Weights & Biases

import { Workspace, MountMode, WandbResource, normalizeWandbConfig } from '@struktoai/mirage-node'

const ws = new Workspace({
  '/wandb': new WandbResource(normalizeWandbConfig({
    entities: ['my-team'],
    api_key: process.env.WANDB_API_KEY,
  })),
}, { mode: MountMode.READ })

The resource also ships in @struktoai/mirage-browser; its implementation uses standard fetch. Browser access depends on the service’s CORS configuration. The mount exposes /<entity>/<project>/<run-id>/ with run.json, config.json, summary.json, history.jsonl, and files/. Configure one or more entities explicitly; project and run listings are fetched on demand and follow service pagination. Run IDs identify directories; duplicate display names remain separate runs. run.json contains the run metadata exposed by the supported W&B API fields:

Fields	Contents
`entity`, `project`, `run`, `path`	Run identity and its `entity/project/run-id` path
`storage_id`, `display_name`, `state`	W&B’s opaque storage ID, display name, and reported state
`tags`, `sweep_name`, `group`, `job_type`	Experiment organization
`commit`, `created_at`, `heartbeat_at`	Source revision, creation time, and last heartbeat
`description`, `notes`, `user`	Run description, notes, and creator’s `id`/`name`/`username`/`email`
`read_only`	W&B’s reported edit permission; the Mirage mount remains read-only
`system_metrics`	Latest system metrics decoded as JSON, retaining nested values
`history_line_count`, `history_keys`, `file_count`	History row count, metric-key metadata including the last step, and file count

For example, the synthetic integration fixture renders this complete run.json:

{
  "entity": "lab",
  "project": "experiments",
  "run": "run-a",
  "path": "lab/experiments/run-a",
  "storage_id": "synthetic-storage-id-a",
  "display_name": "duplicate",
  "state": "finished",
  "tags": [
    "baseline",
    "café"
  ],
  "sweep_name": "sweep-01",
  "group": "ablation",
  "job_type": "train",
  "commit": "0123456789abcdef0123456789abcdef01234567",
  "read_only": false,
  "created_at": "2026-01-01T00:00:00Z",
  "heartbeat_at": "2026-01-01T00:05:00Z",
  "description": "Synthetic experiment metadata",
  "notes": "First line\nSecond line: café",
  "user": {
    "id": "synthetic-user-id-bob",
    "name": "Bob Example",
    "username": "bob",
    "email": "bob@example.com"
  },
  "system_metrics": {
    "cpu": 0,
    "gpu.0.memoryAllocated": 128.5,
    "nested": {
      "missing": null
    }
  },
  "history_line_count": 4,
  "history_keys": {
    "lastStep": 5,
    "keys": {
      "score": {
        "type": "number",
        "count": 3
      },
      "train_step": {
        "type": "number",
        "count": 4
      }
    }
  },
  "file_count": 4
}

Unavailable metadata is null; empty tags, zero counts, false permissions, and original timestamp strings are preserved. These fields are fetched only when reading run.json, in one run query. Listing and stat do not fetch them. The user is the run’s creator, not the identity associated with the API key. When the creator exists, user has id, name, username, and email; W&B requires the creator’s ID and name, while username/email may be null. A missing creator is user: null. Reading these fields does not query the authenticated viewer or discover teams. Machine/runtime details are available at files/wandb-metadata.json when uploaded; reading run.json does not download that file. config.json unwraps W&B’s configuration value envelopes. summary.json contains the stored summary, which may differ from the best value in history. These documents have run-defined keys rather than a fixed Mirage field list. Nested objects, arrays, metric names containing / or ., and null values remain intact. For this fixture, they are {"lr": 0.01, "label": "café"} and {"score": 0.4}. history.jsonl scans full history in _step windows using the public SDK’s unsampled scan protocol. It retains row order, missing keys, nulls, and logged training-step fields. _step and a training-step metric are separate values. An empty window does not end a scan. The last step is captured when a read begins; rows appended later are visible on a subsequent read. One-step query windows are widened to two steps to avoid a live API edge case; results are filtered to the original range without duplicates or later rows. Run files retain their relative paths and are downloaded as raw bytes. Listing and stat do not download histories or files. Rendered JSON and history have unknown sizes; raw file sizes come from W&B’s file metadata. Directory metadata uses Mirage’s index cache (600 seconds by default). A project listing also establishes its runs’ existence; a file catalog populates all nested directories, so traversing them reuses the same metadata. Config, summary, and history reads request their own fields. Downloads resolve the exact filename for a fresh URL on each read. File content and download URLs are not cached in the index. Reads and downloads stream, so head can stop before fetching the entire history. page_size defaults to 100 (maximum 1000); max_pages defaults to 10000. Exhausting a page budget raises an error instead of returning a complete-looking partial result. Recursive du is capped at 1000 entries; glob expansion uses Mirage’s standard traversal limit. API keys use host configuration and are redacted from resource state. No write operations or W&B CLI are registered.

Getting Started

AI Agents

Resources

CLIs

Watch

Runtimes

Weights & Biases