Runner Secrets Live on the Platform Now – the Token Is All You Need Locally

Until this week a runner needed four or five local environment variables: API URL, runner token, Anthropic key, GitHub token, sometimes Bitbucket on top. Now it needs two. Everything else is delivered by the platform at startup.

Secrets live on the platform

The runner authenticates against the Spedy API with its rnt_… token and receives its full runtime bundle back: Anthropic key, git tokens, MCP credentials. Values are encrypted at rest with AES-256-GCM and delivered over TLS — they never touch disk on the runner, only process memory and the env vars of individual job containers.

The only things you still need locally:

SPEDY_API_URL=https://api.spedy.ai
SPEDY_RUNNER_TOKEN=rnt_…

When you create a new token, Spedy shows you the exact snippet to paste into your .env and checks right there whether the Anthropic key already sits in the vault.

New Secrets tab under Settings → Runners

A fourth tab Secrets now lives under Settings → Runners. It manages:

• Anthropic API key — the only officially supported LLM provider today. OpenAI and Gemini are not on the list yet.
• Custom env vars — anything else your runner job needs (project-specific flags, feature toggles, …).

Each entry shows:

• Scope + provider + env-var name
• A masked preview (e.g. sk-a****abcd)
• Timestamps for creation, last rotation, last use

The plaintext leaves the database at exactly one point: the runner's /runners/bootstrap call. Even admins can't read it back after saving — to rotate you just enter a new value.

GitHub and Bitbucket come from Integrations

If you already have GitHub or Bitbucket connected under Settings → Integrations, you do not need to create a secret for them. The runner pulls the token straight from the active integration on every bootstrap (and mints a short-lived installation token for GitHub App installs).

A small banner on the Secrets tab makes this explicit so nobody duplicates the token by accident.

Setup wizard: three steps instead of four

The wizard you get to via spedy-runner setup has shrunk:

1. AI backend — "Anthropic API (platform-managed)" or "Local Claude installation (Max/Pro)"
2. Configuration — pool size (how many parallel runners)
3. Review & Connect — create the token, run a readiness check

The old "Git Connections" step is gone — GitHub/Bitbucket come from Integrations automatically. The Anthropic key is no longer collected in the wizard either; instead the Review step shows a checklist:

If the Anthropic key is missing, the wizard still creates the runner token but won't send the bundle to the local CLI until you add the secret. That prevents runners from going online just to fail on the first job.

Migration for existing admins

If you already have runners running, a migration banner shows up at the top of the runners page with two steps:

1. Store your Anthropic API key in the vault — direct link to the Secrets tab.
2. Restart the supervisor once — the banner prints the exact command:

docker compose pull
docker compose up -d --force-recreate spedy-supervisor

The banner disappears automatically as soon as ANTHROPIC_API_KEY exists in the vault. As long as you leave the old value in your local .env, the local value wins over the platform value — so you can do the switchover at your own pace and remove the line later once everything is green. The runner re-fetches the bundle every 15 minutes, so rotations propagate without another restart.

Your keys, your responsibility

Because the runner now talks to Anthropic directly with your key, the Secrets tab carries a reminder at the top:

• The cost is yours. Spedy does not cap, proxy, or subsidise token spend. Whatever the agent burns bills directly to the account behind the key.
• Set limits on both ends. Budgets and rate limits in the Anthropic dashboard; per-job token budgets under Runner Teams inside Spedy. Agent loops can spiral into six-figure cache-read counts if left uncapped.
• No secrets in the repo. The agent reads the full working tree as context. Any .env, private key, or committed token it finds can end up echoed in logs, commits, or PR descriptions.

Summary