--- title: "Cline's Desktop App Turns the Coding Agent Into a Service" slug: "cline-desktop-app-open-source-coding-agents-beyond-ide" author: "Blanca Amigot" date: "2026-09-15 10:00:00" category: "Premium" topics: "AI coding agents, Cline, open-weight models, agent governance, non-human identity, scheduled agents, enterprise AI security" summary: "Cline announced its desktop app on September 14, 2026, but signed macOS builds have been on GitHub since July 22 and it is still at v0.0.28. The real change is that a coding agent now has its own cron table." banner: "" thumbnail: "" linkedin: | Cline announced a desktop coding agent on September 14, 2026. The framing going around is "the IDE era is over." That is the least interesting thing about it. A few corrections first, because the version circulating is loose. It did not appear this week. Signed, notarized macOS builds have been on GitHub since desktop-v0.0.2 on July 22, 2026, under the app's earlier name, Cline Code. It is at v0.0.28 and Cline itself calls it "an early version." Windows is beta, and there is no Linux build. It is Apache 2.0 — that part checks out, copyright Cline Bot Inc. "No subscription" does not. Cline sells ClinePass at $9.99/month for 13 open-weight models, plus a provider tier and a custom-priced enterprise tier with SSO and RBAC. Bring-your-own-key works. A subscription is sold; it just is not required. And of the three models named in the post going around, two are in the ClinePass lineup. Poolside's Laguna S 2.1 is real — 118B total, 8B active, released July 21, 2026 — but it is not one of them. Here is what actually changes. Inside an editor, a coding agent is a child process of a human session: it starts when someone opens the window, it borrows that person's credentials, and it stops when they close the laptop. Give it parallel sessions and cron scheduling and it becomes a long-lived service. → It runs at 03:00 with nobody watching, so "the developer reviewed it" stops being a control → It holds its own API keys, so the audit log needs to name the agent, not the employee → It reaches the network on its own schedule, so egress becomes an allowlist question → It runs several sessions at once, so blast radius is per-session, not per-person That is not a developer-tools story. It is an operations object — the same class of thing as a build runner or a scheduled job — and it needs identity, allowlisted egress, approval gates and an audit trail before it gets a schedule. With ibl.ai you own all the code and the data — self-hosted inside your own perimeter, model-agnostic across any LLM, usage-based with no per-seat pricing, deployable anywhere from your own cloud to a fully air-gapped network. #iblai #AgenticAI #EnterpriseAI #DevTools #OpenWeights #AgentGovernance --- ## The Short Answer **Cline announced its desktop app on September 14, 2026, though signed macOS builds have shipped on GitHub since July 22 and it is still at version 0.0.28. It is Apache 2.0, with Windows in beta and no Linux build, parallel sessions and cron scheduling. That turns a coding agent into a long-lived service with its own credentials, egress and schedule. With ibl.ai you own all the code and the data, so that service runs inside your perimeter.** The headline everyone is repeating is that the IDE era is over. The part worth an engineering leader's attention is what the agent becomes once it leaves. ## What exactly did Cline ship, and when did the desktop app actually appear? Cline published [Cline Desktop: An open-source app for open-weight models](https://cline.bot/blog/cline-desktop-an-open-source-app-for-open-weight-models) on September 14, 2026, and describes it in that post as "an early version." The app did not appear that day. The tag [desktop-v0.0.2](https://github.com/cline/cline/releases/tag/desktop-v0.0.2) was published on July 22, 2026, shipping macOS `.dmg` builds "signed and notarized for Apple Silicon and Intel" under the app's earlier name, Cline Code. Version [desktop-v0.0.28](https://github.com/cline/cline/releases/tag/desktop-v0.0.28) landed on September 15, 2026. Eight weeks of public releases and still a 0.0.x version number is a more accurate picture than "just shipped." The licence claim checks out. The repository's [LICENSE file](https://github.com/cline/cline/blob/main/LICENSE) is the Apache License 2.0, copyright Cline Bot Inc., and the homepage advertises Apache 2.0 with 250+ contributors and 11M+ installs. Coverage is narrower than the announcement implies. [Cline's documentation](https://docs.cline.bot/usage/cline-desktop) marks Windows as beta and offers no Linux build. ## Which models does Cline Desktop run, and is Laguna S 2.1 one of them? The launch post names DeepSeek V4 Flash, DeepSeek V4 Pro, GLM 5.3 Flash and Kimi K3. Two of the three models the brief listed are therefore real and supported. The third needs correcting. Poolside's Laguna S 2.1 is real, and [released July 21, 2026](https://poolside.ai/blog/introducing-laguna-s-2-1). It is 118B total parameters with 8B activated per token, a 1M-token context window, weights on Hugging Face under the permissive OpenMDW-1.1 licence, and 70.2% on Terminal-Bench 2.1. [VentureBeat reported](https://venturebeat.com/infrastructure/poolside-drops-laguna-s-2-1-an-open-weight-coding-model-that-beats-rivals-10x-its-size) that Poolside went from the start of pre-training on May 22 to launch in under nine weeks on 4,096 NVIDIA H200 GPUs. But it is not part of Cline's hosted lineup. [ClinePass](https://docs.cline.bot/getting-started/clinepass) covers 13 models from Z.ai, Moonshot AI, DeepSeek, MiniMax, MiMo and Qwen. Poolside is not among them. Laguna would reach Cline the same way most of its catalog does — through your own key against a provider that serves it. The desktop release notes put that catalog at 6,079 models across 33 updated provider defaults, which is a statement about breadth, not about what is included. ## Is Cline Desktop really subscription-free, and what does Cline sell? No, and the distinction matters for anyone budgeting this. The app is free and bring-your-own-key. Cline also sells [ClinePass at $9.99/month](https://cline.bot/cline-pass), described as 2–5x the usage of standard API rate limits on open-weight models, plus a Cline provider tier billing inference at cost. [Cline's pricing page](https://cline.bot/pricing) adds a custom-priced enterprise tier with centralized billing, a team management dashboard, role-based access control and SSO. So "no subscription" is wrong as stated, and "no subscription required" is right. That is a real property and worth having: your keys, your provider, your weights if you self-host them. It is also where seat-shaped economics can re-enter. A tier negotiated against team size scales with headcount rather than with work performed, which is the wrong shape at scale no matter how permissive the licence on the client is. ## What changes when a coding agent runs as its own process instead of inside an editor? Inside an editor, a coding agent is a child process of a human session. It starts when someone opens the window, it borrows that person's credentials, it acts while they watch, and it stops when they close the laptop. Every control is implicitly a human control. The desktop app removes all four assumptions at once. Cline's documentation describes parallel sessions and scheduled tasks, and the launch post puts it plainly: "Schedule lets you create cron job automations in Cline Desktop." A nightly repository scan is a thing you can now configure. It is not a shipped feature with that name, and the difference matters when you are writing policy against it. Four properties change, and each one breaks a control an enterprise was relying on. **It runs unattended.** A scheduled run at 03:00 has no reviewer, so "a developer was watching" stops functioning as a safeguard. **It holds its own credentials.** Once the agent authenticates to providers and repositories on its own schedule, the audit trail has to name the agent, not the employee whose token it inherited. **It reaches the network on its own.** Web search is built in, the marketplace installs plugins and MCP servers, and egress is now a property of a background process rather than of a supervised session. **It runs several sessions at once.** Blast radius stops being one person's editor and becomes whatever the concurrent sessions can each touch. That is the same object an enterprise already governs elsewhere. A build runner, a scheduled ETL job and a service account all have identity, scoped permissions, allowlisted egress and an audit trail. A desktop coding agent is that class of thing, arriving through a developer download rather than through procurement. Which is precisely the discovery problem in [shadow agents](/blog/agent-governance-new-shadow-it-day-one-controls): nobody bought anything, so nothing appears in the expense report. ## Which controls does an enterprise need before a desktop coding agent gets a schedule? Four, and none of them is about the model. **A distinct identity per agent.** Credentials minted for the agent and scoped to what it needs, so revocation is one action and the log reads correctly. **Allowlisted egress.** An explicit list of destinations a scheduled run may reach, enforced at the network, not requested in a prompt. **Approval gates on consequential actions.** Read-only and propose-only by default; writing to a repository, opening a pull request or calling a paid API crosses a boundary that a human approves. **An audit trail that survives the run.** What ran, on whose authority, which model, which tools, which files — retained independently of the agent's own session state. The specification discipline in [spec-driven development](/blog/spec-driven-development-ai-coding-agents) is the other half of this. An unattended agent cannot stop and ask, so ambiguity resolved before the run is the only ambiguity resolved cheaply. ## How does ibl.ai govern agents that run as long-lived services? With ibl.ai you own all the code and the data. The platform runs on your own infrastructure with full source code access, is model-agnostic across any LLM so you can switch providers without rewriting anything, is usage-based with no per-seat pricing, and you can deploy anywhere — your own cloud, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity. The governance controls are platform properties rather than client settings. Our [September 4, 2026 platform update](/updates/platform-update-2026-09-04) shipped approval modes, per-agent workspaces and auto-minted platform keys for Code Mode, with agent guidance loaded from an `AGENTS.md` file the agent re-reads every turn. Approval modes are the gate. Per-agent workspaces are the blast-radius boundary. Auto-minted platform keys are what makes the agent's identity its own rather than a borrowed one. None of this is an integration with Cline, and the point is not to replace it. The point is that a coding agent with a cron table is an operations object, and an operations object needs controls that live below the tool a developer chose. 1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University. ibl.ai is family-owned and operated from New York, NY. *Related reading: [shadow IT stored data, shadow agents take actions](/blog/agent-governance-new-shadow-it-day-one-controls) — the day-one controls for agents that arrive without procurement; and [spec-driven development](/blog/spec-driven-development-ai-coding-agents) — why an unattended agent needs its ambiguity resolved before the run.* *Sources: the announcement, date and feature set from [Cline's blog post](https://cline.bot/blog/cline-desktop-an-open-source-app-for-open-weight-models) and [Cline Desktop documentation](https://docs.cline.bot/usage/cline-desktop); release dates and version numbers from [desktop-v0.0.2](https://github.com/cline/cline/releases/tag/desktop-v0.0.2) and [desktop-v0.0.28](https://github.com/cline/cline/releases/tag/desktop-v0.0.28); the licence from [the repository LICENSE file](https://github.com/cline/cline/blob/main/LICENSE); pricing and tiers from [ClinePass](https://cline.bot/cline-pass), [the ClinePass model list](https://docs.cline.bot/getting-started/clinepass) and [Cline's pricing page](https://cline.bot/pricing); Laguna S 2.1's specifications and licence from [Poolside](https://poolside.ai/blog/introducing-laguna-s-2-1) and [VentureBeat](https://venturebeat.com/infrastructure/poolside-drops-laguna-s-2-1-an-open-weight-coding-model-that-beats-rivals-10x-its-size).* ## Why does owning the AI stack matter? **ibl.ai is the agentic AI platform where you own all the code and the data. You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with no per-seat pricing — so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped.** - **You own all the code and the data.** Full source code under a perpetual license, running on your infrastructure. Not API access to someone else's platform — the stack itself is yours. - **Model-agnostic.** Run any LLM — Claude, GPT, Gemini, Llama, Command, or your own fine-tune — and switch providers without rewriting the platform. - **No per-seat pricing.** Usage-based billing against a budget cap you set. Cost tracks what your organization actually uses, not how many people you employ. - **Deploy anywhere.** Your cloud, your VPC, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity. 1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University. ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.