crates:

  • crates/roder-api — extension traits, canonical events, tools, inference, sessions, policy, and shared types.
  • crates/roder-core — runtime loop, event bus, policy gate, tool execution, cancellation, media artifacts, forks, knowledge recall, compaction boundaries, lifecycle recovery, agent-control lifecycles, and provider orchestration.
  • crates/roder-extension-host — default distribution wiring for providers, web search, subagents, sessions, tools, media, knowledge, process extensions, and TUI services.
  • crates/roder-app-server and crates/roder-protocol — local, remote-node, and hosted JSON-RPC control planes plus typed protocol shapes.
  • crates/roder-dynamic-workflows and crates/roder-roadmap — approval-gated workflow scripts, child-agent orchestration, roadmap boards, loop orchestration, checkpoints, and reports.
  • crates/roder-cli and crates/roder-tui — command entrypoint and reference terminal client.
  • sdk/typescript — JavaScript and TypeScript host APIs for local or remote app-server control, streaming, display parts, and external tools.
  • crates/roder-ext-* — provider, storage, search, remote-runner, subagent, plan-mode, memory, notification, Chrome, and Zerolang extensions.

Read next

  • Eval harness — deadlines, task ledgers, plan-first reruns, and run artifacts.
  • Mission — why Roder exists and what it is trying to become.
  • Contact the team — RL tooling, embedded harnesses, product work, and OSS sponsorship.
  • Extension inventory — every implemented first-party extension and when it appears.
  • CLI changelog — user-facing notes for the roder binary, sourced from GitHub releases.
  • TypeScript SDK — local/remote agent hosts, abortable streams, display parts, and runner read roots.
  • Memory & embeddings — SQLite memory plus OpenAI, Google, and ZeroEntropy embeddings.
  • Project knowledge — markdown requirements, decisions, runbooks, links, revisions, and recall.
  • Agent swarm — bounded, rate-limit-aware fanout with live progress and ordered results.
  • Agent control — long-lived teammates with canonical paths, follow-up turns, waiting, and non-destructive interruption.
  • Lifecycle recovery — interrupted-turn records, cleanup proof, runtime drain, and restart diagnostics.
  • Blaxel sandbox runner — isolated remote coding workspaces with pause, resume, detach, and rejoin.
  • Workspace forks — branch a thread or task into isolated writable workspaces.
  • Media generation — OpenAI and Google image providers with artifact persistence.
  • Usage analytics — local-only token, tool, latency, and session statistics.
  • Remote agent nodes — control a full remote Roder runtime over mTLS.
  • Hosted service — multi-tenant Roder gateway with isolated tenant runtimes.
  • Process extensions — non-Rust services through stdio JSON-RPC.
  • Roadmap orchestration — inbox triage, worker dispatch, and evidence-gated completion.
  • OpenRouter provider — Grok Build setup, model discovery, and API-key configuration.
  • roder.cloud provider — hosted inference with roder_ key exchange and Responses-subset turns.
  • Fireworks provider — first-party API-key provider with account-scoped model ids and Responses transport.
  • Inference providers — default provider set, including Codex GPT-5.6 Sol/Terra/Luna, Grok 4.5 on xAI/SuperGrok, Cursor AgentService, Kimi Code auth, Claude Fable 5, and provider compaction boundaries.
  • Claude Code provider — local CLI sessions with optional subscription-backed Claude-in-Chrome browser tools.
  • Google Vertex AI — Gemini-on-Vertex via service-account OAuth and regional endpoints.
  • Provider tool search — provider-native discovery for large tool catalogs.
  • Codex-parity tools — view_image, unified_exec, freeform patches, and eval-loop persistence.
  • Chrome browser control — pair Roder with a logged-in Chrome session.
  • Dynamic workflows — approval-gated scripts for larger multi-agent jobs.
  • Zerolang graph edits — checked ProgramGraph editing tools for Zero tasks.