nightjar
Nightjar / Connect your assistant
Connect your assistant
Delegate a task to Codex or Claude, then collect the result wherever you work. Saved agents are optional.
Before you connect
Sign into Nightjar, choose your workspace, and connect your own Codex or Claude account. You need a ready Environment and an active model subscription with remaining capacity. This is a limited private beta; new users may need an invitation or operator setup. Read the current pilot limits.
Hosted MCP
https://app.usenightjar.com/mcpTransport: Streamable HTTP over HTTPS. Authentication: OAuth with PKCE (S256) or a workspace API key. OAuth discovery and dynamic client registration are available. Connect to this address directly; do not use the marketing domain as your MCP endpoint.
Using a compatible assistant
- Add the hosted MCP endpoint in your assistant’s remote connector settings.
- Sign in to Nightjar and select the workspace to authorize.
- Choose Delegate and manage runs to submit work, or Read only to inspect existing runs.
- Ask for a small task. Keep the returned run link and collect its result.
First request
Use Nightjar to ask Claude for a short launch checklist.
Use no service connections and a two-minute timeout.
Give me the run link, then collect the completed result.For Muse
Nightjar is preparing a connector submission to Meta. A listing or approval is not yet claimed. The submission uses Existing MCP and the same endpoint above; Muse-specific onboarding and end-to-end execution must be verified during review.
Tools
| Tool | Purpose |
|---|---|
get_environment |
Inspect ready accounts, Environments, and available connections. |
create_run |
Start work from a prompt. |
get_run, list_runs |
Follow progress and read results. |
follow_up_run |
Continue from a previous goal and result. |
answer_run |
Answer a saved question. |
recover_run |
Request bounded Environment recovery. |
cancel_run |
Request cancellation. |
list_agents |
Find optional reusable instructions. |
Collect the result
Submission returns a durable run ID. Call get_run with
waitSeconds: 20 and repeat while work is active. Read all result pages and
check lifecycle.resultReady. Show saved questions to the user. Disconnecting
your assistant does not cancel the task, and MCP alone does not wake a closed chat.
Use an idempotencyKey when creating a run. Retry an uncertain submission with
the exact same body and key. Choose service connections explicitly; start with
connectionIds: [].
API and SDK
The HTTPS API and TypeScript SDK provide the same task lifecycle, plus signed webhook delivery. Store keys in your application’s secret manager, never in browser code.
API and SDK documentation ↗
Full MCP guide and client commands ↗
Permissions and limits
Client access is scoped to a workspace and may include its existing run history. Revoking an MCP grant or removing workspace membership blocks future access. Tokens refresh independently of your browser session. Write-capable connections may act without a separate human approval for each tool call.
The pilot runs on one host with SQLite and does not promise high availability. Selected connections limit service access, not machine isolation. Current hosted capacity is time- and budget-limited; consult the pilot contract before relying on it.
Get support · Privacy · Terms