Skip to main content
Glama

Forge by FoundryNet — Industrial Machine Intelligence

The cross-manufacturer industrial MCP server. Talk to any CNC, robot, or industrial machine in natural language — machine identity, telemetry normalization across 52 vendor packs in 20 verticals, plain-English automation, and tamper-evident work records.

Hosted MCP over Streamable HTTP. 36 tools wrap the Forge v1 API: provision a stable machine identity, normalize raw OEM telemetry into a canonical schema, query operational history, parse and activate plain-English automations, predict failures (TimesFM), score fleet health, and record every state-changing action as a verifiable, tamper-evident work record.

What it does

It normalizes raw OEM telemetry from 18 manufacturer families into one canonical vocabulary with thousands of confirmed field mappings — so an agent writes against one set of field names whether the machine is a Fanuc CNC, a KUKA arm, or a Universal Robots cobot. On top of that it turns plain-English instructions into structured automations (review then activate), and records every state-changing action as a tamper-evident work record.

Related MCP server: Industrial MCP Agent Platform

Architecture

Pure HTTP proxy. Every tool is a thin wrapper around https://forge.foundrynet.io/v1/*. No state, no shared imports with forge-prod — separate Railway service, separate dependencies (fastmcp + httpx).

Claude Desktop / agent
        │ Streamable HTTP (/mcp) — or legacy SSE (/sse)
        │ Authorization: Bearer <the CALLER's Forge API key>
        ▼
  foundrynet-mcp on Railway
        │ gate: validate key → tier → cap → bind caller
        │ HTTPS, Authorization: Bearer <the SAME caller's key>
        │        X-Forge-Caller-Key-Id: <their forge_api_keys.id>
        ▼
  forge.foundrynet.io/v1/*  — scopes every row to that caller

Tenancy. The upstream call is made as the CALLER, not as this server. It used to be made with the server's own FOUNDRYNET_API_KEY for every caller, with no caller named: forge-prod scopes rows to the authenticated key's owner, so every MCP tenant read and wrote inside one shared Forge account — and fleet_oee, shift_report, list_agents and prediction_accuracy take no machine argument, so they returned that whole account. The upstream hop now REFUSES to run when it cannot name the caller: there is no fallback, because the only fallback is the shared key. OAuth access tokens carry a key id but not a key, so they cannot be presented upstream and are refused on tool calls with tenant_context_unavailable.

POST /a2a/tasks and GET /a2a/agents go through the same gate function (gating.enforce_call) as tools/call, by skill→tool mapping. They previously authenticated the caller and then dispatched with no tier check, no cap and no payment gate, which made the A2A route a free pass to the paid inference tools.

Tests and the release gate

python3 -m venv .venv && .venv/bin/pip install -r requirements.txt pytest pytest-asyncio
.venv/bin/python -m pytest tests/ -q        # 229 checks
scripts/release_gate.sh                     # tree, deps, import, suite, version

The suite drives the real ASGI app over real Streamable HTTP through the real gate; only Supabase and the upstream hop are substituted, and the upstream substitute RECORDS requests, because what identity this server presents to forge-prod is the thing most worth asserting. tests/test_rest_mcp_parity.py measures free-tier parity against forge-prod's own _endpoint_to_billing_meter, read out of its source at test time rather than copied.

CI runs the whole tree on every push and pull request (.github/workflows/release-gate.yml, pinned to ubuntu-24.04): measured 223 passed, 6 skipped, 229 collected in 2m04s. The 6 skips are the checks that read forge-prod's api.py off the local disk — 5 parity plus the cost-ledger single-writer check — and they skip loudly, with their reason printed, because forge-prod is a separate private repo and is not on the runner. A visible skip beats a green tick over a check that was quietly dropped.

Deploying

cd ~/forge-mcp && git checkout main && git pull --ff-only && ./deploy.sh

Never railway up by hand. deploy.sh does seven things in order, and five of them exist because of a specific way a deploy has gone wrong:

  1. refuses a dirty tree or a branch that is not main — railway up uploads the WORKING TREE, so with uncommitted changes no commit describes what ships;

  2. runs scripts/release_gate.sh and refuses if it is red. There is no override flag, deliberately;

  3. takes a deploy lock (~/.forge-mcp-deploy.lock, atomic mkdir) so two concurrent deploys cannot race on the upload and on the verification;

  4. refuses if railway up would archive the wrong directory. It archives the CLOSEST LINKED project directory from ~/.railway/config.json, walking UP — not $PWD. This directory once inherited $HOME's link, which names a DIFFERENT SERVICE;

  5. runs stamp_build.py, which writes build_info.json (commit, dirty state, content_hash over the uploaded file set, build time);

  6. railway up --service 162115b7-… — the id is explicit, because a bare railway up resolves the service from that same ancestor link;

  7. polls https://mcp.foundrynet.io/health until build.commit_short AND build.content_hash both equal the stamp it just wrote, then asserts the status. Railway reporting SUCCESS only means the upload and build succeeded — a crash-looping new container leaves the OLD one serving and the deploy looks fine. content_hash is checked as well as the commit because railway up uploads a tree: the same commit with a different tree is a different deployment.

/health carries a build block using the same key names as forge-prod's, so one verifier reads both services. When build_info.json is absent it says so in words — build_info.json absent — deploy did not run stamp_build.py, with commit, content_hash and built_at present and null — because a null that looks like a missing field is indistinguishable from a stamped build.

Tools (36)

Identity & data: identify_machine, normalize_telemetry, query_machine_history, get_coverage, correct_mapping. Automation: create_automation, activate_automation, list_automations, disable_automation, delete_automation, restore_automation, query_webhook_history. Prediction (TimesFM): predict, predict_breach, remaining_life, predict_batch, fleet_health, detect_anomalies, machine_intelligence, prediction_accuracy. Operations: calculate_oee, fleet_oee, energy_consumption, shift_report, diagnose_machine, health_index. Guardrails: check_guardrail, list_guardrails. Agents on your kernel: get_agent_card, list_agents. Alerts: set_alert, list_alerts, delete_alert, alert_history. Work records: verify_record. Demo: fire_sandbox (the full watch → fire → settle loop, no card).

get_agent_card and list_agents are scoped to the agents connected to your kernel instance (their capabilities, trust scores, and machine access) and are included in your subscription — there is no per-call charge. They are operational tools for coordinating agents on your own equipment, not cross-platform discovery. For cross-platform agent trust across any framework, see Assay (assay.foundrynet.io). verify_record creates a tamper-evident, hash-verified record of completed work.

Free tier exposes the read-only tools; metered pay-per-use unlocks the premium prediction and diagnostics tools (see https://forge.foundrynet.io/pricing?utm_source=github&utm_medium=readme&utm_campaign=forge-mcp-readme).

Connect (Claude Desktop, Cursor, any MCP client)

claude mcp add --transport http foundrynet-forge \
  https://mcp.foundrynet.io/mcp \
  --header "Authorization: Bearer fnet_YOUR_KEY"

Or via claude_desktop_config.json with the mcp-remote bridge:

{
  "mcpServers": {
    "foundrynet-forge": {
      "command": "npx",
      "args": ["-y", "mcp-remote",
               "https://mcp.foundrynet.io/mcp",
               "--header", "Authorization:Bearer ${FNET_KEY}"],
      "env": { "FNET_KEY": "fnet_…  (get a free key at foundrynet.io/signup)" }
    }
  }
}

Legacy SSE (--transport sse against https://mcp.foundrynet.io/sse) remains supported for existing configs, but new integrations should use the Streamable HTTP /mcp endpoint above.

Get a free fnet_ key at https://foundrynet.io/signup?utm_source=github&utm_medium=readme&utm_campaign=forge-mcp-readme (50 normalize calls, no card).

Required environment (server-side)

Var

Required

Default

FOUNDRYNET_API_KEY

No longer used on the tool path

Kept for the boot-time config report only; tool calls run as the caller's own key

SUPABASE_URL / SUPABASE_SERVICE_KEY

Yes for gating

unset = gating disabled, fail-open

ENABLE_OAUTH_EXCHANGE

No

default false: /oauth/token and its metadata answer 501 and JWTs are not accepted (G9). true re-enables the client-credentials flow

MCP_JWT_SECRET

No

with the exchange enabled: unset = issuance answers 503; key auth unaffected

FORGE_BASE_URL

No

https://forge.foundrynet.io

PORT

No

8080 (Railway sets this automatically)

REQUEST_TIMEOUT

No

120 (seconds)

Files

  • mcp_server.py — the server (36 tools + /health + /.well-known/mcp routes)

  • gating.py — per-client tier gating (Free vs Pro tool/quota enforcement)

  • server.json — MCP registry metadata (name, description, keywords, remote endpoint)

  • smithery.yaml — Smithery listing metadata

  • requirements.txt — fastmcp>=3.4.5,<4, httpx>=0.27, PyJWT>=2.14.0, supabase, stripe

  • tests/ — the behavioural suite (auth, tenancy, parity, idempotency, cache scope)

  • scripts/release_gate.sh — must be green on the exact commit before a deploy

  • deploy.sh — the only path to production; gated, locked, stamped and verified

  • stamp_build.py — writes build_info.json; what /health's build block reports

  • Procfile — Railway start command (stamps, then python mcp_server.py)

  • .github/workflows/release-gate.yml — CI; the whole suite on push and PR

License

Proprietary (commercial). © Foundry Labs LLC. Contact: foundrynet@proton.me

Available Tools

14 tools
activate_automationAInspect

Activate a parsed automation trigger on a machine. Call this AFTER create_automation returns a parsed_trigger and the user explicitly confirms they want to arm it.

Creates a live trigger that monitors the machine's normalized telemetry and fires the listed actions when the condition matches. Each action references a registered tool by tool_id; on fire, the tool's webhook is POSTed with {{variable}} interpolation against the canonical data context (mint_id, oem, model, serial, site, field, value, threshold, plus every canonical field on the matched record).

Inputs: machine_id mint_id ("MINT-…") or internal_id; resolved to canonical mint_id name short human label, ≤ 80 chars (e.g. "high spindle load") condition simple {field, op, value|threshold} OR compound {all: [...]} ops: >, <, >=, <=, ==, != actions list of {tool_id, payload_overrides?, headers_overrides?} enabled defaults to true; pass false to create the trigger paused

Returns the persisted trigger row including id (use it later to pause/edit/delete via the Forge API). Once active, the trigger fires on every subsequent normalize_telemetry call where the condition matches — no further activation needed.

USE WHEN: the user has reviewed the parsed_trigger from create_automation and said something like "yes, activate it" / "go ahead" / "arm it." Never call this tool without explicit confirmation — it changes machine behavior in a way the user can feel (real Slack messages, real ERP work orders).

Inputs: name (required), actions (required), condition (required), machine_id (required), +1 more. Returns: object

correct_mappingAInspect

Teach Forge the RIGHT canonical field for a source column that normalize_telemetry mapped wrong (or abstained on). Each correction is recorded as a corpus-improvement signal the retrainer uses to fix the mapping for everyone — so every agent interaction makes normalization better.

USE WHEN: you or the user can see normalize_telemetry returned the wrong canonical for a field (e.g. it mapped an oil-pressure column to a tire- pressure field), or it abstained on a field whose meaning you know.

  • source_field: the raw column name exactly as it appeared in your data.

  • confirmed_canonical: the canonical field it SHOULD map to.

  • original_canonical: what normalize_telemetry actually returned (pass the canonical from that field's entry; use "abstained" if it abstained).

  • oem: the OEM you passed to normalize_telemetry (improves aggregation).

  • mapping_id: optional — the mapping_id from the normalize_telemetry field entry. If omitted it is derived deterministically from (source_field, oem).

Returns {ok, feedback_id, action:"correct"}. Corrections feed an offline retrain (they don't hot-patch the live corpus), so noisy feedback can't poison other users' mappings.

Inputs: source_field (required), original_canonical (required), confirmed_canonical (required), oem, +3 more. Returns: object

create_automationAInspect

Set up automated monitoring + actions for an industrial machine using natural language. Connect machine telemetry to any business system — ERP, CMMS, MES, Slack, Teams, email, Zapier, n8n — via webhooks already registered as tools on the Forge service.

Examples of instruction: "Alert maintenance Slack when spindle load exceeds 90 percent." "Create a Fiix work order when coolant temperature stays above 35°C for five minutes." "Notify the supervisor when part_count hits 500." "When the maintenance_type changes to CORRECTIVE, post to the ops channel."

Returns a parsed_trigger JSON for HUMAN review — DOES NOT auto-activate. The caller (you, with user confirmation) must explicitly POST the parsed_trigger to /v1/triggers on the Forge API to actually create it. The response includes confirmation_required: true and may include notes if the parser had to make a fuzzy match (e.g. resolved an ambiguous field name to its closest canonical match).

USE WHEN: a user wants to set up monitoring, alerts, or automations for machine state transitions. Always show the parsed_trigger to the user verbatim and ask "Confirm to activate?" before they activate it.

Inputs: mint_id (required), instruction (required). Returns: object

delete_automationAInspect

Soft-deletes the trigger (recoverable for 30 days via restore_automation). The trigger immediately stops evaluating against /v1/normalize calls and is hidden from list_automations, but the row persists with deleted_at set so an accidental delete can be undone. Use restore_automation to undo. For permanent deletion, the API supports ?permanent=true.

Past forge_trigger_executions rows for this trigger remain in either case (audit trail).

USE WHEN: the user wants to remove an automation they no longer need — "delete the coolant alert," "remove that trigger." Safer than hard delete because misclicks are recoverable; tell the user about restore_automation if they later change their mind.

Inputs: trigger_id (required). Returns: object

disable_automationAInspect

Pause an automation trigger without deleting it. The trigger stops evaluating against incoming /v1/normalize calls but its configuration (condition, actions, history) is preserved. Re-enable later by PATCHing /v1/triggers/{id} with {"enabled": true} (or by asking the user to confirm and creating a follow-up tool for resume).

USE WHEN: the user wants to TEMPORARILY stop an automation — e.g. "pause the high-spindle alert during planned maintenance," "stop that alarm for now, I'll re-enable it tomorrow." Distinct from delete_automation, which is permanent.

Inputs: trigger_id (required). Returns: object

fire_sandboxAInspect

Demo the full Forge watch→fire→settle loop against a built-in sandbox endpoint. Free tier; no machine onboarding required.

The MCP server POSTs {message, condition: condition_text, ts} to its own /sandbox/echo route — a real HTTP round-trip with a real response body — then hashes the response and anchors it on Solana mainnet via the MINT relay. Returns the echo body, the tx_signature, and a Solscan verify_url.

USE WHEN: a developer is evaluating Forge and wants to feel the full loop (a webhook actually fires, a real Solana tx actually settles, the Solscan link actually verifies) without onboarding any machines or paying for the Pro tier. 10 fires lifetime per fnet_ key.

Inputs: condition_text (required), message. Returns: object

get_coverageAInspect

Ask Forge what it can normalize BEFORE you try: the recognized OEM verticals (CNC / robot / vehicle / AMR), the canonical-field families, and the field list per family. Optionally pass an oem to see which vertical it resolves to and whether the cross-vertical gate will engage.

USE WHEN: starting a new integration, or deciding whether to call normalize_telemetry — confirm the machine's OEM and your fields are in coverage. Unknown OEMs still normalize (the gate just disables itself), so absence here is a soft signal, not a hard block.

Inputs: oem. Returns: object

identify_machineAInspect

Provision or retrieve a persistent on-chain identity (mint_id) for any industrial machine. Works for CNC machines, industrial robots, PLCs, additive manufacturing cells, injection molders, presses, turbines, pumps, compressors, conveyors — any equipment from any OEM: Fanuc, Siemens, Haas, DMG Mori, Mazak, Okuma, Hurco, Doosan, Makino, ABB, KUKA, Universal Robots, Yaskawa, Stäubli, FANUC Robotics, Komatsu, Caterpillar, John Deere, Trumpf, Bystronic, Amada, EMAG, Bosch Rexroth, Beckhoff, Rockwell Allen-Bradley.

Returns the mint_id (universal handle, format "MINT-xxxxxx") plus its Solana wallet_address. Idempotent — calling again with the same (oem, model, serial) returns the same mint_id with created: false.

USE WHEN: a user references a specific machine by OEM/model/serial and you need a stable handle to attach normalized data, automations, or on-chain settlements to. Always call this first when a new machine is introduced to the conversation, before normalize_telemetry or create_automation.

Inputs: oem (required), model (required), serial (required), site, +1 more. Returns: object

list_automationsAInspect

List all active automations / triggers configured for one machine.

Returns each trigger with: id, name, condition (field/op/value or compound all), actions (each resolved to its tool name + url + method), enabled state, fire_count, last_fired_at, last_error.

USE WHEN: the user asks "what automations do I have on this machine" / "show me my triggers" / "what alerts am I getting" / "what's monitoring this machine right now". Always pass the machine's mint_id (or internal_id — both resolve).

Inputs: machine_id (required). Returns: object

normalize_telemetryAInspect

Translate raw machine telemetry from any OEM's proprietary format into universal canonical FCS (FoundryNet Canonical Schema) data. Maps vendor-specific column names like "Spindle_Speed", "servo_load_x", "CoolantTemp", "FeedRateOverride" into standard fields like spindle_speed_rpm, axes.x_load_pct, sensor_readings.coolant_temp, feed_override_pct.

Accepts a data dict of {raw_field: value}. If machine_id (mint_id or internal_id) is omitted but oem+model+serial are provided, silently auto-provisions the machine identity (same effect as calling identify_machine first).

Each call:

  • Returns canonical_data + a per-field mapping_id (use mapping_id with /v1/feedback/{mapping_id}/correct if a mapping is wrong)

  • Writes a row to forge_normalized_history (visible via query_machine_history)

  • Evaluates active triggers; the response includes a triggers_fired array if any condition matched. The actual webhooks fire async, so the array tells you what was triggered without blocking on remote latency.

USE WHEN: you have raw machine data — a CSV row, a sensor reading, an MES export, an alarm log line — and need to either (a) understand it semantically using canonical field names, (b) feed an automation that watches canonical fields, or (c) build up history for the machine.

Inputs: data (required), oem, site, model, +2 more. Returns: object

query_machine_historyAInspect

Retrieve operational history for an identified machine. Each row is one /v1/normalize call's canonical output (FCS field → value).

Query options: from_dt, to_dt ISO-8601 timestamps to bound the time range fields comma-separated FCS field names to project; omit for full canonical_data limit max rows (1–1000, default 100) summary true → returns aggregate stats only (row_count, time range, avg coverage_pct, fields_covered set) without the raw rows. Always cheap.

USE WHEN: a user asks how a machine has been running, wants utilization or throughput or health trends, looks for patterns in alarms or operational state, compares periods ("how was today vs yesterday"), or wants to know what data is even available for a machine. Prefer summary=true first to orient on volume + which fields are present, then drill in with field projection on a smaller time window.

Inputs: mint_id (required), limit, to_dt, fields, +2 more. Returns: object

query_webhook_historyAInspect

Show webhook delivery history for a trigger — HTTP status codes, response times, retry counts, errors. Use to verify webhooks are actually delivering.

Returns up to limit most-recent execution rows (default 10, max 200), each with: fired_at, http_status, attempt_count, response_time_ms, error (if any), tool_name, target_url, and the on-chain settlement tx (settled_tx) once the row has been Merkle-rooted via batch settle.

USE WHEN: a user asks "did the alert actually go out?" / "why didn't Slack get pinged?" / "is the trigger working?" / "show me the last few fires." Soft-deleted triggers can still be queried — useful for forensic audits after a misclick + restore.

Inputs: trigger_id (required), limit. Returns: object

restore_automationAInspect

Restore a previously soft-deleted automation trigger within its 30-day recovery window. Re-enables the trigger so it evaluates against incoming /v1/normalize calls again.

Returns the restored trigger row plus restored: true and the restored_at timestamp. 410 (Gone) if the trigger was deleted more than 30 days ago and is past the restorable window. 409 if the trigger isn't actually deleted.

USE WHEN: a user accidentally deleted a trigger and wants it back. Also useful as the "undo" half of a "delete then change my mind" flow — pair with disable_automation when the user wants to pause rather than delete in the first place.

Inputs: trigger_id (required). Returns: object

verify_on_chainAInspect

Anchor data on Solana mainnet via the MINT relay for cryptographic proof of work. Two modes:

BATCH MODE (batch=true, requires mint_id): Collects every unsettled event for that machine — normalize calls, trigger fires, webhook executions — since the last batch. Computes a Merkle root of their event hashes and anchors that single root on Solana. ONE transaction proves dozens to thousands of events. Returns: merkle_root, event_count, event_types breakdown, tx_signature, verify_url (Solscan link). Cost-efficient — call this once an hour or once a shift per machine, not per event.

SINGLE-PAYLOAD MODE (batch=false, requires payload): Hashes an arbitrary JSON payload deterministically (sorted keys, no whitespace) and anchors the hash. Returns: payload_hash, tx_signature, verify_url. Use for one-off proofs — inspection records, completed work orders, signed reports — where you want a permanent independent timestamp.

USE WHEN: a user wants tamper-proof evidence — settlement of a completed work batch, proof a maintenance window happened, anchoring a quality report, rolling up a day's machine activity into a single verifiable hash. ALWAYS include the verify_url (a Solscan link) in your reply so the user can independently verify on-chain.

Inputs: batch, mint_id, payload. Returns: object

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.5/5.0

    Scored across 14 tools

    Disambiguation5/5

    Each tool has a clearly distinct purpose covering machine identity, telemetry normalization, mapping corrections, coverage checks, history queries, automation lifecycle (create, activate, list, disable, delete, restore), webhook history, sandbox demo, and on-chain verification. No two tools overlap in their core functionality.

    Naming Consistency5/5

    All tools follow a consistent verb_noun pattern in snake_case (e.g., identify_machine, normalize_telemetry, create_automation). The naming is predictable and intuitive, making it easy for an agent to infer tool purpose from the name.

    Tool Count5/5

    14 tools is well-scoped for an industrial IoT platform. Each tool addresses a specific need without unnecessary redundancy. The count supports a comprehensive workflow without overwhelming the user.

    Completeness4/5

    The tool set covers the core lifecycle: machine identity, telemetry normalization with correction feedback, history queries, automation CRUD plus activation/disable/delete/restore, webhook monitoring, sandbox demo, and on-chain verification. A minor gap is the lack of a dedicated re-enable tool for disabled automations, but the disable description notes it can be done via PATCH, so agents can adapt.

    Maintenance

    ActivityActive
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      Not graded
      quality
      D
      maintenance
      An AI agent system that monitors manufacturing equipment health using the Model Context Protocol (MCP). Enables answering natural language questions about equipment status, maintenance, and anomalies through MCP tools.
      -
    • F
      license
      Not graded
      quality
      D
      maintenance
      Enables engineers and plant managers to interact with manufacturing systems using natural language, providing machine health analysis, KPI dashboards, predictive maintenance, and automated workflow execution via MCP.
      -