MCP connector
MCP 0.2.14 connects your existing agent to Maglev through a local stdio process. Your agent keeps its model, files and permissions. A live key enables real project and delivery workflows; sample capability quotes remain sandbox-only.
Install and connect
Section titled “Install and connect”Connect your agent installs CLI 0.6.14. Run maglev setup from the hardware project folder, choose Claude Code or Codex, then reopen the host and accept its project connection. Setup pins the hosted MCP 0.2.14 archive and stores no key in project configuration.
For another MCP host, install directly:
npm install -g https://trymaglev.com/downloads/maglev-mcp-0.2.14.tgzmaglev-mcp --versionExpected version: 0.2.14. Archive · SHA-256. The npm registry still serves the older 0.1.2 release. Choose stdio with command maglev-mcp; no arguments are needed. Use an absolute executable path if the host cannot find it.
The server reads your local maglev login profile, or MAGLEV_API_KEY from the host’s secret/environment settings. A remote host needs its own login or environment. Never put keys in tool arguments or prompts. Running the process alone waits for a client; there is no hosted HTTP MCP URL.
Start with your project
Section titled “Start with your project”Ask your connected agent:
Check my Maglev identity. Read the project files I have authorized, preserve the current revision and sample quantity, and record what is missing. Find relevant manufacturers and prepare the next action. Show me the recipient and complete message before contacting anyone.
The agent can use get_identity, project tools, list_manufacturers, knowledge and action tools. It should read list_action_kinds before submitting an action, retain record IDs and reuse the same idempotency key after an interrupted action request.
Tools and effects
Section titled “Tools and effects”| Tools | Mode and effect |
|---|---|
read_local_docx |
Selected Word paragraph/table evidence and omissions; no upload, signature or legal verification |
read_local_pdf |
Selected local PDF text/page evidence; no upload, OCR, geometry or signature verification |
inspect_local_bom, review_local_bom |
Local CSV/TSV/XLSX evidence under MAGLEV_TRANSFER_ROOT; no key or network, explicit mapping |
expand_local_bom |
Source-linked assembly and purchase calculation; explicit current batch, parent records, groups and allowances; no purchasing authority |
list_manufacturing_stages |
Read stage requirements and availability; does not advance production |
start_interview, recover_interview, get_interview, answer_interview |
Sandbox/live collection with a retained request UUID, local creation recovery, sourced facts and versioned answers; no automatic publication |
get_identity |
Sandbox or live; returns scopes and expiry, never the key |
list_projects, get_project, create_project, update_project |
Sandbox or live briefs; updates support expectedContextVersion |
list_manufacturers, get_manufacturer |
Directory records; contact visibility depends on scope |
search_manufacturing_knowledge, read_manufacturing_knowledge |
References with source context; documents cannot grant permissions |
list_action_kinds, request_action, list_actions, get_action |
Action schemas, drafts, deterministic outputs or coordinated tasks; request needs confirm: true and an idempotency key |
wait_for_action |
Observe one saved action; only running work is polled, review and coordination states return immediately; no new work or approval |
list_supplier_requests, get_supplier_request |
Read existing buyer-owned RFQs and supplier-submitted quote versions; drafts remain private; no sending or acceptance |
wait_for_supplier_quotes |
Bounded observation of new quote IDs on one owned RFQ; cancellation and errors stop the wait, no dispatch or approval |
list_deliveries, get_delivery |
Read owned drafts and durable dispatch receipts |
review_delivery |
Live only; exact resolved recipient, message and project review; sends nothing |
approve_delivery |
Live only; can send a real message, requires the approved digest and confirm: true |
retry_delivery |
Live only; recovers the same approved dispatch, requires confirm: true |
reject_delivery |
Live only; rejects a draft, cannot cancel an approved send |
get_engineering, preview_engineering |
Live and owner-granted; reviewed package and nonbinding supplier-rule preview |
prepare_engineering_transfer |
Local only; selected technical files under MAGLEV_TRANSFER_ROOT, no upload |
propose_engineering_transfer, get_engineering_transfer, upload_engineering_transfer |
Live, explicit consent and separate server owner authorization; resumable exact-file transfer |
discover_capabilities, inspect_capability, request_quote, list_runs, get_run |
Sandbox catalog and saved sample requests; never supplier contact or production |
There are 45 tools. Catalog samples, directory records, engineering estimates and firm supplier offers are distinct. The connector does not parse native CAD, sign agreements, buy shipping labels, place orders or reserve production capacity.
Approve real communication
Section titled “Approve real communication”Read delivery execution. Show the principal every field from review_delivery, including recipient and copied recipients. Only after approval, pass its exact reviewSha256 as expectedReviewSha256 to approve_delivery with confirm: true.
The API rejects changes to the reviewed contact, content or project. A conflict requires a fresh principal review; the agent must not fetch a new digest and approve automatically. Missing transport leaves the message approved but unsent. A provider-accepted receipt does not prove recipient delivery or a supplier reply.
Recover interrupted work
Section titled “Recover interrupted work”Follow idempotency and recovery. Connector-specific cases:
- Selected files: keep the same transfer ID and inspect its server receipt before resuming.
- Cancellation: stops the local request; accepted server work may still exist.
The connector never retries writes automatically. It preserves API error codes, request IDs and recovery details. Responses over 2 MiB fail explicitly rather than returning a truncated success.
Progress and troubleshooting
Section titled “Progress and troubleshooting”Hosts that request progress receive standard notifications/progress events for the current operation. The host decides how these appear; protocol stdout contains no terminal animation. Activity is scoped to each request and stops on cancellation.
| Symptom | Next step |
|---|---|
missing_key |
Run maglev login on the machine running the MCP process, or supply its environment secret |
sandbox_required |
Catalog samples need a sandbox key; use project/action tools for live work |
live_required |
Use a live organization key for delivery and granted engineering access |
approval_review_changed |
Show a fresh complete review to the principal before another approval |
delivery_reconciliation_required |
Do not resend; reconcile the saved provider outcome |
| Unknown tool | Check maglev-mcp --version; install the hosted 0.2.14 archive and restart the host |
| Different project configuration | Setup preserves existing entries; review the Maglev entry before changing it |
Verification uses a real stdio client, synthetic HTTP fixtures and the platform’s disposable end-to-end fixture. Real supplier execution additionally requires configured transport and authorization for that communication.
Word evidence
Section titled “Word evidence”read_local_docx reads explicitly selected Word files locally. See Word evidence for selection, limits, confidentiality and unsupported content.
Supplier replies
Section titled “Supplier replies”Use list_supplier_requests and get_supplier_request to read existing managed RFQs and supplier-submitted quote versions. Both tools are read-only. See supplier replies for buyer permissions, history and commercial limits.
Wait for new quote records
Section titled “Wait for new quote records”CLI 0.6.14 / MCP 0.2.14 adds a read-only wait for new supplier quote IDs on one owned RFQ. Quote waiting explains the explicit baseline, deadlines, cancellation, exact original records and error handling. Native connector 0.1.6 includes CLI 0.6.14/MCP 0.2.14 and can upgrade recognized original Claude Code/Codex projects after account connection. Custom configurations remain a local review case.
Observe saved actions
Section titled “Observe saved actions”Wait for an action with hosted CLI 0.6.14/MCP 0.2.14. The bounded read follows an existing running action and returns original evidence at draft, approval, coordination or failure states. It never recreates work, retries a write or grants approval. Native connector 0.1.6 includes the same tools. Reconnect or review a project upgrade when using an older installation.
Recovering a lost action ID
Section titled “Recovering a lost action ID”Recover a lost action ID is available in hosted CLI0.6.14/MCP0.2.14. It uses the original project, kind and idempotency key, and reads only recent saved work. A missing match does not establish absence or authorize a replacement. Native connector 0.1.6 includes the same CLI 0.6.14/MCP 0.2.14 tools; an older project may need a reviewed upgrade.