- Installation
- hybrid — common denominator: asynchronous code tasks executed in the publisher's isolated cloud environments, producing diffs and pull requests to review. Differences: Devin adds a local CLI and Desktop plus Outposts on the customer's infrastructure; Codex cloud remains fully managed by OpenAI (the CLI and the IDE extension are its local sides); Jules is used via the web, its CLI and its API, but execution takes place only in Google's cloud.
- Prerequisites
- Common: an account with the publisher and connecting a Git repo. Devin: GitHub, GitLab, Bitbucket or Azure DevOps repos, environment prepared by blueprints (dependencies, variables, secrets). Codex cloud: a ChatGPT account whose plan includes Codex Cloud (Plus, Pro, Business, Enterprise, Healthcare, Education eligible; neither Free nor Go) and a published cloud environment. Jules: a Google account (18 and over), the Jules GitHub app authorized on the repos; paid plans via a Google AI subscription attached to a personal @gmail.com account.
- Hardware
- non_applicable — managed cloud execution at all three publishers; the only documented elements: the Jules VM has 20 GB of disk (official changelog), and Devin's Outposts run on the customer's machines, sized by them.
- Self-hosting
- partial — only Devin offers execution on the customer's infrastructure: Outposts (self-hosted workers, the inference loop staying in Cognition's cloud) and dedicated deployment in a VPC; Codex cloud and Jules are only hosted by their publisher.
- License
- the three cloud services are commercial products, but the exact license of each component (CLI, SDK, servers) was not verified component by component in the official sources consulted.
- API & integrations
- API: yes — Devin exposes a REST API (v3, sessions, organizations, Enterprise) and Jules a v1alpha REST API (sources, sessions, activities; API key, maximum 3 keys per account). For Codex, the documented surface is the SDK and the app-server for local agents, plus the Analytics and Compliance APIs; no documented cloud-task REST API. · SDK: yes — official Codex SDKs in TypeScript (@openai/codex-sdk) and Python (openai-codex) to control local Codex agents. Devin and Jules SDKs: non_documente. · CLI: yes — at all three: Devin CLI (with task transfer to the cloud), Codex CLI (which can also launch cloud tasks) and Jules Tools (@google/jules: remote new/list/pull, TUI dashboard). · Webhooks: yes — Devin Automations can start sessions from webhooks (also Slack, GitHub, Linear, schedules). Codex and Jules webhooks: non_documente. · MCP: yes — at all three, on the client side: Devin connects to MCP servers and exposes its own MCP server; Codex supports MCP servers; Jules only supports a restricted list of approved MCP servers (Linear, Stitch, Neon, Tinybird, Context7, Supabase at launch).
- Models
- Devin: multi-model composed system — recent models from Anthropic, OpenAI, Google and Cognition, plus open models (DeepSeek, Kimi, GLM) in the selector; third-party LLM keys not supported (official deployment FAQ). Codex: OpenAI models only, the default model depends on the plan and configuration. Jules: Google's Gemini models — Gemini 2.5 Pro on free, higher (Pro) or priority (Ultra) access to the latest model, starting from Gemini 3 Pro.
- Data
- Common: the repo is cloned into the publisher's isolated cloud environment, which installs the dependencies there, modifies the code and returns a diff or a PR. Codex: ChatGPT's data controls apply to the processed content; for Business, Enterprise and Edu, OpenAI states it doesn't use inputs and outputs to improve its models by default. Jules: no training on private repo content (official FAQ). Devin: training and retention policy non_documente in the official pages consulted (Trust Center page inaccessible during the research).
- Documented limits
- Jules publishes precise quotas: 15 tasks/day and 3 concurrent (free), 100/day and 15 concurrent (Pro), 300/day and 60 concurrent (Ultra), per rolling 24-hour window. Codex: no separately published cloud-task limit — usage counts against the ChatGPT plan's Codex allowance and credits, with no distinct VM charge at launch. Devin: consumption measured in plan quota or ACUs (Enterprise), number of concurrent sessions depending on the plan, sleep after 30 minutes of inactivity (5 minutes for Lite sessions and those created via API).
- Operations
- Devin: SSO, SCIM, RBAC and custom roles, ACU limits per organization and per user, Session Insights and audit API in Enterprise; dedicated deployment in a single-tenant VPC possible. Codex: workspace RBAC, compliance API (usage logs, local and cloud clients), Agent Security policies (sandbox, approvals, network access by domain) and shareable cloud environments. Jules: plan approval before execution, activity feed and notifications; no documented SSO or SCIM (individual Google accounts).