What model providers see
Which language model providers receive data during a run, what they receive, and how to use your own contract.
kanman does not train or host its own language model. The coding tool and the independent reviewer use models from established providers. This page explains which provider sees what, so you can assess it before you connect a repository.
Which provider is involved
The provider depends on the executor your team uses and on how you pay for compute.
| Executor | Provider (pay through kanman) | Provider with your own key |
|---|---|---|
| Claude Code (default) | Anthropic | Your Anthropic contract, Claude models through your Amazon Bedrock contract, or your company’s AI gateway with an Anthropic-compatible endpoint |
| Codex | OpenAI | Your OpenAI contract, your Azure OpenAI contract, or your company’s AI gateway with the OpenAI Responses API |
The executor is chosen when you set up a team and can be changed on the Team tab under Models > Coding tool.
The reviewer deliberately uses a model from a different vendor than the executor, so that the work is checked by an independent second opinion. If your executor is Claude Code, the reviewer runs on an OpenAI model, and the other way round.
Teams that use their own key get the reviewer on that key too, so their code only reaches their own provider contract. Because one key reaches one vendor, the reviewer then uses a model of that vendor (with Azure OpenAI or Amazon Bedrock, the deployment or model you configured) instead of the other vendor’s. If the key is missing or does not work, the review reports an error; kanman never switches to its own model access for it.
Teams on a self-hosted runner can also leave model access entirely to the runner’s environment, for example through their own Amazon Bedrock, Google Vertex AI or Microsoft Foundry account. See Model access on self-hosted runners.
Intake (turning a requirement into stories) also uses a language model. It uses your own key when the team or the workspace has one, otherwise kanman’s model access.
What new teams use
Owners and admins choose how new teams reach the models in Workspace settings > Model access (/<workspace>/settings/models), under New teams use:
- Your own key (the default): your provider bills you directly. You add the key once, on the same page.
- Pay through kanman: provider cost plus 15 percent on the team’s invoice, nothing to set up. Coming soon: until it is available, new teams use your own key.
- Your runners’ model access: only where runs execute on your own runners. The runners use the model access of their own environment, see Model access on self-hosted runners.
The choice saves at once. The team setup starts each new team with it, and a team can still choose otherwise in its Team > Models > Model access. There the field shows Workspace default: with the current default, and Use the workspace default brings a team that chose otherwise back to it. Existing teams keep what they use until you change it for them.
Using your own key
To get a key, create an API key in your provider’s console: the Anthropic Console (API keys), the OpenAI platform (API keys), your Azure OpenAI resource (Keys and Endpoint) or Amazon Bedrock with an access key of your AWS account. Then add it here.
A workspace has one provider key, which teams that are set to Use your own key share. Only workspace admins can add, test, replace or remove it.
- Open Workspace settings > Model access (
/<workspace>/settings/models) and find Your provider key. - Click Add provider key.
- Choose the Provider and enter the key first. Each provider is listed with the executor it works with: Anthropic and Amazon Bedrock for Claude Code, OpenAI and Azure OpenAI for Codex. For your company’s AI gateway or a proxy, choose Use a gateway or proxy instead below the list (see Use your company’s AI gateway).
- Fill in the other fields for that provider (see the table below). To pin a specific model, open Choose a model; otherwise kanman picks a fitting model for each task.
- Click Test connection to check the key, then Save key.
| Provider | What you enter |
|---|---|
| Anthropic | Anthropic API key. A Model is optional, under Choose a model; without one, each team’s policy chooses. |
| OpenAI | OpenAI API key. A Model is optional, under Choose a model. |
| Azure OpenAI | Azure endpoint (the https address of your Azure OpenAI resource, for example https://your-resource.openai.azure.com), the Azure OpenAI key and the Deployment name. |
| Amazon Bedrock | AWS region where Bedrock model access is enabled (for example eu-central-1), a Bedrock API key and the Model or inference profile ID. Use a Bedrock API key, not IAM access keys. |
| OpenAI-compatible gateway | The gateway’s OpenAI-compatible base URL, the Gateway API key, how the key is sent, optional extra headers, an optional Anthropic-compatible endpoint and the model names. See Use your company’s AI gateway. |
After saving, the key shows its status (Working or Check failed), the model and when it was last checked. Use Test connection again at any time, Replace to enter a new key, or Remove key.
The key must fit the team’s executor: a team on Claude Code needs an Anthropic or Bedrock key (or an AI gateway with an Anthropic-compatible endpoint), a team on Codex an OpenAI or Azure OpenAI key (or an AI gateway with the OpenAI Responses API). kanman never falls back silently to its own model access. If the key is missing, removed or does not fit, the team’s runs stop with a clear message until you add a fitting key.
Keys are stored encrypted and are only handed to the run’s sandbox as environment values. They never appear in logs, run records or the audit log.
Use your company’s AI gateway
Many companies send all model traffic through one AI gateway, for example LiteLLM, Portkey, Kong AI Gateway, Azure API Management, Cloudflare AI Gateway or an internal proxy. If your gateway offers an OpenAI-compatible API, kanman can use it with your company’s key: the gateway applies your company’s rules, routing, logging and contracts, and the provider bills you through it.
Add the gateway
- Open Workspace settings > Model access and click Add provider key (or Replace next to the current key).
- Below the provider list, choose Use a gateway or proxy instead. The form then shows Gateway or proxy (an OpenAI-compatible gateway).
- Fill in the fields:
| Field | What you enter |
|---|---|
| Gateway API key | The key your gateway issued, for example per project or profile. |
| OpenAI-compatible base URL | The address your tools use as the OpenAI base URL, for example https://ai-gateway.corp.example/v1. Any path works, including a profile or tenant in the path. kanman adds /chat/completions (and /responses for Codex) to it. |
| How the key is sent | Authorization: Bearer (the usual) or In its own header, for example api-key or x-api-key, as your gateway expects it. |
| Extra headers | Optional, for example a tenant or project header your gateway requires. Up to 10. The values are stored encrypted, like the key. |
| Anthropic-compatible endpoint | Optional. Only if your gateway also offers the Anthropic Messages API, for example https://ai-gateway.corp.example/anthropic. The Claude Code coding tool needs it. |
| Model names | The names your gateway expects, which often differ from the vendors’ names, for example openai/gpt-5 or vertex/claude-sonnet. Drafting and conversation is required. Reviewer and spec author and Coding tool are optional and use the drafting model when left empty. |
- Click Test connection, then Save key.
What Test connection checks
kanman sends a tiny request to the gateway and lists what it offers:
| Line | Meaning |
|---|---|
| Chat Completions | Required. kanman’s own steps (drafting stories, replies in the conversation, the reviewer) use it. |
| JSON mode | Whether the gateway accepts structured answers. If not, kanman asks for the structure in the prompt instead; nothing else changes. |
| Tool calling | Whether the gateway’s model can call tools through Chat Completions. Shown for your information. |
| Responses API (Codex) | Whether the gateway offers the OpenAI Responses API. The Codex coding tool needs it. |
| Anthropic Messages API (Claude Code) | Whether the Anthropic-compatible endpoint answers. The Claude Code coding tool needs it. |
Below the list, kanman names the coding tools that work with this gateway. The key card keeps the result of the last check.
Which coding tool works with which gateway
| Your gateway offers | Coding tool |
|---|---|
| An Anthropic-compatible endpoint (Anthropic Messages API) | Claude Code |
The OpenAI Responses API (/responses) |
Codex |
| Only Chat Completions | No coding tool. Drafting, conversation and reviews work. |
Current versions of Codex no longer support Chat Completions, so a gateway with only Chat Completions cannot drive a coding tool. In that case, ask your gateway team to also offer the Anthropic Messages API (/v1/messages) or the OpenAI Responses API. LiteLLM, for example, offers both. Until then, a team can keep its assistant steps on the gateway and run the coding tool with another model access (see below).
kanman does not let a team combine a coding tool with a key it cannot run on: the team setup and the Team tab say why and do not save the choice.
Choose it per team, per purpose
In Team > Models > Model access a team chooses its model access twice. Both start with the workspace’s choice under New teams use (see What new teams use):
- Model access of the coding tool: pay through kanman, your own key, or (on self-hosted runners) the model access of the runner environment.
- My own steps use other access (kanman’s assistant steps): drafting stories, replies in the conversation, the reviewer and the spec author. By default Same as the coding tool. You can choose another, for example your company’s AI gateway for the assistant steps while the coding tool uses your runner’s model access.
Cost and usage
Gateways rarely report prices. For runs on a model kanman knows (also with a provider prefix, for example openai/gpt-5), kanman estimates the cost as usual. For models without a known price, kanman records the tokens and the run page shows Model usage in tokens instead of euros; euro budgets cannot stop such runs, the turn limit of each run still applies. Your gateway’s own reporting is the place for the real cost.
Keys of the type “OpenAI-compatible” from earlier versions of kanman are treated as gateway keys: test them once to see which coding tools work.
Gateways inside your network
kanman’s cloud only calls gateways at a public HTTPS address. If your gateway is reachable only inside your network or VPN, turn on Run everything on your runners: your runners then call the gateway, internal addresses (also plain http inside your network) are accepted, and Test connection runs on your runner. The form shows Checking with your runner… until it answers; this needs a current runner version. To configure a gateway directly on the runner instead, see AI gateways in the runner environment.
Models for kanman’s helpers
kanman starts helpers for a story: workers that write the change and reviewers that check it. Models for kanman’s helpers decides which models they run on:
- Automatic (the default): kanman picks Haiku 5.5, Sonnet 5.5 or Opus 5.5 for each task by its complexity. Haiku is fast for mechanical edits and lookups, Sonnet handles normal code changes and reviews, Opus is for architecture, hard debugging and risky changes.
- At most …: kanman still picks per task, but never above the model you choose. Use it to keep the cost of a team down.
- Always …: every helper runs on the model you choose.
Owners and admins set the workspace’s choice in Workspace settings > Model access. A team follows it unless it chooses its own in Team > Models > Models for helpers: there Workspace default shows what the workspace uses, and the team’s own choice overrides it. A helper never runs on a larger model than kanman itself uses for the team.
Your people’s own questions
A question someone asks kanman about a team they are in runs on that team. A question about no team (small talk, a question about the workspace) runs on a small allowance per person and month. Owners and admins set it in Workspace settings > Model access > Your people’s own questions, in whole euros: 3 € by default, 0 turns such answers off. When a person’s allowance is used up, kanman tells them privately. Such questions are billed like a team’s own use: to the person’s first team with a paid plan. For people in no paid team, pick the team that pays for them in the same card (Team that pays for people in no paying team). Without that team, the workspace’s own model key answers them; without either, kanman tells the person that such answers are off and an admin where to change it. In a workspace that runs everything on its own runners there is no allowance: such questions run on the person’s home team. See Who pays for your questions.
Weekly pace and Boost
Some model access comes with a weekly limit. When a team’s model access has one and kanman has seen how much of it is used, kanman spreads the team’s usage over the week: each day adds a share, and unused days carry over. On a busy day the team slows down instead of using up the week and stopping for days.
Team > Models > Weekly pace shows where the team stands:
- On pace: kanman works at its normal speed.
- Slower today: today’s share is nearly used, so kanman works on one thing at a time.
- Paused: today’s share is used. The card shows when kanman continues, and kanman says so in the team chat once. Nothing is lost; it picks the work up again on its own.
- Full speed: Boost is on.
A bar shows how much of the week is used against how much the team may have used by the end of today, and when the new week starts. Teams whose model access has no weekly limit are not paced.
The card also says how much of the team’s model use in the last seven days went to people’s own questions about the team (“Personal asks: 4 %”), in total only, never per person. Each person’s questions share the team’s day with a small share; when it is used up, the team’s own work goes first and kanman tells the person privately. See Who pays for your questions.
Owners and admins can switch on Boost to run at full speed instead, optionally until a day (Until). Leave the day empty to keep Boost on until you switch it off. With Boost the week’s allowance can run out before the week ends.
What the coding tool sends
While it works, the coding tool sends its provider the context it needs for the task:
- the story: title, description, acceptance criteria and test plan,
- the policy summary for the run (denied paths, limits, budget),
- the entries of your team’s memory and skills that apply to the work,
- the files it reads from the repository, and the commands and outputs of its own tool calls (for example test output),
- the changes it writes.
It does not receive secret values (see Sandbox and secrets), other teams’ data, or other repositories.
What the reviewer sends
The reviewer receives the story, its acceptance criteria, the test plan and the diff of the change. It answers with a verdict per criterion, with file and line references.
What intake sends
When you submit a requirement in the app, in a chat with kanman or from a Slack thread, that text is sent to the model that drafts the stories, together with the team context needed to draft them.
What chat and team context send
When kanman takes part in your team’s conversations in Slack or Microsoft Teams, the messages it answers are sent to the model, together with the board state needed for the answer. When the team has team context switched on, kanman reads each source with the team’s own key, or with kanman’s model access if the team has none, and sends the passages that fit a question along with it. Only the team’s own OpenAI key is used to search sources by meaning. If your workspace runs everything on its own runners, reading sources and answering questions happen on your runners.
What providers do with it
What a provider may do with the data is governed by that provider’s terms:
- Pay through kanman: kanman calls the provider under kanman’s business agreement. These agreements do not allow the provider to train models on your data. The providers kanman uses are listed as subprocessors in the DPA (see DPA and subprocessors).
- Your own key: calls run under your own contract with the provider. Your terms, your region choice and your data processing agreement apply. This is the usual choice for regulated teams.
Reducing what is sent
- Keep secrets out of stories and repositories, as you would for any developer.
- Use
denied_pathsto keep the coding tool away from areas it should not change, for example infrastructure or customer data fixtures. - Use your own Azure OpenAI or Amazon Bedrock contract if model calls must stay in a specific region, and switch on EU-only processing so that kanman refuses every route outside the EU.
- Use a self-hosted runner if the checkout and the builds must stay in your network. Note that model calls still go to the provider you choose.
- Turn on Run everything on your runners if kanman’s cloud must not call a model provider at all: the reviewer, the spec author, intake drafts and maintenance then call the model from your runner, with your account.
Related pages
Last updated: January 1, 0001
Open kanman