MCP Server

Jungle Grid MCP Server

Jungle Grid MCP lets MCP-aware agents estimate, submit, monitor, cancel, and retrieve artifacts from Jungle Grid workloads. It supports local stdio clients and hosted Streamable HTTP deployments that forward tool calls to the Jungle Grid API.

Use it for asynchronous AI workload execution, batch processing, training, fine-tuning, uploaded file or script backed jobs, lifecycle diagnostics, workload logs, and managed output artifacts.

Installation

Requirements:

• Node.js 18 or newer
• A Jungle Grid API key for local stdio, or an OAuth bearer token for hosted HTTP
• API scopes that match the tools you want to call

Run the local stdio server with npx:

JUNGLE_GRID_API_KEY=jg_placeholder npx -y @jungle-grid/mcp

Install globally if you prefer a stable executable:

npm install -g @jungle-grid/mcp
junglegrid-mcp

Configuration

Local stdio uses environment variables:

Hosted HTTP gateway deployments also support:

Never commit API keys, OAuth tokens, signed upload URLs, signed artifact URLs, or callback secrets.

Connection Modes

Local stdio

Local clients launch the package and communicate over stdio.

{
  "mcpServers": {
    "junglegrid": {
      "command": "npx",
      "args": ["-y", "@jungle-grid/mcp"],
      "env": {
        "JUNGLE_GRID_API_KEY": "jg_placeholder"
      }
    }
  }
}

Claude Desktop

Add the same mcpServers block to claude_desktop_config.json, then fully quit and reopen Claude Desktop.

macOS:

~/Library/Application Support/Claude/claude_desktop_config.json

Windows:

%APPDATA%\Claude\claude_desktop_config.json

Cursor

For project config, avoid checked-in secrets. Put the key in the environment used to launch Cursor:

{
  "mcpServers": {
    "junglegrid": {
      "command": "npx",
      "args": ["-y", "@jungle-grid/mcp"]
    }
  }
}

For a local uncommitted Cursor config:

{
  "mcpServers": {
    "junglegrid": {
      "command": "npx",
      "args": ["-y", "@jungle-grid/mcp"],
      "env": {
        "JUNGLE_GRID_API_KEY": "jg_placeholder",
        "JUNGLEGRID_API_BASE": "https://api.junglegrid.dev"
      }
    }
  }
}

Hosted HTTP

The HTTP server exposes:

• GET /healthz
• GET /.well-known/oauth-protected-resource
• POST /mcp

Start it locally:

MCP_TRANSPORT=http PORT=3000 JUNGLEGRID_INTERNAL_SERVICE_TOKEN=service_token_placeholder npm start

Hosted MCP clients must send Authorization: Bearer <oauth_access_token> to POST /mcp. The server introspects tokens at /oauth/introspect on the configured API base and requires tool-specific scopes.

Minimal Working Example

Ask your MCP client to call the tools in this order:

{
  "tool": "estimate_job",
  "arguments": {
    "workload_type": "batch",
    "image": "python:3.11-slim",
    "command": ["python", "-c", "print('hello from Jungle Grid')"],
    "routing_mode": "balanced"
  }
}

If the estimate is acceptable, submit the job:

{
  "tool": "submit_job",
  "arguments": {
    "name": "mcp-hello",
    "workload_type": "batch",
    "image": "python:3.11-slim",
    "command": ["python", "-c", "print('hello from Jungle Grid')"],
    "expected_artifacts": ["/workspace/artifacts/output.txt"]
  }
}

Use the returned job_id with get_job, get_job_events, get_job_logs, list_artifacts, and get_artifact.

MCP Tools

The current tool registry exposes these exact tool names:

Accepted workload_type values are inference, training, fine_tuning, and batch. The MCP server forwards fine_tuning to the REST API as fine-tuning. Accepted routing_mode values are cost, speed, and balanced.

Tool Details

estimate_job

Returns classification, route status, capacity source, estimated cost range, availability, and screening details when returned by the API. An estimate is not a reservation and does not guarantee immediate startup.

Common errors: missing workload_type, invalid enum value, authentication failure, forbidden scope, invalid request, upstream API error.

{
  "workload_type": "inference",
  "model_size": 7,
  "image": "pytorch/pytorch:2.4.0-cuda12.1-cudnn9-runtime",
  "command": ["python", "infer.py"],
  "routing_mode": "balanced",
  "notes": "single model inference run"
}

submit_job

Creates an asynchronous job. model_size is an optional size in GB used to select suitable GPU capacity and is forwarded as REST model_size_gb. command is preferably an array of strings. env must be an object with string values and is forwarded as REST environment. input_files and script_files accept arrays of { "input_id": "..." }; string IDs are normalized for compatibility. The current REST implementation supports one uploaded script reference.

Expected response includes job_id, status, queued_at or submitted_at, routing fields, input/script details, and artifact contract fields when returned by the API.

Common errors: missing name, image, or workload_type; invalid workload type; command or args too long; invalid environment values; missing or incomplete input IDs; insufficient funds; unavailable capacity; maintenance; authentication or scope failures.

{
  "name": "transcribe-audio",
  "workload_type": "inference",
  "model_size": 7,
  "image": "python:3.11-slim",
  "command": ["python", "/workspace/scripts/transcribe.py", "/workspace/inputs/audio.ogg", "/workspace/artifacts/transcript.txt"],
  "script_files": [{ "input_id": "inp_script123" }],
  "input_files": [{ "input_id": "inp_audio123" }],
  "expected_artifacts": ["/workspace/artifacts/transcript.txt"],
  "routing_mode": "balanced",
  "metadata": {
    "request_id": "req_123"
  }
}

upload_job_input

Creates a signed upload slot. It does not upload file bytes by itself. Upload the bytes to upload.upload_url using upload.method, then complete the upload with upload.complete_url and the returned upload.token.

kind is an arbitrary string accepted by the API. Use input for normal input files and script for scripts by convention. Script uploads mount under /workspace/scripts/<filename>; input uploads mount under /workspace/inputs/<filename>.

Expected response:

{
  "upload": {
    "input_id": "inp_123",
    "filename": "transcribe.py",
    "method": "PUT",
    "upload_url": "https://signed-upload.example",
    "token": "upload_token",
    "expires_at": "2026-06-11T12:15:00Z",
    "complete_url": "https://api.junglegrid.dev/v1/job-inputs/inp_123/complete"
  }
}

Common errors: missing filename, invalid filename, file too large, upload storage unavailable, authentication or scope failure.

list_job_inputs

Returns uploaded inputs with input_id, filename, content_type, size_bytes, kind, status, ready, mount_path, and timestamps when available.

list_jobs

Returns jobs, limit, next_cursor, and has_more. limit is capped by the API. status is a free-form filter string passed to the API; do not assume the MCP schema restricts it to a fixed enum.

get_job

Returns the current job status and details. Status, execution phase, lifecycle events, runtime details, and workload logs are separate surfaces.

Important response fields include status, phase, execution_phase, status_message, status_reason, phase_started_at, phase_last_updated_at, wait_duration_seconds, delayed_start, delay_reason, scheduling, startup_diagnostics, provider, artifacts_ready, failure, input_files, script_file, and artifact_contract when present.

get_job_events

Returns lifecycle events before and during execution. Events may exist before workload logs begin. Events include IDs, types, phases, titles, messages, source, level, timestamps, sequence, and a generated timestamp.

Use events to diagnose queueing, route selection, scheduling, provider provisioning, input preparation, startup, retries, failures, and cancellation.

get_job_logs

Returns stored log entries with items, next_cursor, has_more, failure_highlight, and usage_hint when available. Entries include entry_id, source, category, stream, message, truncated, and created_at when returned by the API.

Logs can be empty while a job is queued, scheduling, provisioning, or preparing. Call get_job_events when logs are empty but the job is not terminal. This MCP tool fetches persisted logs; it does not provide true streaming.

cancel_job

Requests cancellation for a pending, queued, assigned, starting, or running job. Completed, failed, rejected, or already cancelled jobs return a conflict from the API.

Expected response includes job_id, status, and status_reason when returned by the API. Cancellation may trigger managed teardown, but do not assume immediate infrastructure shutdown.

list_artifacts

Returns managed artifacts for a job. Artifacts include artifact_id, job_id, filename, content_type, size_bytes, status, ready, and timestamps when returned by the API. Failed jobs may have no artifacts or partial artifacts.

get_artifact

Creates temporary download information for one artifact. The API returns artifact metadata, a signed URL, and expires_at. Treat the URL as a secret.

Common errors: artifact not found, artifact not ready, artifact storage unavailable, forbidden job, authentication failure.

Production Workflows

Simple Job

1. Estimate:

{
  "workload_type": "batch",
  "image": "python:3.11-slim",
  "command": ["python", "-c", "from pathlib import Path; Path('/workspace/artifacts/output.txt').write_text('done')"],
  "routing_mode": "balanced"
}

1. Submit:

{
  "name": "simple-artifact-job",
  "workload_type": "batch",
  "image": "python:3.11-slim",
  "command": ["python", "-c", "from pathlib import Path; Path('/workspace/artifacts/output.txt').write_text('done')"],
  "expected_artifacts": ["/workspace/artifacts/output.txt"],
  "routing_mode": "balanced"
}

1. Monitor:

{ "job_id": "job_123" }

Call get_job, get_job_events, and get_job_logs with the same job_id until the status is terminal.

1. Retrieve:

{ "job_id": "job_123" }

Call list_artifacts, then:

{
  "job_id": "job_123",
  "artifact_id": "art_123"
}

File-Backed Job

1. Create upload slots:

{
  "filename": "transcribe.py",
  "content_type": "text/x-python",
  "kind": "script"
}

{
  "filename": "audio.ogg",
  "content_type": "audio/ogg",
  "kind": "input"
}

1. Upload each file to the returned signed upload_url, then complete it:

curl -X PUT "$UPLOAD_URL" \
  -H "Content-Type: text/x-python" \
  --data-binary @transcribe.py

curl -X POST "$COMPLETE_URL" \
  -H "Authorization: Bearer $JUNGLE_GRID_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "token": "upload_token",
    "filename": "transcribe.py",
    "content_type": "text/x-python",
    "size_bytes": 1234,
    "etag": "optional-etag"
  }'

1. Submit with input IDs:

{
  "name": "file-backed-transcription",
  "workload_type": "inference",
  "image": "python:3.11-slim",
  "command": ["python", "/workspace/scripts/transcribe.py", "/workspace/inputs/audio.ogg", "/workspace/artifacts/transcript.txt"],
  "script_files": [{ "input_id": "inp_script123" }],
  "input_files": [{ "input_id": "inp_audio123" }],
  "expected_artifacts": ["/workspace/artifacts/transcript.txt"]
}

1. Monitor with get_job_events, get_job, and get_job_logs.
2. Retrieve /workspace/artifacts/transcript.txt with list_artifacts and get_artifact.

Error Shape

REST MCP routes return an envelope:

{
  "ok": false,
  "error": {
    "code": "INVALID_REQUEST",
    "message": "name, image, and workload_type are required"
  }
}

The MCP server converts API errors into tool errors like:

submit_job failed: INVALID_REQUEST: name, image, and workload_type are required

Common API codes include UNAUTHORIZED, FORBIDDEN, INVALID_REQUEST, JOB_INPUT_NOT_FOUND, JOB_INPUT_NOT_READY, ARTIFACT_NOT_READY, NOT_FOUND, CONFLICT, INSUFFICIENT_FUNDS, MAINTENANCE_ACTIVE, and INTERNAL_ERROR.

Security

• Keep API keys and OAuth tokens out of prompts, source control, browser bundles, logs, and issue trackers.
• Prefer host secret stores or local-only MCP config files for JUNGLE_GRID_API_KEY.
• Treat signed upload and artifact URLs as temporary bearer secrets.
• Do not print environment variables that contain tokens from workload code.
• Review submit_job and cancel_job requests before allowing an agent to execute them, because they can spend credits or stop active work.

Development

npm install
npm run build
npm test

Run stdio from the built package:

JUNGLE_GRID_API_KEY=jg_placeholder node dist/index.js

Run HTTP locally:

MCP_TRANSPORT=http PORT=3000 JUNGLEGRID_INTERNAL_SERVICE_TOKEN=service_token_placeholder node dist/index.js

Inspect with MCP Inspector:

JUNGLE_GRID_API_KEY=jg_placeholder npx @modelcontextprotocol/inspector node dist/index.js

Full Documentation

Public Jungle Grid documentation: https://junglegrid.dev/docs

MCP documentation page: https://junglegrid.dev/docs/mcp

License

MIT