The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Agent Utilities MCP listing page.
Connect a local stdio MCP client to 36 paid utility APIs for HTML extraction, JSON validation, agent configuration checks and product data. This MIT-licensed adapter runs locally; the hosted API is a paid service.
Tool catalog · Installation guide · Pricing · Policies and support
VS Code users: open the setup link and private-key configuration. Detailed instructions are included in this source repository.
Clone this repository and preview the supplied cart with Node.js 22+:
Preview validates the input and prints the three-step plan offline, with no API key, network request or charge. It does not execute the tools. The included expected output is a local fixture: German price strings become decimal amounts, guarded patches update the supplied cart snapshot, and cost reconciliation returns a known total of €45.92. Tax is unknown, so the final total remains null.
To run those three operations against the hosted service, privately set AGENT_UTILITIES_API_KEY in your process environment and explicitly choose:
Execution spends existing service credits. The recipe enforces per-call price ceilings totaling at most $0.0013 (0.13 cents) across its three logical calls. It stops if a discovered price is too high or a later price exceeds the authorized ceiling. Completed steps stay charged if a later step fails. Each step returns its debit receipt and recovery information; do not rerun the entire recipe after an uncertain response. Read the recipe instructions before execution. This example does not buy credits or place a merchant order.
The recipe is a separate standalone download, not part of the immutable v0.3.0 MCPB bundle. You can also get it from the workflow page. The generic MCP adapter's spending behavior remains as described below.
No account or key is needed to read contracts, examples, prices and guides:
These public GET endpoints support cross-origin browser reads. Tool execution remains paid. Custom HTTP clients can send X-Max-Credit-Micro-Usd to cap a new prepaid credit debit; read the quickstart for retry semantics. Adapter 0.3.0 caps each new call at its startup catalog price and requires server-advertised support.
The Python HTTP client uses only Python 3.10+ standard libraries. See examples/python for offline preview, free discovery and explicit paid calls with per-call ceilings and stable retry identities. No Node or MCP process is required for this route. It is separate from the MCPB bundle.
Also listed on Smithery as a local Node.js MCPB bundle. Its download matches the v0.3.0 GitHub artifact. Connect the adapter for the complete current tool schemas and prices; directory metadata is a discovery summary.
For clients that support MCP bundles, download agent-utilities-mcp-0.3.0.mcpb and its checksum from the v0.3.0 release. Verify the checksum before importing the bundle through your client's extension settings. It contains the adapter, manifest and license notices. The bundle is unsigned; the release checksum establishes file consistency, not an independent signature. Bundle schema and standalone execution are tested; installation in each desktop client is not verified.
The bundle's optional Agent Utilities API key setting is marked sensitive. Leave it empty for free discovery, or enter only your API key to enable paid calls. Node.js 22+ is required; use a compatible runtime provided by your client or installed locally. If your client cannot import MCPB files, use the manual setup below.
Requires Node.js 22 or newer. Download agent-utilities-mcp.mjs and its .sha256 file from the installation guide, then verify the file from its folder:
Alternatively, clone this repository and use its bundled agent-utilities-mcp.mjs directly. No npm install is needed to run the bundle. Source is in src/; npm ci && npm run build rebuilds it. The public bundle is tested as a standalone process outside the application's dependency tree.
For a client that uses the common mcpServers configuration format:
Replace the file path. Configure the API key privately using the client's secret settings where supported; never commit a real key or paste it into an agent prompt. Use the absolute path to Node if the client cannot find it. Remote-only MCP clients cannot launch this stdio adapter.
Omit the key to discover tools without spending credits. Obtain an API key and separately saved recovery key in the workspace. Only give the API key to the adapter. Live mode sells $5 USD prepaid service credits. Check the displayed mode before buying.
Try a useful task such as: “Use commerce_gtin_validate to check barcode 036000291452.” One successful call costs $0.0003. Current tool prices range from $0.0003 to $0.002; these are experimental prices, not fixed forever. Every new operation spends credits. Version 0.3.0 caps each new debit at the price discovered at startup. There is no total session budget; repeated new calls can spend the available balance. A higher server price returns HTTP 412 instead of increasing the ceiling. Restart only after reviewing refreshed prices. Review prices and approve spending in your client.
Results and uncertain outcomes include a requestId. Call agent_utilities_retry with that ID, the original MCP tool name, and the original arguments:
Within ten minutes a completed request returns the saved result without another debit. In version 0.3.0 the recovery helper sends a zero ceiling: a request that never reached the service is rejected without a new reservation. This does not cancel a previous reservation or refund a charge. The automatic transport retry of a new call retains its original nonzero ceiling and can still execute that authorized operation once. Do not generate a new ID or change the input to recover an uncertain result. Expired request identities cannot make another debit. Reusing an MCP request ID within one process also retains its original debit identity.
After a process restart you need the returned request ID and original input for recovery. If the process died before your client received the ID, inspect the balance or contact support before repeating the operation. No durable local input or result history is stored. Each process retains up to 5,000 MCP request identities, then refuses new calls until restarted; finish pending recovery first.
The release allows 1,000 new paid calls per day across the service. Inputs are bounded to 128 KiB including protocol overhead. Network tools are restricted to the hosts shown in the catalog; they are not a general web browser. Security checks are heuristics. Successful results are encrypted for ten-minute retry recovery, with request tombstones retained for one day. See the policies for retention and refund details.
Automated tests verify lost-response recovery against a local Cloudflare credit ledger. They do not establish customer demand or prove a live customer purchase. Directory validation and free tool discovery are not sales.
Adapter code: MIT. Bundled dependencies: see THIRD-PARTY-NOTICES.txt. The license covers the adapter, not free access to the hosted APIs.
The pack planner finds the minimum item subtotal for a required count of interchangeable items using explicit pack limits. Twelve items can cost $15.98 as two six-packs at $7.99, even when a ten-pack at $11.99 has a lower unit price. The operation costs $0.0008 in prepaid credits. Shipping, tax, coupons and product equivalence are outside its optimization.
Call commerce_pack_plan through MCP, or use the Python client with tool ID commerce.pack-plan and an explicit 800-micro-dollar ceiling. Inspect the free JSON contract before executing. The worked workflow explains stock limits and how pack counts map into cart reconciliation. The adapter discovers this tool from the service; new tools do not require replacing the adapter.
Try the Python pack-planning example with its three sample cases. Run python3 pack_plan_example.py from examples/python, or select --scenario limited-stock / --scenario insufficient-stock. The default is an offline display of fixed fixtures; it sends no request and computes no new answer. Adding --execute explicitly authorizes one hosted call capped at $0.0008 using existing credits. A valid infeasible result is billable too. See the Python guide for checksums, input limits and recovery before executing. Each new run creates a new ID; do not rerun to recover an uncertain paid result.
Install 0.3.0 for per-call price ceilings and recovery-only manual retries. Version 0.2.1 does not enforce these adapter protections. Existing 0.2.1 release bytes remain available and unchanged; upgrading requires replacing the installed adapter. The per-call ceiling does not limit the number of new calls or provide a total session budget.
The standard-library Python recipe in examples/python/compare_carts.py compares supplied merchant checkout snapshots using one capped reconciliation per cart. Default preview is offline; execution requires --execute and existing credits. It ranks complete totals, retains ties and declines to name an overall winner when any cart has unknown shipping, tax or fees. The two-cart default ceiling is $0.001. See the comparison guide for sample inputs, checksums, explicit budgets and recovery.