Skip to content
Trust center

Security whitepaper

How kanman is built and run: architecture, data flows, credentials, isolation, access control, logging, backups, change and incident handling.

Version 1.0 Updated November 15, 2026

This document is for security, privacy and procurement reviews of kanman. It describes the controls that run today. Use Print or save as PDF above to keep a copy. For questions, or answers to your own questionnaire, write to [email protected].

1. What kanman is

kanman is an AI teammate for engineering teams. It takes stories from your tracker, works on them in isolated runs with a coding tool (Claude Code or Codex), has every change checked by your CI and by an independent reviewer against the story’s acceptance criteria, and opens a pull or merge request. It talks with your team in the app, in Slack and in Microsoft Teams.

kanman does not train or host its own language model. It uses models from established providers, either under your own contract or, if your team chooses, through kanman’s model access.

2. Architecture

Component What it does Where it runs
Web app and API The conversation, the board, settings, the audit log kanman Cloud: EU. Self-hosted edition: your servers
Database and file storage Accounts, stories, runs, decisions, audit log, run evidence kanman Cloud: Stockholm, Sweden. Self-hosted edition: your servers
Run machines A fresh machine per run that checks out the repository and runs the coding tool kanman Cloud: Frankfurt, Germany
Runners The same run host as a container image in your network; pulls work over outbound HTTPS Your infrastructure
Model providers Language models for the coding tool, the reviewer and kanman’s own steps The provider and region your team’s model route names

Runners authenticate with a workspace token, need no inbound ports and hold no database credentials. Nothing calls into a runner.

3. Deployment options and data flows

kanman Cloud Your own runners Everything on your runners Self-hosted edition
Plans All plans Enterprise Enterprise Enterprise
Repository checkout and coding tool Fresh machine per run, Frankfurt Your network Your network Your network
Repository actions (push, merge requests, comments) The run’s machine Your runner Your runner Your runner
kanman’s own model calls kanman’s cloud, on your team’s route kanman’s cloud, on your team’s route Your runner Your runner
What reaches kanman’s cloud Run records Progress, diff summary, commit information, transcript Status, summaries, file names and links; no transcript No application data. An instance you link to kanman fetches releases and reports its versions and update jobs

With “Run everything on our runners” kanman’s cloud never calls a model provider and never reads or writes the workspace’s repositories. When no runner is online, work waits; it never falls back to the cloud.

The self-hosted edition runs the application, database and runners on your servers (Docker Compose or Kubernetes). Model calls go only through your runners, email through your mail relay, and public sign-up is off. Updates are pulled from kanman’s registry or, with Self-hosted Plus, installed from an offline bundle. A signed licence file is verified on the instance itself.

4. Data kanman stores

Data Purpose
Workspace, members and roles Sign-in and access control
Stories mirrored from your tracker Intake, readiness checks, runs
Run records: stages, progress, cost, policy decisions, model and region Run page, reports
Audit log Accountability and export
Team memory and skills Knowledge for later work
Chat messages kanman received and posted Answers and history
Team context and incident data, with connection tokens (encrypted) Answers with sources, incident handling

kanman keeps no copy of your repositories. Code is checked out per run and discarded with the machine. Diffs, file references and short excerpts appear on run pages because reviewers need them.

Retention. Run records, decisions and team data are kept while the team exists. A deleted team or workspace is erased completely and automatically 30 days after deletion; only a minimal erasure record remains. Audit log retention is a workspace setting from 90 days to 5 years. Invoices are kept as long as the law requires.

5. Data residency

  • Application data: EU, Stockholm, Sweden.
  • Cloud run machines: Frankfurt, Germany.
  • Model calls: the provider and region of your team’s route. EU-only processing limits a workspace to EU routes and refuses every other route before a request leaves; each model call and run records the region its model ran in.
  • Runner-only and self-hosted: code, data and model calls stay in your network, apart from what the table in section 3 lists.

6. Credentials

  • The coding tool never holds a code host token, a model key or a run secret’s value. It reaches the model and the repository through a credential proxy on the run’s host, which adds credentials on the way out.
  • Only the host pushes and opens merge requests, after scanning the diff for credentials; a hit stops the push.
  • Each run gets a token that works for that run only and expires with it. With the kanman GitHub App, repository tokens cover one repository and expire after an hour.
  • Run secrets reach only kanman’s own checks, never the coding tool.
  • From the repository, the coding tool loads instructions only; project hooks, environment files and tool servers stay off unless the team allows them.
  • Model keys, connection tokens and secret values are stored encrypted; API tokens only as a hash.

7. Run isolation

  • A fresh machine per cloud run, removed when the run ends; no credentials for other workspaces or for kanman’s own systems.
  • Only repositories the team’s policy allows are cloned. Denied paths and change size limits are enforced while the coding tool works and again on the final diff.
  • Internet access during a run is limited to the code host, package registries, kanman, the model and hosts the team lists. Networks marked as production are refused, also when a name resolves into them. The same limits apply on self-hosted runners.
  • Budgets per run, day and month. Stop all runs stops every run of a workspace within a minute, and nothing is pushed after it.

8. Untrusted text and who may command kanman

  • Text from people without write access to the repository or team (public issue comments, shared channels) never becomes an instruction; at most it becomes a finding for a person.
  • Text nobody on the team wrote (tracker text, review comments, CI logs, dependency docs, tool output) reaches a model cleaned, capped and marked with its source. kanman’s own steps that read it answer only in checked fields.
  • kanman’s own model steps have read-only tools. Actions with effects are proposals that pass a rights check and executed checks before they happen.
  • A change that adds a dependency never merges on kanman’s own decision.

9. Merge control

kanman merges only when your CI is green and a reviewer session that did not write the change approved that exact commit. The reviewer runs the acceptance check or reproduces the bug before and after the change. Your code host’s protected branches and approval rules apply on top, so a team that wants a person to approve every merge sets that on its code host. Acceptance specs, test files and CI configuration are read-only to the coding tool.

10. Redaction and privacy rules

Text from runs, coding tools and models passes a redaction step at every point where kanman stores it, and a run’s exact secret values are removed from everything it reports. Privacy rules per workspace, team and connection mask personal and customer data before any model call, in the cloud and on runners, and a privacy audit shows what was masked.

11. Access control

  • Workspaces are isolated at the database: every table checks workspace membership per row.
  • Roles: owner, admin, member, viewer. Model access, policies, members, tokens, runners and the audit export are limited to owners and admins.
  • Sign-in by password, email link or single sign-on. Single sign-on (Enterprise) supports Microsoft Entra ID, Okta, Google and other SAML or OIDC providers, with group-to-role mapping for Entra ID; enforcing it needs a passing sign-in test. Directory sync (SCIM) is coming.
  • Removing a member ends access immediately; past actions stay attributed in the audit log.
  • Support access to a self-hosted instance is granted by the customer, with a reason and a time limit.

12. Logging and audit

  • Audit log: append-only entries for policy changes, answers, stops, resumes, sign-on events and every action kanman takes; export as CSV or JSON; retention 90 days to 5 years.
  • Run records: stages, progress, cost, policy decisions, model and region.
  • Model call records: step, model, region, tokens, duration and cost for every call; no prompts or answers, only a fingerprint of the input.
  • Tracing: one trace id per run across cloud, runner and coding tool; runners can send traces to your own OpenTelemetry collector.
  • AI marking: commits, merge requests and tracker comments by kanman are marked as AI work.

13. Encryption

All connections use TLS. Stored data is encrypted at rest by the hosting provider. Keys, tokens and secret values are additionally encrypted before storage; API tokens are stored as a hash.

14. Backups and recovery

The kanman Cloud database is backed up daily by the hosting provider; backups are kept for seven days in the same EU region. The self-hosted edition ships backup and restore scripts for Docker Compose and Kubernetes; the restore is tested in kanman’s release pipeline. See the backup policy.

15. Change management

Every change to kanman goes through version control and automated checks: a secret scan, linting, type checks and the tests affected by the change before it is merged, and the full test suite before it reaches production. New versions of the run machines first take a quarter of the runs and are compared with the previous version; if they do worse, runs go back to the previous version automatically. See the change policy.

16. Incident handling

Incidents are contained first (stopping runs, a model route or a run machine version), then assessed for severity and affected workspaces. For an actively exploited vulnerability or a severe incident we follow the EU reporting path (an early warning within 24 hours, a notification within 72 hours) and inform affected customers. Personal data breaches are reported to you as the controller without undue delay, as the DPA sets out. Status and incidents are published at status.kanman.ai. See the incident policy.

17. Sub-processors

The current list with purposes, locations and safeguards is on the sub-processor page (Annex 2 of the DPA). New sub-processors are announced to workspace owners at least 14 days in advance. With your own model key, AI gateway or runner model access, the model provider works under your contract. See the supplier policy.

18. The security floor in plain words

These rules hold on every team and cannot be switched off:

  1. Instructions are not controls. Every limit that matters is enforced outside the model: by kanman’s host, your code host and the network.
  2. Only people with access command kanman. Text from anyone else is data, never an instruction.
  3. The agent proposes, the host writes, credentials stay outside. The coding tool never holds tokens or keys; only the host pushes, after a credential scan.
  4. Redaction at every storage boundary. Nothing from a run, tool or model is stored unredacted.
  5. Untrusted text is data. It reaches a model fenced and marked with its source, and can never add an action.

19. Contacts