Coding agents with training opt-out, compared on price

On this page

1. The short answer: Zed ranks first, Augment Code for a contract
2. Opt-out, price, open source and admin controls, checked Sep–Oct 2026
3. How we ranked
4. 1. Zed: off by default, and Business members cannot opt in
5. 2. JetBrains Junie: excluded on every tier, free included, by contract
6. 3. Kilo Code: never trains, and Teams can filter out providers that do
7. 4. Augment Code: no-training in the contract of every paid plan
8. 5. Claude Code: no-training on commercial terms, a toggle on Pro
9. 6. Gemini Code Assist: nothing to opt out of on the business editions
10. 7. goose: open source and free, but your provider decides training
11. 8. OpenCode: paid gateway models don’t train; some free ones may
12. Whose training terms apply when you bring your own key
13. Who can lock no-training for the whole team
14. Which to choose
15. Questions buyers ask
16. Sources

• This page as Markdown

Six of the 8 coding agents we compared let you keep your code out of training: Junie, Zed, Kilo Code, Augment Code, Claude Code and Gemini Code Assist [1]. goose and OpenCode answer “partly”: the model you route to sets the terms [2][3].

On 28 October 2026 we listed 197 agents: 90 confirm a training opt-out and 32 sit under Coding agents [1]. We compared 8 of them, reading every fact from vendor pages and our listings between 2 Sep and 24 Oct 2026 [4].

The short answer: Zed ranks first, Augment Code for a contract

Ranked on the training answer first and entry price second, Zed comes first: Pro is $10 a month, and nothing is retained or trained on unless you rate a response or opt into training collection [5][6]. Augment Code is the pick when you need no-training written into the contract, since every paid plan excludes training on customer data by contract, from $20 a month [7].

• Cheapest paid entry plan in US dollars a month: Zed at $10, of the 4 plans priced that way [1].
• No-training on every tier including free, by contract: JetBrains Junie, from €10 a seat a month (about $11) [8][9].
• Admin filter that removes training providers: Kilo Code Teams, $15 a user a month [10][11].
• Partly, not yes: goose and OpenCode, where the model provider’s terms govern your prompts [2][3].

Opt-out, price, open source and admin controls, checked Sep–Oct 2026

Open source and admin controls do not follow the training answer: 4 of 8 are open source, and 3 of 8 answered yes on admin controls [1].

Facts read 2 Sep to 24 Oct 2026 [1].

Of the 8, 5 have a free plan, 1 has none (Claude Code) and 2 are not recorded (Augment Code and Gemini Code Assist) [1].

How we ranked

We compared 8 of the 197 agents we listed on 28 Oct 2026 [4][1]. The training answer weighs most (yes before partly), then the maker’s own published entry price. Zed, Augment Code, Claude Code and OpenCode publish an entry price in US dollars a month, and those four are compared on it. Junie’s euro seat price and the per-seat prices of Kilo Code and Gemini Code Assist are named as such and left out of that comparison [1].

1. Zed: off by default, and Business members cannot opt in

Best for teams that want a paid plan where training is off unless someone opts in. Pro is $10 a month, and nothing is retained or trained on unless you rate a response or opt into training collection [5][6]. On Business, at $30 a seat a month, both are off by default and members cannot opt in [5][6].

• Who it’s for: developers who want an open-source editor that runs Claude Code, Codex, Gemini CLI or other vendors’ agents natively through the Agent Client Protocol [14].
• Model and key: bring-your-own-key works on every plan including the free one, with keys stored in the system keychain [5]. GPT models are reachable as Zed-hosted inference or through your own OpenAI key [5].
• MCP: Zed works as an MCP client [40]. “Zed uses the Model Context Protocol to interact with context servers.” [40]
• Plans: Personal is free forever; Business has no minimum seat count, and order-form contracts begin at 25+ seats [5]. The Pro trial runs 14 days [5].
• Open source: the editor, its Zeta edit-prediction model and the ACP protocol are open source [12].

Skip it if your team needs VS Code’s extension ecosystem or a managed cloud development environment [14]. Business admin controls do not reach BYOK, gateways, local models, external agents or extensions [13].

2. JetBrains Junie: excluded on every tier, free included, by contract

Best for JetBrains IDE teams that want no-training on every tier, including free. Training on your code is off unless you switch it on, the exclusion is ticked on every tier including free, and the commitment is contractually extended to model subcontractors [8]. It starts at €10 a seat a month (about $11), priced in shared JetBrains AI Credits that also cover other AI Assistant features [9].

• Who it’s for: teams already standardized on a JetBrains IDE who want the agent to share its debugger, project index and version control [16].
• Model and key: six providers accept your own key, billing bypasses JetBrains, and a supplied key wins over the subscription when the same model is reachable both ways [41]. In JetBrains’ words, “you can connect directly to providers like OpenAI, Anthropic, Google, xAI, OpenRouter, or GitHub Copilot.” [41]
• Plans: AI Free is $0 with 3 credits per 30 days; organization plans run AI Pro at $20 a month per seat and AI Ultimate at $60 a month per seat [9]. A 30-day free trial is offered [9].
• Proof: the Trust Center lists a SOC 2 Type II report scoped to Junie [42].

Skip it if you don’t use a JetBrains IDE day to day: Plan Mode’s UI and Agentic Debugging are built into the JetBrains IDEs [16]. Skip it too if you need the agent’s source; it is not open source [15].

3. Kilo Code: never trains, and Teams can filter out providers that do

Best for teams that switch models and want an admin filter on training. Kilo states it never trains on your code and does not retain prompts for its own purposes [10]. Its Teams page lists a control to “Filter out models and providers that train on your data or retain prompts, across every surface your developers use.” [10] Teams is $15 a user a month [11].

• Who it’s for: developers who want to switch models freely, local ones included [18].
• Model and key: the gateway reaches hundreds of models through one endpoint and supports BYOK [43]. “You can bring your own API keys for OpenAI, Anthropic, Google, Groq, and any OpenAI-compatible provider, and pay those providers directly at list prices.” [44]
• Open source: the VS Code extension, JetBrains plugin and CLI are MIT; the Gateway and Cloud backend are source available [17].
• Proof: SOC 2 Type II is claimed, with the report in the Trust Center [44].
• Seats: priced per user with no seat minimum or ceiling stated [11].

Skip it if you need one predictable bill [18]. Inference is billed on top through your own key or Kilo Pass credits from $19 a month to $199 a month, with a 5% processing fee on credit purchases [11].

4. Augment Code: no-training in the contract of every paid plan

Best for engineering leaders who need the no-training promise as a contract term. Every paid plan excludes training on customer data by contract, and the security page repeats it with an indemnification clause behind the claim [7]. In Augment’s words: “Uncompromising security, backed by SOC 2 Type II, rigorous third-party testing, and ironclad IP protection. We never train on our customer’s proprietary code.” [45]

• Who it’s for: organizations running AI agents through full SDLC loops with humans at the quality gates [21].
• Plans: $20 a month, and Business at $100 a month, each with up to 50 seats; larger teams move to Enterprise [7]. A 14-day free trial is listed [7].
• Proof: SOC 2 Type II on every paid plan with annual audits, a stated ISO/IEC 42001 certified AI management system, and CMEK on Enterprise [45].
• Admin: a repo-committed settings file constrains every cloud agent’s tools, and Enterprise admins can set enforced monthly budgets [20].
• Model and MCP: several OpenAI GPT models are available [46]. The agent is an MCP client, and Augment warns that not all MCP servers work with its models [47].

Skip it if you want a standing free plan or open-source agent code: the pricing page publishes no standing free plan, and the agent, Cosmos platform and CLI are closed [7][19]. Usage meters in dollars with a 40% fee on model spend [7].

5. Claude Code: no-training on commercial terms, a toggle on Pro

Best for teams buying Team, Enterprise or API access who want settings an admin can pin. “Anthropic does not train generative models using code or prompts sent to Claude Code under commercial terms, unless the customer has chosen to provide their data to us for model improvement” [22]. Managed and server-managed settings pin permission rules, model allowlists, telemetry and login method, and local settings cannot override them [25].

• Who it’s for: developers and teams who want a terminal-native, scriptable agent that composes with CI [26].
• Plans: Pro is $20 a month billed monthly or $17 billed annually; Max starts at $100 a month [23]. Team is self-serve for 2 to 150 seats [23].
• Model choice: Claude models only; the alias table lists only Claude families [48][26].
• Proof: SOC 2 Type 2 and ISO 27001 artefacts are at the Trust Center, with zero data retention as an Enterprise option [49].
• Residency: US-only inference is selectable on the Claude API at a 1.1x rate [50].

Skip the Pro plan if training must be off by contract. “We will train new models using data from Free, Pro, and Max accounts when this setting is on (including when you use Claude Code from these accounts).” [22] The consumer terms describe training as opt-out in account settings [51]. Even then: “Even if you opt out, we will use Materials for model training when: (1) you provide Feedback to us regarding any Materials, or (2) your Materials are flagged for safety review” [51]. Consumer data from users who allow model improvement is kept 5 years, against 30 days for those who don’t [22].

6. Gemini Code Assist: nothing to opt out of on the business editions

Best for teams already on Google Cloud. “Gemini doesn’t use your prompts or its responses as data to train its models.” [27] Data shared through the Trusted Tester Program “is used for product improvements, not for training Gemini models.” [27] Standard is $22.80 a user a month billed monthly or $19 billed annually [28].

• Who it’s for: teams standardized on VS Code or JetBrains who are already on Google Cloud [32]. Each developer needs an account managed by Cloud Identity, Google Workspace or a federated identity provider [31].
• Plans: Enterprise is $54 a user a month billed monthly or $45 billed annually, and both editions include a 30-day free trial for up to 50 users [28]. Enterprise has a 10-license floor [52].
• Proof: SOC 1, SOC 2 and SOC 3 plus ISO/IEC 27001, 27017, 27018 and 27701, listed for both editions [28].
• Admin: IAM roles gate access, but the agent’s tool allow and deny lists live in each developer’s settings.json [31].
• Limits: each user gets their own allowance rather than a pool, and there is no data residency choice for Code Assist processing [53][31].

Skip it if you want a free individual tier: it stopped serving requests on 18 June 2026 [29]. Skip it too if you want models other than Gemini [28].

7. goose: open source and free, but your provider decides training

Best for developers who want a free agent on their own machine with a choice of providers. goose is Apache 2.0, governed by the Agentic AI Foundation at the Linux Foundation, and works with 15+ providers including local Ollama [33]. You install it with no account and no vendor payment [33].

It ranks “partly” because goose’s own telemetry is opt-in and excludes your code and conversations, but your LLM provider governs what happens to prompts [2]. The docs say “your conversations, prompts, and information accessed by goose might be sent to the provider” [2].

• Who it’s for: developers who want an open-source agent that reuses an existing Claude or ChatGPT plan [35]. A ChatGPT Plus or Pro subscription signs in directly for GPT-5 Codex models [54].
• MCP: goose connects to 70+ extensions, which use MCP [33].
• Admin: an MCP allowlist URL and permission modes, set in per-machine config rather than a central console [34].

Skip it if you need a fully managed setup with vendor support; you choose, configure and pay for your own LLM provider [35].

8. OpenCode: paid gateway models don’t train; some free ones may

Best for teams that want an MIT-licensed agent with 75+ providers and a central config that can pin one approved gateway [37][55][38]. Paid gateway models carry a zero-retention, no-training commitment [3]. OpenCode Go is $10 a month and Go Plus is $40 a month; OpenCode Zen is pay-as-you-go per token [36].

It ranks “partly” because several free and stealth models may use your data to improve the model [3]. For Big Pickle: “During its free period, collected data may be used to improve the model.” [3] Flint 40B Free adds: “Data is retained for 30 days.” [3] NVIDIA’s free Nemotron endpoints state: “Your use is logged for security purposes and to improve NVIDIA products and services.” [3] Other free models, GLM-5.3-Flash Free and DeepSeek V4.1 Flash Free, are documented as zero-retention with no training [3]. Requests to OpenAI models through Zen are kept 30 days under OpenAI’s data policies [3].

• Model and key: paste in your own API key, or log in with a ChatGPT Plus or Pro account [55][37]. Local models are supported [55].
• MCP: OpenCode is an MCP client for local and remote servers [56].
• Proof: a SOC 2 Type 2 report is offered through a trust centre under NDA [57].
• Getting out: refunds within 14 days on unused credits, but a balance left when a workspace is deleted is lost [58].

Skip it if your team might pick a free model: several free and stealth models may use your data to improve the model [3].

Whose training terms apply when you bring your own key

With your own key, the model provider’s policy governs what you send. In goose, conversations, prompts and information it accesses may go to the LLM provider, whose own policies then apply [2]. OpenAI states: “As of March 1, 2023, data sent to the OpenAI API is not used to train or improve OpenAI models” [59].

The route matters too. In Kilo, a ChatGPT subscription connected in the BYOK section goes through Kilo’s gateway: “Requests and their code context pass through the Kilo Gateway before reaching OpenAI.” [60] Kilo also offers a direct option: “Use this option when you want the local OpenAI provider to use your subscription directly, without the Kilo Gateway.” [60]

Who can lock no-training for the whole team

Zed Business locks data sharing so members cannot opt in [6][13]. Claude Code managed settings cannot be overridden by local settings [25]. OpenCode’s enterprise config can restrict the agent to one approved gateway [38].

Kilo Code Teams adds a filter on models and providers that train or retain prompts [10]. Augment Code’s admin lever is a repo-committed settings file for cloud agents’ tools, plus enforced budgets on Enterprise [20]. goose’s controls are per-machine config, and Gemini’s tool lists live in each developer’s settings.json [34][31].

Which to choose

• Paid plan with training off by default, from $10 a month: Zed, with Business locking it for the team at $30 a seat [5][6].
• No-training on a free tier, by contract: JetBrains Junie, with AI Free at $0 [8][9].
• No-training as a contract term with indemnification: Augment Code, from $20 a month [7].
• Admin-pinned settings on commercial terms: Claude Code on Team or Enterprise [25][22].
• Google Cloud identity already in place: Gemini Code Assist [32][31].
• Full control of the model route: goose, or OpenCode with paid gateway models or your own key [2][3]. Kilo Code fits when you want model freedom plus a Teams filter on training providers [18][10].
• Skip all of these for training reasons only if your rule forbids any code leaving the machine; then run a local model, which Kilo Code, goose and OpenCode support [10][33][55].