Developers
Agents your code can call
Start a task from a job, a script or your product. Simple answers come straight back; anything that runs code, a shell or files gets a fresh sandbox of its own. Read its transcript, result and cost. Stop it from your own alerting. Every action stays inside the rules you set.
The routes are built and tested; the hosted endpoint opens with the hosted service.
curl -X POST https://api.rykon.ai/v1/runs \
-H "Authorization: Bearer $RYKON_AI_KEY" \
-H "Content-Type: application/json" \
-d '{
"prompt": "Read the three supplier quotes in the brief and record the cheapest per unit, with the quote it came from.",
"agentId": "8f3a6d21-7e4b-4c90-a5d2-1b6e9c0f7a35",
"maxTurns": 8,
"deadlineSeconds": 900
}'{
"run": {
"id": "0b7c5f2e-4c1a-4f0e-9d2b-6a8e3f1c9d40",
"tenant_id": "5e1d7c40-2b9a-4f6e-8d31-0c7a9e2b4f18",
"state": "queued",
"prompt": "Read the three supplier quotes in the brief and record the cheapest per unit, with the quote it came from.",
"model": "claude-sonnet-5",
"max_turns": 8,
"deadline_seconds": 900,
"sandbox_driver": "kubernetes",
"sandbox_addr": null,
"created_by": "k7m2q9xd",
"created_at": "2026-10-30T14:02:11.408Z",
"started_at": null,
"deadline_at": null,
"finished_at": null,
"stopped_by": null,
"result": null,
"failure_reason": null,
"exit_code": null,
"num_turns": null,
"model_calls": 0,
"input_tokens": "0",
"output_tokens": "0",
"cost_usd": "0",
"reported_cost_usd": null,
"agent_id": "8f3a6d21-7e4b-4c90-a5d2-1b6e9c0f7a35"
}
}Three ways in
Your code, your tools, your machines
- Built
The REST API
58 routes over runs, agents, approvals, the record, kill switches, keys and the account. JSON, scoped keys, one error shape.
API reference - Built
MCP server
Connect Claude Code, Cursor or Claude Desktop with one URL, signed in with your Rykon account: your teams and agents, start work, follow a run, see what is waiting for you.
Connect a client - Your machine
Self-hosted runner
Run agents on your own machine, connected outbound only, under the same approvals, record and kill switches.
How it works
The reference
Everything the console does
The console runs on the same routes your code calls, with one exception kept on purpose: deciding an approval is a person's act.
Browse the reference- Runs6Start a task, answered directly or in a fresh sandbox, follow its transcript, read its result and cost, and stop it.
- Agents6Named agents with a brief, a model, tools from the catalogue and an autonomy level.
- Approvals5Actions waiting for a person, with the exact input that will run. Read them from code; a person decides them in the console.
- The record1Every model call, tool call, approval step, agent change and stop, append only.
- Kill switches3Stop one run, one agent or the whole account, at once, and release a switch when it is safe.
- Spend4What the account’s AI costs this month, against a monthly budget the AI gateway enforces.
- Webhooks15Signed webhooks to your own endpoints when a run ends or an approval waits or is decided, every delivery and attempt kept and replayable, and triggers your systems call to start a task.
- API keys5Scoped keys for your own code: list, open, make, change and revoke.
- Members4The people on the account and their roles.
- Account8The account itself: its name, plan, isolation and run isolation policy.
- Audit log1Who did what to the account, including every refusal, append only.
Runner hosts
Where your runs execute
Anything that runs code, a shell or files gets a fresh sandbox whose network reaches one place, our relay. Your account chooses where those sandboxes live.
The default
Shared
Sandbox nodes set aside for sandboxes only. Every sandbox isolated from every other.
On request
Dedicated
A machine of your own that we run, so your sandboxes share hardware with nobody.
Your machine
Self-hosted
A runner on your machine, connected outbound, for code and data that cannot leave it.
Keys hold only what you give them
Scoped, shown once, stored as a hash, revoked on the next request. A key never holds more than the person who made it.
A person decides
Your code can read every approval. Only a person in the console can approve, edit or reject one, so no script approves its own work.
One source for the docs
The reference, the samples and the OpenAPI file are drawn from one table of the control plane’s routes. Nothing is documented that does not exist.
Webhooks and your own tools
Signed webhooks when a run finishes or an approval is waiting, and your own MCP servers as tools behind an allowlist.
A stop button anywhere
Engage a kill switch on a run, an agent or the account from your own alerting.