rivet-dev

rivet-dev/agentos

Rust Apache-2.0 AI & ML

A faster, lighter, cheaper alternative to sandboxes. Run any coding agent inside an isolated Linux VM, with agent orchestration built in.

3.6k stars
172 forks
active
GitHub +88 / week

3.6k

Stars

172

Forks

49

Open issues

4

Contributors

v0.2.8-rc.1 08 Jul 2026

AI Analysis

agentOS is a lightweight, in-process runtime for AI agents that runs inside your application with near-zero cold starts (~6 ms) and significantly lower resource costs than traditional sandboxes. It's purpose-built for developers embedding coding agents (like Claude Code or Pi) directly into backend services, offering granular security controls and direct function bindings without network overhead. This is a specialized tool for AI application builders and backend engineers; it is not a genera...

AI & ML Runtime Discovery value: 7/10
Documentation 9/10
Activity 10/10
Community 8/10
Code quality 7/10

Inferred from signals mentioned in the README (tests, CI, type safety) — not a review of the actual code.

Overall score 8/10

AI's overall editorial judgment — not an average of the bars above, can weigh other factors too.

ai-agents lightweight-runtime v8-sandbox code-execution in-process-vm
Actively maintained Well documented Niche/specialized use case Apache-2.0 licensed Production ready
Deep Analysis · Based on README and public signals
2w ago

In-process VM for AI agents: 92x faster cold starts, 8x smaller than sandboxes

agentOS is a lightweight, in-process Linux VM runtime designed to execute coding agents (Pi, Claude Code, OpenCode) with millisecond cold starts and minimal memory overhead. Built by Rivet, it positions itself as complementary to traditional sandboxes rather than a replacement, integrating agents directly into backend services via function bindings. Primary users appear to be teams building AI agent infrastructure who prioritize latency and cost over full OS capabilities.

Origin

agentOS emerged from Rivet's broader agent orchestration platform (Feb 2024). The project represents a deliberate architectural choice: lightweight in-process execution for agents rather than full-VM isolation. It evolved as part of a suite including sandbox extensions and multiplayer agent coordination.

Growth

270 stars in the past 7 days (as of 2026-06-28) suggests recent visibility surge, likely tied to agent framework adoption. The project sits at 3,298 stars with 160 forks—modest but non-trivial. Growth appears correlated with AI agent infrastructure maturation in 2025–2026 rather than viral adoption; comparable repos (nono, sandbox-agent, NVIDIA OpenShell) in the 2.8k–7.3k range indicate this is an emerging category segment.

In production

Adoption not verified. README references 'Rivet Cloud' managed hosting and mentions deployment to 'Railway, Vercel, Kubernetes,' but provides no public customer names, case studies, or production usage metrics. The benchmarking data is author-published (Rivet), not third-party validated. No GitHub discussions, blog posts, or ecosystem integrations visible in metadata. SDK packages (@rivet-dev/agentos-core, @agentos-software/pi) exist on npm but package download stats not accessible from repository metadata.

Code analysis
Architecture

Based on README: likely built in Rust (container runtime efficiency) with a TypeScript/Node.js SDK surface. Appears to use capability-based security model (deny-by-default permissions) similar to browser sandboxing. Likely embeds a WASM or minimal Linux kernel to achieve sub-10ms boot. Can mount external storage (S3, Google Drive, host) as filesystems and exposes bindings for host functions as CLI commands inside the VM. Includes built-in orchestration primitives (sessions, workflows, agent-to-agent delegation, multiplayer sync). README does not document internal architecture details; inference from features and claimed performance.

Tests

Not documented in README. No benchmarking methodology details provided beyond hardware specs (Intel i7-12700KF for agentOS, E2B for sandbox baseline). Benchmark comparisons lack statistical rigor (CI, sample sizes for sandbox runs unclear).

Maintenance

Last push 2026-06-28 indicates active maintenance as of report date. Project is 2.3 years old (created Feb 2024). Weekly star gain (270/7 ≈ 39/day) suggests active feature development and community interest. No issues-closed or PR-merge rate visible from metadata, but README breadth (features, docs, quickstart) indicates ongoing investment. Relative to creation date, velocity appears healthy but not explosive.

Honest verdict

ADOPT IF: building agent infrastructure where latency (<10ms), cost efficiency, and tight backend integration are critical; you control deployment and accept the constraint that agents run inside your process (not arbitrary binaries). AVOID IF: agents need full OS capabilities (native compilation, browsers, system services); you require proven, third-party-validated production references; or you prioritize ecosystem maturity over performance. MONITOR IF: you're evaluating AI agent frameworks (2025–2026): agentOS addresses a real performance gap but lacks public production proof; ecosystem adoption is emerging, not yet widespread.

Independent dimensions

Mainstream potential

5/10

Technical importance

7/10

Adoption evidence

2/10

Risks
  • Adoption unverified: no public case studies or production user testimonials; success metrics unknown.
  • Narrow scope: optimized for agent execution, not general-purpose workloads; may become obsolete if agent runtimes converge to different architectures.
  • Benchmark credibility: performance claims are author-published; third-party validation absent. Cold-start benchmarks may not account for JIT warm-up or real-world workload patterns.
  • Dependency on parent org (Rivet): project tight coupling to Rivet Cloud ecosystem could limit independent adoption.
  • Early maturity: 2.3 years old, stars 3.3k (vs. 6k+ peers). Mainstream enterprise adoption unproven; may remain niche if sandbox providers optimize cold-start.
Prediction

agentOS likely remains a specialized tool for agent-infrastructure builders (startups, enterprise AI teams) through 2027. Growth depends on (a) mainstream adoption of multi-agent orchestration patterns, (b) cloud sandboxes failing to match its latency/cost, and (c) ecosystem adoption (bindings, extensions, MCP integration). If agent infrastructure commoditizes, in-process runtimes may fragment; if latency becomes table-stakes, agentOS gains mainstream traction.

0 found this helpful

Newsletter

Get analyses like this every Monday

Free weekly digest of the most interesting open-source discoveries.

Languages

Rust
50.9%
Vim Script
45.2%
C
1.7%
TypeScript
1.3%
Shell
0.4%
Makefile
0.2%
HTML
0.2%
Dockerfile
0%

Information

Language
Rust
License
Apache-2.0
Last updated
11h ago
Created
29mo ago
Analyzed with
anthropic/claude-haiku-4-5

Stars over time

Loading…

Contributors over time

Top 100 contributors only — repos with more will plateau at 100.

Loading…

Similar repos

rivet-dev

rivet-dev/sandbox-agent

Sandbox Agent is a universal HTTP API and TypeScript SDK for running and...

1.5k TypeScript AI & ML
TencentCloud

TencentCloud/CubeSandbox

CubeSandbox is a high-performance sandbox runtime built on RustVMM and KVM,...

9.4k Rust AI & ML
superradcompany

superradcompany/microsandbox

Microsandbox is a lightweight microVM runtime for executing untrusted workloads...

6.9k Rust DevOps
mattpocock

mattpocock/sandcastle

Sandcastle is a TypeScript library for orchestrating AI coding agents in...

6.7k TypeScript AI & ML
nolabs-ai

nolabs-ai/nono

Nono is a zero-latency, zero-setup sandbox for running AI agents securely...

3k Rust Security
vs. alternatives
E2B (not in similar repos, inferred from README)

Full Linux sandbox with native binary support. agentOS claims 92–516x faster cold starts and 8x lower memory. Trade-off: E2B supports arbitrary binaries and dev servers; agentOS optimizes for agent execution only.

CubeSandbox (TencentCloud, 6537 stars)

Broader sandbox ecosystem. agentOS is narrower, focused on agent orchestration with built-in bindings and workflows; CubeSandbox appears more general-purpose containerization.

OpenShell (NVIDIA, 7294 stars)

Shell/CLI environment. agentOS extends beyond shell to multi-agent coordination (multiplayer, agent-to-agent, workflows); OpenShell likely single-agent focused.

nono (nolabs-ai, 2841 stars)

Also Rust-based agent runtime. Insufficient details from similar-repos metadata. Both appear to target agent-centric workloads; differentiation unclear from available data.

sandbox-agent (rivet-dev, 1451 stars)

Sibling project from same org. TypeScript vs Rust. Likely sandbox-agent is higher-level orchestration; agentOS is the runtime layer.