What Is BYOK?
BYOK stands for "Bring Your Own Key." It means you supply your own model-provider API key — OpenAI, Anthropic, Google, OpenRouter, AWS Bedrock, or any OpenAI-compatible endpoint — and the provider bills you directly at list price. The AI coding tool itself sits in the middle as a client, not as a reseller, and adds no token markup. That definition is consistent across how vendors document it: JetBrains describes it the same way in its AI Assistant documentation, and Cursor explains the same relationship in its API-key help page.
BYOK is often confused with two adjacent ideas, so it is worth separating them precisely:
- BYOK (bring your own key): you authenticate to a provider you chose, and that provider meters and bills you. The tool passes your requests through.
- BYO-endpoint: you point the tool at a self-hosted inference endpoint — vLLM, Ollama, or any OpenAI-compatible gateway. This can be entirely local.
- BYO-weights: you run the model weights yourself and never call a hosted API at all.
The distinction matters because only the last two can reach zero marginal cost per request, while BYOK alone still means you pay a provider — just without a reseller's margin on top.
Why BYOK Matters (Cost, Choice, Privacy)
BYOK is not a niche preference. Four structural arguments explain why developers and teams seek it out.
1. Cost pass-through with no markup
With BYOK you pay the model provider's list price instead of a bundled subscription margin. Warp explicitly markets its BYOK option as "no markup," and other vendors have pointed out that some competitors charge a handling fee on top of provider usage where BYOK tools do not. Your inference spend becomes visible at the source rather than hidden inside a plan tier.
2. Provider and model choice
A BYOK tool is model-agnostic by construction. You can swap between model families — or between a frontier cloud model and a local one — without changing tools or negotiating a new plan. This is the flexibility that made VS Code's BYOK rollout notable: a broad ecosystem of providers through a single client.
3. Privacy and data control
Using your own key changes whose privacy policy governs your data. Cursor's documentation states that its zero-data-retention policy does not apply when you use your own key — your provider's policy takes over. For teams who must reason about where prompt data lands, BYOK keeps that decision in your hands. For the deeper version of this argument, see local AI coding and code privacy.
4. No lock-in, portable spend
BYOK avoids paying a subscription markup for model access and keeps your spend portable: the keys, the billing relationship, and the model choice belong to you, not to the tool vendor. If you switch tools, your provider relationship does not reset. That portability matters when a tool changes direction — a theme we return to in perpetual license vs subscription.
BYOK Support by Tool (2026)
The honest answer to "which AI coding tools support BYOK?" is: fewer than the marketing implies, and often only partially. Here is the support matrix, classified as Yes / Partial / No. Classifications and limits below mirror the source research exactly.
| Tool | BYOK | Notes / Limit |
|---|---|---|
| Cline | Yes | "Bring your own key, your own endpoint, or your own weights." Apache-2.0, free to install. |
| Continue.dev | Yes | Open-source (Apache 2.0); primary targets are local Ollama / LM Studio; cloud tiers optional. Note: acquired by Cursor in 2026 and the repo is now read-only. |
| Aider | Yes | Open-source; works with Ollama / local OpenAI-compatible APIs or any cloud key. |
| Cursor | Partial | Add OpenAI / Anthropic / Google / Azure / Bedrock keys, but chat models only — Tab completion still uses Cursor's models, and some agent models "cannot be billed to an API key." |
| GitHub Copilot | Partial | BYOK since VS Code v1.99 (Mar 2025); 2026 expansions add Anthropic, Microsoft Foundry, OpenAI, xAI, AWS Bedrock, Google AI Studio and any OpenAI-compatible provider. Some setups are enterprise-gated. |
| Claude Code | Partial | Not a classic BYOK tool (it is Anthropic's client), but supports third-party / OpenAI-compatible providers and API-key billing distinct from Pro/Max. |
| Zed | Partial | Pro / Business plans use Zed-hosted models (billed at API list price +10%); direct provider keys and Ollama / LM Studio are configurable. |
| Windsurf / Devin Desktop | Partial (unverified specifics) | A BYOK feature existed during the Windsurf era; exact 2026 status after the Cognition rebrand (Jun 2, 2026) is not confirmed in the sources reviewed. |
| Tabnine | No | Enterprise-only model; no self-serve BYOK. A 5% handling fee has been noted in comparisons versus BYOK tools. |
| Replit | No (unverified) | Bundled credit model; no documented BYOK. |
| Lovable | No (unverified) | Credit-based, bundled. |
| Bolt.new | No (unverified) | Credit / token-based, bundled. |
Two clarifications keep this table honest. First, "Partial" is not a small gap: for Cursor, BYOK covers chat but not the tab-completion experience that many users value most. Second, "No" is not always a criticism — credit-based tools like Replit, Lovable and Bolt.new deliberately bundle inference so that users never manage keys at all. That is a real trade-off, not a flaw. It simply is not BYOK.
BYOK + Local Models: The Full-Control Setup
BYOK is powerful on its own, but it becomes a different category of tool when combined with local models — in that setup the "provider" is your own machine.
The pattern is well established. Cline, Continue.dev and Aider all support pointing at a local runtime, typically Ollama or LM Studio, which expose an OpenAI-compatible API. Aider with Ollama, for example, is widely described as costing nothing per request after the one-time model download. Zed likewise accepts local models via configuration.
The trade-off is hardware. The rough rule of thumb is about 1 GB of VRAM per 1 billion parameters on a quantized model: 7B needs roughly 4–5 GB, 14B around 8–10 GB, 32B around 20–24 GB, and 70B-class models 48 GB or more. For Apple Silicon, 16 GB of unified memory is the practical minimum; 32 GB is where local work becomes genuinely useful.
So the honest framing is this: BYOK gives you cost transparency and provider choice; adding local models eliminates marginal cost and gives you full data control. The two are complementary, not competing.
How Doforu Does BYOK
Doforu is a local-first desktop agent orchestrator for macOS / Windows / Linux, and BYOK is not a bolted-on feature — it is the cost model.
- Direct provider connection, no markup. Doforu connects straight to model providers using your keys. Requests travel from your machine to the provider you chose; there is no Doforu server in the request path taking a cut. You can verify this yourself with a network monitor.
- 15+ model families — OpenAI (GPT-6 Astra), Anthropic (Claude Opus 5.5), Google (Gemini 4 Argon), DeepSeek (V4.1), Qwen (Qwen3.8-Max), Moonshot (Kimi K3), MiniMax (M3), Zhipu (GLM-5.3), plus local models via Ollama/vLLM.
- Local models for zero marginal cost. Point Doforu at Ollama or vLLM and the whole workflow runs on your machine, including offline.
- One-time pricing. Doforu is a $49 one-time perpetual license (including 1 year of updates; renewal $49/yr), with a free trial available. Because the license is not tied to inference, your model spend goes to providers at list price rather than through a subscription tier.
Architecturally, Doforu adds something beyond BYOK in a single agent: an orchestrator that splits a goal and dispatches multiple subagents simultaneously, across Plan / Agent / Orchestrator modes. For how that compares to BYOK tools like Cline and Zed, see Doforu vs Cline and Doforu vs Zed; for the editor-native alternative, Doforu vs Cursor.
FAQ
Final Thoughts
BYOK is best understood as a structural property rather than a feature checkbox. It determines who bills you, whose privacy policy applies, and whether your spend is portable when you change tools. In 2026 the honest landscape is that only a handful of tools offer it fully — Cline, Continue.dev and Aider — while the most popular editor and assistant products offer it partially, sometimes only for chat.
The tools that do BYOK well tend to share one trait: they were built to be model-agnostic from the start. That is also the design principle behind Doforu — direct provider connections with no markup, 15+ model families, and local models via Ollama/vLLM as a first-class option. If you are weighing BYOK against the alternatives, our breakdown of perpetual license vs subscription covers the cost side.
Ready to try a local-first desktop agent orchestrator? Download Doforu for macOS, Windows & Linux
