Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

Agent Orchestration SDK

Orchvia: local engine with TS and Python SDKs that runs Claude Code and Codex agents as a team, routing tasks by code and typed decisions.

LibrarybetaMIT
View on GitHub ↗

Overview

Orchvia runs Claude Code and Codex agents as a team from your own application, using one local engine with TypeScript and Python SDKs. Agents are long-lived warm sessions that keep history and can be forked. Tasks, messages, approvals and token usage are stored in SQLite before anything runs, and after a crash work with an unknown outcome waits for you rather than being retried. A result counts as done only when a person or a registered check accepts it.

The list summary describes it as routing each task with a typed decision instead of a manager LLM. The README excerpt stresses that your code and configured limits decide what runs next, not a manager model, but does not describe a Jev call directly, so the role of a decision model here is unclear.

What the decision model decides

TypeQuestionAnswers
ChoiceWhich agent or session should take this task?(inferred)—

How it works

  1. Application submits a task through the SDK
  2. Engine persists it in SQLite
  3. Routing picks an agent session (code rules / typed decision)
  4. Agent runs; result waits for acceptance by a person or check

Strengths

  • Durable state survives crashes without silent retries
  • Warm sessions reuse context
  • Quickstart runs with a fake runtime and no account

Caveats

  • Jev usage not described in the README excerpt
  • Uses Node's experimental built-in SQLite

Getting started

git clone https://github.com/masonlee39/orchvia.git && cd orchvia && node examples/typescript/quickstart.ts

At a glance

Type
Library
Status
beta
Language
TypeScript
Runtime
Node engine with TS and Python SDKs
Has code
yes
License
MIT

Use cases and integrations

How curated lists describe it

  • Orchestration engine for multi-agent work with durable mailboxes and warm sessions, routing each task with a typed decision instead of a manager LLM.

    MrJev/awesome-jev

Frequently asked questions

What does it orchestrate?

Teams of Claude Code and Codex agent sessions run from your own application.

Does a model decide what runs next?

The README says your code and limits decide, not a manager model.

Can I try it without an account?

Yes, the quickstart uses a fake runtime and calls no model.

Related projects

  • agent-stack

    Glue that turns one Linux machine into a team of coding agents, with Jev making typed calls like who builds a task and whether it may merge.

    ProjectbetaTypeScript
  • decision-gate

    npm library that rate-limits, spend-caps, caches and redacts every Jev request made in a loop.

    LibrarybetaJavaScript
  • jevable

    Bash-native CLI that filters any stream by meaning with Jev, like grep for semantics, so an agent wakes only for lines that matter.

    ProjectbetaTypeScript
  • AutoGPT

    AutoGPT agent platform, listed as adding Jev decision blocks for typed routing, filtering, scoring and confidence-gated next actions.

    ProjectproductionPython
  • jeff

    Go CLI for Jev that returns numbers to shell scripts, with threshold exit codes and a weighted multi-dimension rank command.

    CLIbetaGo
  • jeq

    Pipe-friendly CLI that lets shell scripts and agents ask Jev typed questions and gate on the answers with exit codes.

    ProjectbetaGo