Open Interpreter is an Apache-2.0 terminal coding agent relaunched in July 2026 as a Rust fork of OpenAI’s Codex CLI, positioned as the agent layer for low-cost and open-weight models like Kimi K3 and GLM.
The 2023 code-execution chat agent is gone: today’s Open Interpreter bets that the harness, not the model, is the product, and emulates whichever harness each model was trained on.
What it is #
A terminal agent installed by a curl script and run as interpreter or i, built on the Codex CLI core and inheriting its interface, app-server protocol, and cross-platform sandboxing.
Its distinguishing feature is /harness, which switches the active emulation among native, claude-code, kimi-code, qwen-code, deepseek-tui, swe-agent, and minimal profiles.
The pitch is that Claude drives a Claude Code-like harness, GPT drives the Codex harness, and Kimi, Qwen, and DeepSeek drive emulations of their provider-recommended harnesses instead of one generic tool set.
The portability posture is explicit: repository AGENTS.md, shared .agents/skills directories, MCP, ACP for editors, and a one-line codexPathOverride for existing Codex SDK users.
A separate Open Interpreter Desktop app covers cross-app knowledge work (spreadsheets, PDFs, email, browsers) with profiles, voice mode, approvals, and subagents.
The fork is documented rather than hidden: FORK_BRANDING.md defines the distribution as variant: "open-interpreter" on Codex internals, with rust-v release tags and ~/.openinterpreter as the product home.
Status #
Active and fast-moving again after roughly nineteen quiet months: 68,451 stars, 5,881 forks, and a push on 2026-09-25 as of 2026-09-27, with only 4 open issues, which tells you issue intake is tightly curated. The latest release is rust-v0.0.45, published 2026-09-20, and the 0.0.43 notes describe a maintenance release on the “stable upstream Codex rust-v0.154.0 compatibility baseline” with presets for glm-5.1, glm-4.7-flash, Gemini, Claude, and OpenAI IDs. The legacy Python line is stranded: PyPI shows open-interpreter 0.4.3, the last of 130 releases, uploaded 2024-10-26. The history matters: a Show HN on 2023-08-30 (82 points), a 117-point thread in November 2023, a 72-point “natural language interface for computers” thread in November 2024, then silence until the relaunch post of 2026-07-13. The README now opens with “Today: Kimi K3 is here”, saying the project reimplemented the provider-recommended Kimi Code harness in Rust.
Strengths #
- Harness emulation is a genuinely different bet: instead of one generic agent loop, each model gets the tool behavior, system instructions, and compaction workflow its RL environment resembled.
- The portability list is checkable rather than aspirational: AGENTS.md,
.agents/skills, MCP, ACP, and the Codex exec protocol, with a script (scripts/test-codex-sdk-compat.sh) that verifies SDK compatibility without a provider. - BYOK breadth reaches the cheap end of the market, down to Z.AI’s lightweight
glm-4.7-flash, and local models work through any OpenAI-compatible Chat Completions endpoint. - A computer-use QA skill (driving web apps through agent-browser and native apps through trycua) ships in the box, a niche most terminal agents leave to plugins.
Cautions #
- The star count flatters the present: most of the 68k stars date to the 2023 hype cycle, and the current user base is whoever adopted a July 2026 fork, not the historical total.
- It is a downstream fork of Codex with a
rust-v0.154.0baseline, so OpenAI owns the core and Open Interpreter’s differentiation is a provider, harness, and branding layer on top. - The evidence for harness emulation is vendor-run: the relaunch post reports the project’s own Terminal-Bench 2.1 measurement of DeepSeek V4 Flash, unverified independently.
- The identity churn is three pivots deep (code interpreter, OS-mode natural language interface, Codex fork), so anything built on its platform story should expect another turn.
- An installer that pipes
curl | shfrom the project’s own domain, for a binary fork of a fast-moving upstream, is a supply-chain surface worth pinning.
Pricing #
Free and open source under Apache-2.0, and I found no Open Interpreter-operated hosted or metered tier anywhere I fetched: the sitemap scan of 2026-09-27 lists no pricing pages, and the desktop FAQ describes a pure BYOK posture where data goes only to the provider in your active profile. Usage therefore bills at whatever provider the profile points at (Z.AI, DeepSeek, OpenAI, Anthropic, Google, or a local Ollama or LM Studio endpoint), so no prices belong in this note and no price history table applies.
Compared to #
- OpenCode: the established open terminal agent with an original codebase; choose OpenCode for community governance and a longer track record, Open Interpreter if the open-model harness-emulation pitch fits your stack.
- Grok Build: the other fork-lineage terminal agent in this section; Grok Build is a vendor monorepo sync that takes no contributions, while Open Interpreter is an Apache-2.0 fork you can patch.
- Codex: the upstream itself; the shared exec protocol and SDK override mean you can run both, and Open Interpreter is best described as a tuned Codex distribution for cheap models.
Bottom line #
Recommended for engineers standardizing on open-weight or low-cost models who want a maintained terminal agent without building one, and who accept that the core is Codex’s. Not for anyone needing a mature plugin ecosystem today (the old Python one is stranded), and not as evidence of community health via its star count. My disagreeable claim: harness emulation is a configuration layer rather than new research, and its long-term value depends on whether upstream Codex eventually makes per-model harnesses easy for everyone.
Changes #
- 2026-09-27 - Created when the owner’s GitHub-stars scan surfaced it.
See also #
- OpenCode - the category’s reference open terminal agent, and the comparison point for original codebase versus fork
- Grok Build - the other fork-lineage harness, with the opposite contribution posture
- OpenRouter - the open-model access layer this agent’s BYOK profiles typically bill through
- Model selection for coding tasks - the section-root guide to picking the models these harnesses drive
References #
https://api.github.com/repos/openinterpreter/openinterpreter - 68,451 stars, 5,881 forks, Apache-2.0, pushed 2026-09-25, 4 open issues, as of 2026-09-27
https://raw.githubusercontent.com/openinterpreter/openinterpreter/main/README.md - the Kimi K3 note, the
/harnesslist, ACP and Codex SDK compatibility, and the portability posturehttps://www.openinterpreter.com/blog/open-interpreter - the 2026-07-13 relaunch post: Apache-2.0, built on Codex CLI, and the harness-emulation rationale (the page is JS-rendered, so I extracted its text through a proxy; the direct fetch returned 200 with only the title)
interpreter/icommands, rust-v tag prefixhttps://pypi.org/pypi/open-interpreter/json - the legacy Python line: 0.4.3, 130 releases, last upload 2024-10-26
https://hn.algolia.com/api/v1/items/37315866 - the Show HN of 2023-08-30, 82 points, the original CodeLlama terminal CLI