Key takeaways
- Open-source clone of Ramp's internal 'Inspect' background coding tool
- Uses Modal or Daytona for cloud sandboxes and OpenCode as the agent runtime, with Anthropic, OpenAI, and OpenCode Zen models
- Single-tenant only — designed for internal team use, not multi-tenant SaaS
- 1,931 GitHub stars and 293 forks as of June 2026, up from 537 in its first two weeks; actively maintained
FAQ
What is Background Agents?
An open-source system that runs AI coding agents in the background using Modal sandboxes and Cloudflare Workers.
Who created Background Agents?
Cole Murray, an AI/ML consultant and former Amazon Sr SDE, based in San Francisco.
How does Background Agents compare to Devin?
Background Agents is open-source and self-hosted; Devin is a commercial hosted service with more polish but less transparency.
What coding agent does Background Agents use?
OpenCode (opencode.ai) as the in-sandbox agent runtime, with a choice of models from Anthropic (Claude), OpenAI (GPT/Codex via ChatGPT subscription OAuth), and OpenCode Zen.
Executive Summary
Background Agents (also called "Open-Inspect") is an open-source implementation of a background coding agent system, directly inspired by Ramp's internal Inspect tool. The system enables AI agents to work on coding tasks asynchronously while developers focus on other work. Built with Modal (or Daytona) sandboxes and Cloudflare Workers, it's designed for internal team use with single-tenant deployment. As of June 2026 it supports models from Anthropic, OpenAI, and OpenCode Zen, plus Slack, GitHub, Linear, webhook, and cron-triggered sessions.[1]
| Attribute | Value |
|---|---|
| Company | Open source (Cole Murray) [2] |
| Founded | 2026 |
| Funding | None (open source) |
| Employees | Individual creator |
| Headquarters | San Francisco, CA [2] |
Product Overview
Background Agents launched in late January 2026 and gained 537 GitHub stars in its first two weeks; as of June 2026 the repo has 1,931 stars, 293 forks, and remains actively maintained (last commit June 10, 2026).[1] The project is directly inspired by Ramp's internal Inspect tool.[3] It now has an official website at backgroundagents.dev, which also offers managed deployments on request.[4]
The system enables AI agents to work on coding tasks asynchronously while developers focus on other work. Unlike real-time pair programming tools, background agents process tasks independently and deliver completed PRs.
Key Capabilities
| Capability | Description |
|---|---|
| Async Execution | Agents work independently while developers do other work |
| Fast Startup | Near-instant sessions using Modal snapshots |
| Multiplayer | Multiple users collaborate in same session |
| Commit Attribution | Every commit attributed to prompt author |
| Multi-Provider Models | Anthropic Claude, OpenAI GPT/Codex, OpenCode Zen with per-session reasoning controls |
| Automations | Cron schedules, Sentry alert auto-triage, webhook triggers |
| Sub-Task Spawning | Parallel sub-tasks in separate sandboxes |
| In-Browser VS Code | Optional code-server connected to session workspace |
| Repo Setup | Custom .openinspect/setup.sh for environment config |
Product Surfaces / Editions
| Surface | Description | Availability |
|---|---|---|
| Web UI | Browser-based interface with real-time streaming and terminal panel | GA |
| Slack Bot | @mention or DM to start sessions; threaded replies | GA |
| GitHub Bot | Auto-review on PR open or @mention in PR comments | GA |
| Linear Bot | Assign agent to an issue to start a coding session | GA |
| Webhooks | Trigger sessions from external systems via authenticated HTTP POST | GA |
Technical Architecture
Background Agents separates concerns across two planes:
Control Plane (Cloudflare):
- Workers handle API requests
- Durable Objects maintain per-session state
- D1 Database stores repo-scoped secrets
Data Plane (Modal or Daytona):
- Modal provides the primary cloud sandbox infrastructure[5]
- A Daytona sandbox backend was added as an alternative[1]
- Each session gets isolated container
- OpenCode runs as the coding agent runtime[6]
Key Technical Details
| Aspect | Detail |
|---|---|
| Deployment | Self-hosted (Cloudflare + Modal or Daytona) |
| Model(s) | Anthropic Claude (Haiku/Sonnet/Opus 4.x), OpenAI GPT 5.x and Codex, OpenCode Zen (Kimi K2.5, MiniMax M2.5, GLM 5) |
| Integrations | GitHub, Slack, Linear, Sentry, webhooks |
| Open Source | Yes (MIT license) |
Strengths
- Fully open-source — MIT licensed, inspect and modify everything
- Modern architecture — Cloudflare edge + Modal sandboxes is solid foundation
- Fast iteration — Modal snapshots enable near-instant session starts
- Multiplayer support — Collaboration features rarely seen in OSS agent tools
- Active development — Rapid commits, responsive maintainer
- Model flexibility — Anthropic, OpenAI (via ChatGPT subscription OAuth), and OpenCode Zen models with per-session reasoning controls
- Commit attribution — Proper authorship tracking via GitHub OAuth
Cautions
- Single-tenant only — Not suitable for multi-org or commercial deployment
- Infrastructure dependencies — Requires Modal or Daytona for compute plus Cloudflare
- OpenCode runtime coupling — Locked to OpenCode as the agent runtime, though models are now configurable across providers
- No enterprise features — Missing Jira, signed commits, and audit logs
- Early stage — Roughly four months old as of June 2026; expect rough edges and breaking changes
- Self-hosting complexity — Requires Modal/Daytona and Cloudflare account management
Pricing & Licensing
| Tier | Price | Includes |
|---|---|---|
| Open Source | Free | Full functionality (MIT license) |
| Managed Deployment | Contact for pricing | Custom deployment via backgroundagents.dev [4] |
Licensing model: Open source (MIT)[1]
Hidden costs: Modal or Daytona compute (estimated ~$0.50-2/hour as of February 2026), Cloudflare (free tier may suffice). Expect $50-200/month for an active team, plus model/API costs.
Competitive Positioning
Direct Competitors
| Competitor | Differentiation |
|---|---|
| Devin | Devin is commercial hosted; Background Agents is self-hosted OSS |
| Tembo | Tembo has multi-tenancy and enterprise features; Background Agents is single-tenant |
| Codex App | Codex is first-party OpenAI; Background Agents uses OpenCode (though it can run OpenAI Codex models via ChatGPT OAuth) |
| Emdash | Emdash is local-first; Background Agents is cloud sandboxes |
When to Choose Background Agents Over Alternatives
- Choose Background Agents when: You want open-source, self-hosted background agents for internal team
- Choose Devin when: You need commercial support and hosted service
- Choose Tembo when: You need multi-tenancy and enterprise features
- Choose Emdash when: You prefer local-first without cloud dependencies
Ideal Customer Profile
Best fit:
- Engineering teams wanting to self-host background coding agents
- Developers studying modern agent architecture patterns
- Startups willing to operate their own infrastructure
- Teams with DevOps capacity to manage Modal/Cloudflare
Poor fit:
- Multi-tenant SaaS requirements
- Enterprise customers needing compliance features
- Teams without DevOps capacity for infrastructure
- Organizations requiring signed commits or BYOK
Viability Assessment
| Factor | Assessment |
|---|---|
| Financial Health | N/A — individual open source project |
| Market Position | Niche — best OSS background agent implementation |
| Innovation Pace | Rapid — active development |
| Community/Ecosystem | Growing — 1,931 stars and 293 forks as of June 2026 [1] |
| Long-term Outlook | Uncertain — depends on maintainer commitment |
Cole Murray's track record (claude-code-otel, Amazon background) is solid, but single-maintainer OSS projects have inherent sustainability risks.
Bottom Line
Background Agents is a well-architected open-source implementation of Ramp's Inspect concept. The Cloudflare + Modal architecture is production-worthy, and the multiplayer features show thoughtful design.
However, the single-tenant security model limits applicability beyond internal team use, and the agent runtime remains coupled to OpenCode even though model choice now spans Anthropic, OpenAI, and OpenCode Zen.
Recommended for: Teams wanting to self-host background coding agents, developers studying agent architecture, and startups comfortable managing infrastructure.
Not recommended for: Multi-tenant deployments, enterprise compliance requirements, or teams without DevOps capacity.
Outlook: Valuable as reference implementation and for internal team use. Commercial/enterprise deployment requires additional work on multi-tenancy and security.
Research by Ry Walker Research • methodology