Documentation
OpenAPIAgent index

Purchase to first run

Dharma AI gives one company an organization-scoped control plane for local agents, managed agents, evaluations, remediation, and usage. Start at Create an account. An invitation is not required for the first organization owner.

1. Create the organization

  1. Register with your business email.
  2. Choose a trial or production package.
  3. Name the organization and provide the GitHub username that should receive access to its private control repository.
  4. Open Portal -> Agent Fabric -> Instructions.

The portal is the source of truth for the organization ID, entitlement, provisioning state, control repository, policy revision, enabled runtimes, and copyable setup commands. Do not obtain these values from a public example.

2. Choose an execution mode

| Mode | Where the model runs | Credential boundary | Usage | | --- | --- | --- | --- | | Local agent | On an enrolled customer device | Provider credentials stay local | Dharma analysis and orchestration | | Dharma-managed ADK | In the organization runtime managed by Dharma | Customer receives no cloud credential | Model, runtime, tools, traces, storage, analysis | | GCP Vertex BYOK | In the customer's GCP project | Workload Identity Federation; no JSON key | Dharma orchestration and analysis; provider usage remains in customer GCP |

Provisioning is asynchronous. Treat ready as the activation state; do not infer readiness from organization creation alone.

3. Create a developer token

Open Portal -> Developer API -> Keys, choose the minimum scopes needed, and generate one organization token. The token is shown once. Store it in a secret manager or environment variable, never in a repository, screenshot, prompt, or browser URL.

export DHARMA_ORG_ID="<organization-id>"
export DHARMA_API_TOKEN="<token-shown-once>"

Common scopes are agents:read, agents:run, evals:read, evals:run, traces:read, skills:read, skills:write, and usage:read.

4. Submit a managed run

Create or select a managed agent in the portal, then submit work asynchronously. Use a unique idempotency key for every intended operation.

curl --request POST \
  "https://www.dharma-ai.io/api/orgs/$DHARMA_ORG_ID/agent-runs" \
  --header "Authorization: Bearer $DHARMA_API_TOKEN" \
  --header "Content-Type: application/json" \
  --header "Idempotency-Key: first-run-$(date +%s)" \
  --data @request.json

A successful submission returns 202 Accepted and a run identifier. Poll the returned run resource and its ordered events until it reaches a terminal state. A 202 response means accepted, not completed.

5. Verify the operating loop

Before production traffic, verify all of the following in the authenticated portal:

  • the run is linked to the expected organization, agent, revision, endpoint, and trace;
  • input and output artifacts are visible only to authorized organization members;
  • usage is settled once under the expected rate-card version;
  • the run appears in the selected evaluation scope;
  • a failed run can be inspected without exposing provider credentials or private runtime URLs.

Continue with API and SDK, Managed agents, or CLI and local agents.