For agents

AgentCloud is built so an autonomous agent can go from a single URL to a deployed, database-backed service with no human in the loop. This page is the one-stop map. Everything here is reachable from https://api.agentcloud.ac.

The one thing to fetch first

GET https://api.agentcloud.ac/skill.md

skill.md is the machine contract: every verb, every error code, the conventions, and a runnable quickstart. If you read nothing else, read that. A terse machine index also lives at https://docs.agentcloud.ac/llms.txt, and the full typed API is at https://api.agentcloud.ac/v1/openapi.

Pick a surface — they're equivalent

All three drive the same API; choose by how you operate.

1. MCP (recommended for tool-using agents)

The conway-mcp server runs locally over stdio, so conway_deploy reads your local filesystem and returns the final URL in one tool call.

# conway-mcp ships with the CLI install (below)
claude mcp add conway \
  -e CONWAY_ENDPOINT=https://api.agentcloud.ac \
  -e CONWAY_TOKEN=<token> \
  -- conway-mcp

Tools: conway_get_skill, conway_list_projects, conway_create_project, conway_status, conway_deploy, conway_operation, conway_logs, conway_scale, conway_rollback, conway_set_env, conway_add_datastore, conway_db_query, conway_create_repo, conway_list_repos, conway_commit_files.

2. CLI

curl -fsSL https://api.agentcloud.ac/install.sh | sh   # macOS/Linux, amd64/arm64
export CONWAY_ENDPOINT=https://api.agentcloud.ac
export CONWAY_TOKEN=<token>
conway --json init my-app          # --json forces machine output (auto when piped)

Direct binary download (no installer): https://api.agentcloud.ac/dl/conway-<os>-<arch> where <os> is linux/darwin and <arch> is amd64/arm64.

3. REST directly

curl -H "Authorization: Bearer $TOKEN" https://api.agentcloud.ac/v1/projects

The canonical flow

conway init blog                                  # idempotent
conway add postgres                               # injects DATABASE_URL everywhere
conway deploy --dir ./blog                        # build + health-gate + live URL
# → https://app-blog.apps.agentcloud.ac

Via REST, deploy returns 202 with an operation_id and stream_url; poll GET /v1/operations/{id} until status is succeeded|failed, or consume the SSE stream (log, phase, done — replays losslessly on reconnect).

Rules that make you self-correcting

• Every error is {code, message, hint, retryable}. The hint tells you the next action. Don't guess — read it. (codes)
• Your app must listen on $PORT (default 8080) or the deploy auto-rolls back with HEALTH_CHECK_FAILED and the log tail.
• Builds need no config: a Dockerfile is used if present, else railpack autodetects Node/Python/Go/Rust/etc.
• Idempotent creates return {"created": false} — safe to retry.
• dry_run=true previews any mutating verb without applying it.
• Datastore credentials are injected for you — never hardcode a connection string; read DATABASE_URL / REDIS_URL / MONGO_URL from the environment.

Delegating to sub-agents

A coordinator can mint a narrower token for each worker — capped to its own capabilities and projects, so a worker can never escalate:

conway tokens issue --principal worker-7 --caps deploy,scale --projects blog

See Tokens & delegation.