OpenAI fully open-sources Codex Harness under Apache-2.0 with three integration tiers

On August 19 OpenAI published a blog post titled "Codex as a platform" announcing the full open-source release of the Codex agent harness — the runtime framework behind Codex App, the Codex CLI, and the IDE extensions — under the Apache-2.0 license. The repository is at github.com/openai/codex. OpenAI president Greg Brockman emphasized on X that Codex can drive far more than coding tools; it can be embedded in different business applications as an underlying agent execution framework. As of August 21 the repository had reached 107,443 stars and 16,354 forks, with the latest stable release v0.149.0 shipped on August 20.

The open-source release centers on three integration tiers: the first is codex exec, a lightweight non-interactive CLI suitable for CI pipelines, batch scripts, or one-shot background jobs, installable via a single curl invocation; the second is the Codex SDK, with TypeScript and Python interfaces to start, resume, and stream Codex tasks programmatically and switch models to any OpenAI-compatible endpoint with auto, manual, or suggest approval policies; the third is codex app-server, which talks to the local Codex process over JSON-RPC, supporting persistent conversation state, real-time event streams, mid-run task interruption, custom tool calls, and human approval handling.

OpenAI's data shows the harness design is as critical as the model: in ARC-AGI-3 testing, only two harness adjustments — reasoning preservation and context compression — lifted GPT-5.6 Sol from 13.3 to 38.3 percent while reducing output token volume by roughly 6x. The order-of-magnitude improvement on the same model validates the emerging agent-engineering consensus that execution framework, not raw model, determines ceiling. Among case studies, Thrive Holdings embedded the harness in a professional tax-preparation system, processing 7,000 filings while reducing preparation time by a third; Cisco integrated the Codex SDK into its cloud control plane so users can build custom apps from natural language; and OpenAI demoed Relay, a virtual logistics dashboard where users select delayed shipments on a business surface while an agent assembles data, calls tools, drafts recovery plans, and awaits human approval at consequential steps. The release coincided that week with the rc.6 through rc.8 series of updates to DeepSeek Harness, opening a parallel open-agent-foundation race.