ColeMurray/background-agents

An open-source background agents coding system

Open‑Inspect (Background Agents)

What it is – An open‑source platform that lets AI coding agents run in the background on real development environments (Node, Python, git, browsers, VS Code). Users start a session via a web UI, Slack, GitHub, Linear or a webhook, and the agent works on the task while they do other things.

Core ideas

  • Single‑tenant, trusted‑team deployment – all users belong to the same organization; no tenant isolation.
  • Full dev sandbox – each session gets a fresh sandbox with Node 24, Python 3.12, git, browser automation, etc.
  • Multi‑model support – Anthropic Claude, OpenAI Codex, xAI Grok, OpenCode Zen, etc., selectable per session.
  • Real‑time collaboration – multiple people can view and prompt the same session; commits are attributed to the prompting user.
  • Automation – cron‑style schedules, Sentry alerts, GitHub workflow triggers, or custom webhooks can spawn sessions without human input.

Key features

  • Fast startup – snapshots and pre‑built images let a session resume instantly.
  • Multi‑repo sessions – up to 10 repos can be cloned side‑by‑side; environments can be saved with their own secrets and images.
  • Managed skills – reusable instruction bundles (skills) that are automatically injected when a session starts.
  • Child sessions – an agent can spawn parallel sub‑sessions that run in separate sandboxes.
  • Client integrations – Web UI (Next.js), Slack bot, GitHub bot (auto‑review, @mentions), Linear bot, generic webhooks.
  • Commit attribution – git commits record the name/email of the user who issued the prompt.
  • Secure sandbox – AES‑256‑GCM encrypted secrets, per‑repo or per‑environment scope, isolated runtime.

Architecture snapshot

Clients (Web, Slack, GitHub, Linear, Webhooks)
   │
   ▼
Control Plane (Cloudflare Workers + Durable Objects)
   │   – stores session metadata, websockets, event streams, GitHub integration
   │   – D1 DB holds repo‑scoped secrets
   ▼
Data Plane (Sandbox Backend)
   │   – Supervisor → Harness (Claude / OpenCode) → Bridge
   │   – Full dev environment (Node, Python, git, browser automation)
   ▼
External services (GitHub, Slack, Linear, Sentry, etc.)

Getting started

  1. Follow docs/SETUP_GUIDE.md for local development and contribution.
  2. Deploy using docs/GETTING_STARTED.md (Cloudflare Workers + chosen sandbox infra – Modal, Daytona, E2B, OpenComputer, etc.).
  3. Create the first team (step 10 in the guide) and test a session.
  4. Configure AI model credentials (Claude Agent, OpenAI OAuth, Grok subscription) as described in the model docs.

Typical use cases

  • Automated PR generation for routine code changes (dependency upgrades, lint fixes).
  • Continuous triage of Sentry alerts or GitHub Action failures.
  • Scheduled maintenance tasks (e.g., nightly linting, code‑base health checks).
  • Collaborative pair‑programming where a human guides an AI assistant.
  • Rapid prototyping of multi‑repo refactors without leaving the UI.

License – MIT.

Credits – Inspired by Ramp’s Inspect; built with Modal, Daytona, Vercel Sandbox, OpenComputer, E2B, Cloudflare Workers, OpenCode, Claude Agent SDK, and Next.js.

Related

  • Project
    NVIDIA/OpenShell

    OpenShell is an open‑source, policy‑driven sandbox runtime that lets autonomous AI agents read files, install packages, call APIs, and use credentials safely. It enforces restrictions at the kernel level, formally verifies policy changes, and provides SDKs for Python, TypeScript, Go, and Rust.

  • Project
    open-thoughts/OpenThoughts-Agent

    OpenThoughts‑Agent (OT‑Agent) is a research‑focused toolkit for generating training data, fine‑tuning, and evaluating small agentic language models. It bundles a datagen pipeline (Ray + vLLM), SFT support via LLaMA‑Factory, and a Harbor‑based evaluation framework, with launchers that work on SLURM clusters, cloud VMs (SkyPilot), or locally. The codebase is aimed at researchers building and benchmarking agentic LLMs.

  • Project
    hhyqhh/inno-agent

    Inno Agent is an open‑source personal learning AI built on the Pi coding‑agent SDK. It adds three memory layers (learner profile, wiki, session logs), a natural‑language scheduler, IM channel support (Feishu/WeChat), and an in‑browser practice lab. Available as an Electron desktop app, a React web UI, or a terminal CLI, it works with any OpenAI‑ or Anthropic‑compatible model and can be extended via a skill hub. The project is MIT‑licensed and includes Docker, sandboxing, and a session‑replay showcase.

  • Project
    The-PR-Agent/pr-agent

    PR‑Agent is an open‑source AI bot that automatically reviews pull‑requests (or issues) using LLMs. It can summarize changes, give a full code‑review, suggest improvements, or answer free‑form questions. The tool works with GitHub, GitLab, Bitbucket, Azure DevOps, and Gitea, and can be run as a GitHub Action, Docker container, or local CLI. Prompts are fully configurable, and it supports many LLM providers via LiteLLM. Self‑hosting keeps data private, and the project is community‑maintained with sponsorship from Qodo.

  • Project
    deepagent-ltd/deepagent-code

    DeepAgent Code (DAC) is a desktop‑plus‑terminal workspace that lets you use an LLM as a coding assistant for multi‑step tasks. It offers persistent sessions, inspectable project memory, sub‑agents for independent jobs, a managed browser, and optional instant‑messaging integration, all configurable with any compatible LLM provider. Licensed under AGPL‑3.0.