Replit Agent / Code Execution API

license

Proprietary

source

closed

service

proprietary(Replit managed Autoscale Deployments, no self-host)

open-part

third-party isolation primitive omegajail(github.com/omegaup/omegajail) + an ARCHIVED thin client (replit-code-exec, ISC, read-only since 2025-02-11)

Closed managed service. The sandbox primitive (omegajail) is open source but it is a third-party dependency rather than Replit’s own open core, and the thin client is archived; there is no self-hostable open product, and an open dependency under a closed service is not open core. github.com/replit/replit-code-exec is archived and read-only under ISC, and its README describes the execution service as a Repl the user forks and deploys onto Replit Autoscale Deployments, sandboxed by the third-party omegaup/omegajail project - nothing Replit publishes is a self-hostable execution service. The two replit.com blog URLs cited beside it answer 403 to every fetch attempt and could not be read.

• https://replit.com/blog/ai-agents-code-execution recorded 2026-06-04

  Code Execution API uses omegajail (open source unprivileged container sandbox) on Replit's proprietary Autoscale Deployments; replit-code-exec PyPI client; Python-focused

• https://github.com/replit/replit-code-exec recorded 2026-08-13

  Replit code-exec client: stateless API for AI agents, runs untrusted Python in ephemeral unprivileged containers using omegajail (github.com/omegaup/omegajail); ISC license; repo ARCHIVED read-only 2025-02-11

• https://raw.githubusercontent.com/replit/replit-code-exec/main/README.md recorded 2026-08-13

  the client is an interface to a stateless Replit-hosted eval-python API; standing up a copy means forking a Repl and deploying it on Replit Autoscale Deployments, with omegaup/omegajail as the code sandbox, so no off-platform self-hostable service exists

• https://blog.replit.com/ai-agents-code-execution recorded 2026-06-04

  canonical Replit blog: Code Execution API uses omegajail on Autoscale Deployments, ~100ms, Python-focused (Feb 2024)

The Replit platform serves more than 50M users, and runs more than 1M live projects on Google Cloud Run. The code-execution API specifically discloses no standalone usage figures, so adoption is attributed to the Replit Agent and platform surface it powers rather than to the raw API, and placed at 4 on that platform-surface reach; the API alone would be lower. Neither Replit nor the code-execution API exposes a countable artifact, so the band rests on a vendor-quoted figure rather than on a recomputable count. replit.com/products/agent answers 403 to every fetch attempt.

• https://cloud.google.com/blog/products/serverless/whats-new-for-cloud-run-at-next26 recorded 2026-08-13

  Replit runs >1M live projects (on Cloud Run), primary-source scale signal for the Replit surface

• https://replit.com/products/agent recorded 2026-06-04

  Replit Agent product page (the agent surface the code-exec API backs)

Fast container-level execution tuned for agent code evaluation, but container rather than microVM isolation, a narrow language focus and no persistence put it mid-tier against microVM-backed peers. The public self-serve client (replit-code-exec) is archived and read-only, and the capability now lives inside the Replit Agent surface. That picture comes from two independent live channels - the archived README and the package’s own PyPI metadata - both describing execution inside an ephemeral unprivileged container created on the fly on Replit Deployments, with omegaup/omegajail as the code sandbox, a stateless API optimized for AI agents, Python, and the container image swappable via `evalctl`. Two claims that once travelled with this score are not carried, because nothing readable supports them: a ~100ms cold start, and a discontinued prior stateful-Repl experiment. Both came only from the Replit blog post, which answers 403 to every request under every header set tried and has no Wayback capture at all, so no surviving copy of the document states them. Losing the 100ms claim removes a point in the product’s favor, so if anything it argues against raising the band. The dead blog source is kept below so a later reader can see what was lost rather than rediscovering the 403. The 3 rests on the same three grounds it always did - container rather than microVM isolation, narrow language focus and no persistence.

• https://replit.com/blog/ai-agents-code-execution recorded 2026-06-04

  DEAD SOURCE, retained as a record. Formerly the sole basis for the ~100ms start figure and the discontinued-stateful-Repl note, neither of which the value asserts any more. It answers HTTP 403 to every fetch attempt under every header set tried, and the Wayback availability API reports no capture of the post at all.

• https://pypi.org/pypi/replit-code-exec/json recorded 2026-08-14

  the package's own metadata, an independent channel for the same description - execution "inside an ephemeral unprivileged container created on the fly running in Replit Deployments using omegajail as code sandbox"; summary "A library for interacting with Replit's code-exec API"; carries no latency figure.

• https://raw.githubusercontent.com/replit/replit-code-exec/main/README.md recorded 2026-08-14

  stateless API optimized for AI agents, executing untrusted Python inside an ephemeral unprivileged container created on the fly on Replit Deployments using omegaup/omegajail as the code sandbox; the container image is swappable via `evalctl image`; Autoscale Deployments only; no latency figure stated.