ankra CLI, the Ankra MCP server, the REST API and the Terraform provider. They are not four products. The CLI, MCP clients and your own automation all call the same bearer-authenticated /api/v1 API, they authenticate with the same personal and service tokens, and what they do lands in the same audit log. What differs is who drives them and what each one is shaped for.
What each is best at
The four combine. A platform team can declare clusters and organisation settings in Terraform, run the CLI in CI to deploy and wait, and connect engineers’ editors to the MCP server read only.
What each cannot do
Some limits hold for every surface, because they are limits on tokens rather than on a tool:- No token changes the Kubernetes access policy. The CLI and the MCP server can only propose a change and answer an apply link, which an organisation owner or admin opens, signed in, to apply it. See Organisation access policy.
- A pipeline definition’s protected sections are approved by a person. A service account’s token is refused.
- A service token stays in its organisation. Naming another organisation answers “Service account tokens cannot act outside their organisation”, and a token pinned to a project answers “Pinned tokens cannot act outside their organisation”.
- MCP tokens reach the MCP server and nothing else: the REST API, the CLI and Terraform refuse them. An MCP token is pinned to one project and cannot switch organisation (“MCP-scoped tokens cannot override their organisation”). It cannot approve or reject an Ankra AI action that waits for a person, or resume the organisation’s paused AI actions. Tools that only work inside an Ankra chat conversation are not offered. A call has a 55-second deadline, except
workspace_exec; a tool that starts longer work answers with an operation id thatwait_for_executionwaits on. See Ankra MCP server. - The REST API serves tokens under
/api/v1. The portal’s own routes under/organd/chatauthenticate with the portal’s session cookie, never with a token, and need anX-Ankra-CSRFheader on every change. Some portal reads also answer under/api/v1, to a personal token with no scopes and no project pin and to no other token. A request acts in the token’s pinned project, else in the project theX-Ankra-Project-Idheader names, else in the organisation’s Default project. It never follows the project selected in the portal. - The CLI acts with one credential at a time, and a saved
ankra loginwins overANKRA_API_TOKEN. A coding agent that runsankrain your terminal therefore acts as you. See Give a coding agent least privilege. - Terraform manages the objects the provider has resources for. The Terraform page lists them and the provider’s own limits.
How they share tokens
There is one set of tokens, made in two places:- Personal tokens are made under Your account → API access (
/next/account/api), or withankra tokens createandankra mcp connect. They act as you, with your role. - Service tokens are made under Identity & access → API and automation (
/next/organisation/automation), or withankra tokens create --service. They belong to the organisation, act with their service account’s role, and keep working when the person who made them leaves. See Service tokens.
Where the state lives
Ankra keeps the live record of every cluster, Stack and setting. Two of the surfaces add a second record outside Ankra:- A GitOps repository holds each cluster’s configuration as YAML, and sync runs both ways. A change committed there reaches the cluster, and a change made through the portal, the CLI, the API or Ankra AI is written back to a GitHub or Bitbucket Cloud repository, so Git stays the record whichever surface made it. Ankra does not push to a Bitbucket Data Center repository.
- Terraform keeps the objects it manages in its own state, and an apply changes Ankra to match the configuration. A change made to such an object anywhere else differs from that configuration until the next apply.
One audit log
Every surface writes to the same audit log, under the person or service account whose token was used. A service account’s rows name it asservice_account:<name>. Requests from the CLI and from MCP clients also say which client sent them: the CLI names its version and the command, never its arguments or flags, and under a coding agent the agent and a digest of its session. Every call to a tool an MCP token can see, refused calls included, has its own row with the action mcp_tool_call.