AI Developer ToolsIn Development

opentine

Git for Agent Runs · Framework + Agent Kernel + Three Visual Consoles

opentine is a complete ecosystem for version-controlling AI agent execution. The core framework turns every agent run into a content-addressed DAG of steps that can be paused, resumed, forked from any step, and diffed side-by-side — like git branches for code, but for agent reasoning paths. Around the core sits an event-sourced agent kernel, a deterministic replay engine, and three first-party visual consoles (terminal, native desktop, and browser) so teams can supervise, audit, and intervene in agent runs through whatever surface fits the workflow. Model-agnostic by design, with one-line swaps across Anthropic, OpenAI, Google, Ollama, and any OpenAI-compatible endpoint.

Visit opentine.com All Projects

At a Glance

opentine — core CLI (`tine run / show / fork / diff / replay / resume`), ~250 LOC of execution logic, <5 MB install

opentine-agent — minimal event-sourced agent kernel with Read / Write / Edit / Shell primitives and planner/executor split for approval gates

opentine-replay — deterministic replay engine: re-execute from any checkpoint, fork alternates, generate side-by-side diffs of diverged runs

opentine-tui — Textual terminal dashboard with three-pane layout (runs · step tree · details), live filtering, keyboard-driven fork / resume / diff

opentine-gui — native cross-platform desktop console (Dear PyGui) with interactive node-editor DAG, minimap, and pause/resume/fork from disk

opentine-web — Starlette + Mermaid.js browser dashboard with REST API (`/api/runs`, `/api/runs/{id}`, `/api/runs/{id}/mermaid`) for embedding in internal tooling

Content-addressed steps — every step is hashed from its inputs, so identical inputs always produce the same ID and replays are bit-exact

Pause / Resume across machines — runs serialize to portable `.tine` files; pick up an in-flight agent on a different host without state loss

Universal provider layer — swap Claude, GPT, Gemini, Llama, or any OpenAI-compatible endpoint without touching agent code

Built-in tool library — web fetch, search, filesystem, shell, Python REPL — and per-step cost / cache tracking to eliminate token waste

The opentine Stack — One Framework, Six Surfaces

opentine ships as a small core framework wrapped by a family of first-party tools. Pick the surface that fits the workflow — script it from CI, watch it from a terminal, drive it from a desktop GUI, or embed the browser dashboard in internal tooling. Every surface speaks the same `.tine` file format, so a run paused in one console resumes in another.

CORE FRAMEWORKAlpha

opentine

tine — the run engine

Python 3.11+ · ~250 LOC core · <5 MB install

CLI verbs: tine run · show · fork · diff · replay · resume

Content-addressed steps — identical inputs always hash to the same ID

Pluggable model layer — Anthropic, OpenAI, Google, Ollama, any OpenAI-compatible endpoint

Built-in tools: web fetch, search, filesystem, shell, Python REPL

AGENT KERNELEarly Release

opentine-agent

Event-sourced minimal kernel

Python · SQLite event store · adapter pattern

Four primitives — Read, Write, Edit, Shell

Every operation emitted to the event store — no hidden state

Planner / executor split enables approval gates and dry-runs

Canonical message + tool + response types translated per provider

REPLAY ENGINEAlpha

opentine-replay

Deterministic time-travel

Python · msgspec · blake3

Re-execute any run from any checkpoint, bit-exact

Fork branches for alternative reasoning paths

Side-by-side structural diff of diverged runs

Underpins the fork / replay / diff verbs in every UI surface

TERMINAL CONSOLEProduction

opentine-tui

Textual three-pane dashboard

Python · Textual 3.0+ · runs anywhere SSH does

Three panes: runs list · step tree · per-step detail

Live preview — auto-refresh while agents are mid-flight

Keyboard-driven: filter (/), fork (f), resume (R), mark + diff (m / D)

Status colours, cost, and duration shown inline per run

DESKTOP CONSOLEBeta

opentine-gui

Native node-editor GUI

Python · Dear PyGui 2.0+ · cross-platform

Interactive DAG node editor with parent–child links and minimap

Three-pane layout: runs · run/step detail · DAG canvas

Run actions write back to disk — pause, resume, fork-from-step

Auto-discovers new .tine files as agents emit them

BROWSER CONSOLEProduction

opentine-web

Embeddable HTTP dashboard

Python · Starlette · uvicorn · Mermaid.js

Localhost dashboard with status, cost, and model indicators

Mermaid.js renders forks as branches in the run graph

REST API: /api/runs · /api/runs/{id} · /api/runs/{id}/mermaid

Polls every 5s — drop into existing internal tooling via iframe or API

Provider-Agnostic by Design

One-line model swap · zero code changes

Claude (Anthropic)GPT (OpenAI)Gemini (Google)Llama / Ollama (Local)Any OpenAI-compatible endpoint

First-Party Surfaces

CLI · TUI · GUI · Web · Agent · Replay

~250

LOC in Core

Lines of execution logic in tine

<5 MB

Install Footprint

Core CLI, full Python install

15+

Provider Endpoints

Out-of-box, plus any OpenAI-compatible

Let's Talk

Interested in opentine?

Whether you're an investor, a prospective partner, or an engineer who wants to use or integrate opentine, reach out — we'll get back to you within a business day.

Contact Visit opentine.com

JaCode

AI coding assistant (in development)

Logos

Fine-tuning as a service