exp-scheduler GPU Lease and Queue Control

    2

    Keep GPU experiment management in the same conversation as your coding work.

    Free

    Security scanned

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    exp-scheduler GPU Lease and Queue Control

    Example session with this skill installed

    Actual user request, translated from Chinese:

    Test first-use checks, installation/startup, and reuse of an existing backend on a real Linux GPU server. Do not package or upload the skill yet.

    • Read your context and instructions
    • Compiled the exp-scheduler gpu lease

    Observed result from the integration test on 2026-10-27, summarized from real command responses. Server addresses, account names and local paths are omitted.

    The server had two NVIDIA GeForce RTX 4090 GPUs. Its default Python was 3.10, so the agent selected an existing Python 3.13.5 interpreter instead of upgrading system Python. A working scheduler was detected even though the selected interpreter did not contain the server packages and no scheduler command was on PATH. Preflight returned ready: true and use_existing_service, so the existing backend was reused without reinstalling or restarting it.

    For the fresh-installation case, the agent used a separate directory, venv, configuration, state database, logs and loopback port. Preflight first reported the absent endpoint as unreachable. After checking existing processes and the authorized target, the agent cloned the public backend, installed its runtime dependencies and initialized the isolated state. A broken proxy configuration and a pip download timeout were diagnosed; command/session-scoped download settings resolved them without changing global network configuration.

    The new backend started as a user-level systemd transient service. All three preflight API checks passed and the backend detected both GPUs. Running preflight again returned use_existing_service; its PID and configuration remained unchanged. A lightweight task submitted only to the isolated queue received ID 1 and completed successfully with exit code 0:

    [exp-scheduler] task=1 gpu=1 started
    skill-first-use-smoke
    CUDA_VISIBLE_DEVICES=1
    [exp-scheduler] task=1 finished status=succeeded exit_code=0
    

    The original backend's PID and configuration were unchanged, and its 353 history records remained intact. The isolated service and temporary proxy forwarding were stopped after the test. This verified installation, startup, GPU discovery, queue execution and service reuse. The smoke task printed a marker; it did not perform CUDA computation or a training workload. Agensi-delivered execution and cross-agent automatic triggering were not exercised.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Keep GPU experiment management in the same conversation as your coding work. This skill connects a shell-capable AI agent to your existing exp-scheduler service so it can inspect queue state, submit experiments, and manage temporary GPU leases using a readable Python helper.

    It supports normal and urgent task submission, GPU and environment-profile selection, task dependencies, metadata changes, queue reordering, cancellation, requeueing, and explicitly requested global pause/resume operations. For direct tests on the GPU server, lease workflows temporarily exclude selected GPUs from this scheduler's launches and attempt to release them after the command exits.

    You provide the Linux NVIDIA GPU server and an agent session with Python command execution and server access. First-use checks inspect Python/API compatibility and, on the GPU host, runtime prerequisites and the existing installation. When the backend is confirmed missing or stopped, the agent can install/start it within your authorization using a dedicated venv and preserving existing state. It verifies the API before continuing and reports the launch method and stop commands. Read-only requests do not trigger setup; privileged OS/driver changes and automatic upgrades are outside this first-use flow. Remote clients can submit jobs through an authorized SSH tunnel. Direct run commands execute on the helper's host, so run those on the GPU server itself.

    A lease coordinates with exp-scheduler; it is not a hardware lock or guaranteed idle-GPU allocator. It cannot block external GPU processes, reserve physical memory, or guarantee cleanup after forced termination. The helper does not wait for idle GPUs or renew leases. TTL expiry does not stop an ongoing command.

    The skill includes instructions, standard-library checker/client scripts, setup and operation examples, and an MIT license. It accompanies the publicly available exp-scheduler project. This is a companion to a self-hosted scheduler; GPU hosting and access credentials are supplied by you. Background setup runs on your authorized server, not on Agensi infrastructure.

    Requires Python 3.11+ and a user-supplied Linux NVIDIA GPU server. Agensi does not supply a shell, GPU, SSH credentials or access to private servers. Setup needs authorized server access and working drivers; privileged OS/driver changes and automatic upgrades are not included. Direct run commands execute on the helper host and must run on the GPU server. Leases coordinate only with exp-scheduler: no hardware lock, external-process eviction, physical-memory reservation, idle-GPU waiting or automatic renewal. TTL expiry does not stop a command; forced termination/network loss can prevent cleanup. The assessed API has no built-in authentication: keep loopback binding and use SSH. Linux integration was tested on Python 3.13.5 with two RTX 4090 GPUs; CUDA training, cross-agent triggering and Agensi-delivered execution were not tested. Skill release: 0.1.0; MIT applies only to the skill directory.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 21 days ago

    • Free with an Agensi account

    Requires

    Python 3.11+
    Linux NVIDIA GPU server
    NVIDIA drivers
    nvidia-smi
    exp-scheduler HTTP service or SSH access
    Git
    Python venv
    ensurepip
    Internet access
    Python

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Network Access
    Environment Variables

    Access is limited to the user-configured exp-scheduler HTTP endpoint and authorized server sessions. Client scripts bypass proxy discovery and have no telemetry. First-use setup downloads the public backend from GitHub and runtime packages through pip, creates a dedicated venv/new configuration as needed, and starts one background process or user service. Existing state is preserved; starting it may resume queued work. No login/boot autostart or automatic upgrades are enabled by default. Agensi-delivered scripts run as unchanged temporary copies. User-supplied commands may access files/network and inherit environment variables; run sets CUDA_VISIBLE_DEVICES by default. Task/lease/queue changes persist on the server. Cancellation, preemption and stop-running interrupt work; deletion removes records and logs; global controls affect the scheduler. These require authorization for the affected scope. The user supplies the GPU server and credentials; keep the unauthenticated API on loopback and use SSH. MIT applies only to the skill directory.

    Listed21 days ago

    What's inside

    Frequently Asked Questions

    Usually not if the agent has shell/SSH access to your GPU server. It checks the environment first, reuses an existing service, or executes installation/startup after the host and scope are authorized. You provide the server and credentials; missing access, drivers or privileged OS changes may need your assistance.

    Yes for queue/API operations when the agent's execution session reaches the server, for example through SSH forwarding. Direct run execution belongs on the GPU host; an HTTP tunnel does not forward process execution.

    Ordinary task submission and leases without stop-running leave existing work running. Interruption commands require authorization for the affected scope. Leasing alone does not make an occupied GPU available for a direct test.

    No. It only changes this scheduler's eligibility rules. Check other leases and actual processes before direct work. TTL is finite and is not automatically renewed.

    The assessed service has no built-in authentication. Keep it on loopback and use SSH access; do not treat it as a public multi-tenant service.

    The skill directory is MIT licensed. That license does not change the license of the scheduler application or other repository files. Marketplace availability does not imply exclusive source access.

    New in Workflow Automation