MewCP vs Arcade: Comparing Actions Runtimes for AI Agents
by Rohit Gite, Co-Founder @MewCP6 min read
FAQ
FAQ
Yes. MewCP covers the same four layers as Arcade: an MCP gateway, a tool catalog, managed auth and governance. It offers 280+ providers, keeps the model's tool list at four, enforces action-level permissions per agent, never stores payloads in its logs, and runs on-prem for Enterprise customers. Arcade is ahead today on OIDC identity provider sign-in, custom policy hooks, a SOC 2 claim and air-gapped deployment.
An actions runtime is the layer between an agent and the systems it acts in. It resolves which user the agent is acting for, injects that user's credentials, checks the call against policy, runs the tool and logs the result, so the agent never holds credentials or calls APIs directly. It's sometimes called an MCP runtime when agents connect over MCP. Arcade and MewCP are both actions runtimes.
On 24 October 2026, Arcade's Free plan includes 2,000 tool calls and 2,000 auth events a month. The Team plan is $25 a month plus $0.01 per tool call and $0.10 per auth event. Enterprise uses annual bundles with discounted rates. MewCP's self-serve plans are $9 a month for 100,000 actions and $24 a month for 300,000, with custom Enterprise pricing.
Yes, both. MewCP offers on-prem deployment on its Enterprise plan, alongside its managed cloud. Arcade offers VPC, on-prem, hybrid and fully air-gapped deployment on its Enterprise plan.
MewCP is SOC ready. Arcade states it's SOC 2 compliant. If your security review needs a specific attestation report, ask each vendor for it during evaluation.
You can build your own multi-tenant MCP servers today with MewCP's open-source `fastmcp-credentials` package, the same credential layer every server in MewCP's fleet uses. It keeps credentials out of the model and isolates them per request. Adding your own servers to the MewCP gateway and catalog is coming soon.
Support for SOC tools, including SIEM, is coming soon. Today MewCP logs each call's tool, time, user, result, latency and failure reason in its dashboard, without storing payloads. Arcade's audit logs can be pulled into a SIEM through its REST API today.
Yes. With MewCP your backend sends the user's ID in the `x-mewcp-enduser-id` header, and calls run on the accounts that user connected. With Arcade, users sign in through Arcade or your identity provider, or your client sends an `Arcade-User-ID` header, and the agent acts with that user's delegated access.
With MewCP, add approval in your agent code, for example the Vercel AI SDK's `toolApproval` policy, and keep risky actions out of the toolset entirely if the agent shouldn't run them. Arcade's pre-tool-call hooks can block or hold sensitive actions at the runtime. Either way, keep destructive actions behind an explicit yes.
Arcade has gone deepest on identity federation and policy hooks. MewCP has gone wide on the catalog and strict on control, with a fixed four-tool context, per-agent action permissions, payload-free logs, on-prem deployment and an open-source credential layer. If OIDC sign-in, runtime hooks or air-gapped deployment is a hard requirement today, Arcade fits. Otherwise, run both against your own apps, policies and traffic. You can try the [MewCP runtime](https://mewcp.com/) on the free plan, or talk to MewCP about an Enterprise deployment.
Keep reading
Build an agent that reads live Polymarket odds, checks a wallet's positions and trade history, and places limit orders only after explicit approval and a hard risk cap, through the Polymarket MCP server on MewCP.
A working walkthrough of an AI agent that finds the latest commit on GitHub, deploys it to Vercel through MewCP, polls until the build finishes, and explains failures from the logs.
A layer-by-layer comparison of MewCP and Arcade.dev, two enterprise actions runtimes that give AI agents an MCP gateway, a tool catalog, managed auth and governance, with deployment, compliance, pricing and clear advice on when each one fits.
Arcade and MewCP are the same kind of product for the same buyer. Both are actions runtimes for enterprise AI agents: an MCP gateway the agent connects to, a tool catalog of maintained integrations, managed auth that keeps credentials away from the model, and governance over what each agent may do. The differences are in the choices each makes inside those four layers.
We build MewCP, so keep that in mind. Arcade facts come from arcade.dev and docs.arcade.dev, checked on 24 October 2026. Where Arcade is ahead today, we say so.
| Choose MewCP for | Choose Arcade for |
|---|---|
| 280+ providers behind one gateway | Sign-in through your own OIDC identity provider |
| A fixed four-tool context, however many apps you connect | Pre- and post-call policy hooks |
| Action-level permissions per agent | A SOC 2 compliance claim |
| Logs that never store payloads | VPC, hybrid or air-gapped deployment |
| On-prem deployment and an open-source credential layer | Pulling audit logs into a SIEM today |
Layer by layer:
| MewCP | Arcade | |
|---|---|---|
| Catalog | 280+ providers, 7,100+ tools | 8,000+ tools (docs: 7,500+ across 81 MCP servers) |
| Model context | 4 fixed tools, about 540 tokens | Selected tools, or Tool Recommendation |
| Auth | Managed OAuth, AES-256-GCM vault, your own OAuth client | OIDC sign-in, delegated user scopes |
| Governance | Action-level permissions per agent | Policy hooks, versioned registry |
| Logs | Payloads never stored | OpenTelemetry audit logs, REST API |
| Custom tools | Open-source fastmcp-credentials. Own servers on the gateway coming soon | Open-source framework, or bring your own MCP server |
| SOC and SIEM | Coming soon | SIEM pull through the audit log API |
| Deployment | Cloud, on-prem on Enterprise | Cloud, VPC, on-prem, air-gapped on Enterprise |
| Compliance | SOC ready | SOC 2 |
| Self-serve price | Free 20K, $9 for 100K, $24 for 300K actions | Free 2K, then $25/mo + $0.01 per call + $0.10 per auth event |
MewCP. Your agent connects to one URL: https://gateway.mewcp.com/personal/mcp for your own accounts, or https://gateway.mewcp.com/<toolset-id>/mcp for an agent you ship to users or employees. Requests carry the project key and, when the agent acts for many people, the user's ID:
Authorization: Bearer <MEWCP_API_KEY>
x-mewcp-enduser-id: <your user's ID>On every call the runtime resolves the user, picks their connected account, injects the decrypted credential, checks the toolset's permissions, runs the call and logs it. If the user hasn't connected the app, the agent gets a connect link to pass on. The model only ever sees four tools: search, get_schema, list_accounts and call_tool.
Arcade. MCP Gateways at https://api.arcade.dev/mcp/{your-gateway-slug} federate the tools you select, including remote MCP servers Arcade doesn't host. Users sign in through Arcade Auth, your own OIDC provider (User Source, recommended for production), or Authorization plus Arcade-User-ID headers. Authorization is delegated, and pre- and post-call webhooks can block actions or redact PII.
MewCP fixes the surface at four tools, about 540 tokens, and search only returns tools in that agent's toolset. The same query returns the same candidates, which makes behavior easy to test and audit. Arcade's Tool Recommendation uses retrieval and ranking models to return tools that fit a plain-language task, and Arcade reports up to 91% fewer tokens per session on its own workloads.
By each vendor's published numbers, MewCP covers more distinct apps (280+ providers) and Arcade counts slightly more tools (8,000+). Check that your specific apps are in each.
For your own tools, Arcade has an open-source framework in Python or JavaScript. MewCP publishes fastmcp-credentials, the Apache-2.0 credential middleware every server in MewCP's fleet runs on. It lets one FastMCP server serve many tenants: credentials are injected server-side, scoped to each request, and read by tool code instead of passed as parameters, so the model never sees them. Adding your own servers to the MewCP gateway is coming soon, as is support for SOC tools, including SIEM.
Example: a multi-tenant tool with fastmcp-credentials.
import requests
from fastmcp import FastMCP
from fastmcp_credentials import CredentialMiddleware, HeaderCredentialBackend, get_credentials
backend = HeaderCredentialBackend()
mcp = FastMCP("My Service", middleware=[CredentialMiddleware(backend, "oauth")])
@mcp.tool()
def call_api(resource_id: str) -> dict
The gateway injects each user's credentials as headers. For local development, EnvCredentialBackend reads them from environment variables with no change to the tool. Requires Python 3.11+ and FastMCP 4.
Both handle OAuth at the gateway, refresh tokens and keep credentials in a vault. MewCP adds your own OAuth client per provider, a hosted form for API-key apps and AES-256-GCM encryption, and takes user identity from your application. Arcade can take identity straight from your corporate OIDC provider, which helps when agents act for thousands of employees managed in Okta or similar.
MewCP puts permissions on the toolset: switch apps and single actions on or off per agent in the dashboard, enforced at the gateway. Its logs record tool, time, user, result, latency and failure reason, and never store payloads. Arcade adds programmable hooks for custom policy, versions tools in a registry, and exposes OpenTelemetry audit logs through a REST API that a SIEM can pull from.
MewCP runs in its cloud and on-prem for Enterprise customers. Arcade runs on Arcade Cloud and, on Enterprise, in your VPC, on-prem, hybrid or air-gapped. MewCP is SOC ready. Arcade states it's SOC 2 compliant.
The two price differently. MewCP sells flat bundles of actions, with custom Enterprise terms. Arcade charges a platform fee plus metered tool calls and auth events, with custom Enterprise terms.
| Plan | MewCP | Arcade |
|---|---|---|
| Free | 20,000 actions/month | 2,000 tool calls and 2,000 auth events/month |
| Entry paid | Plus, $9/month, 100,000 actions | Team, $25/month platform fee |
| Usage | Pro, $24/month, 300,000 actions | $0.01 per tool call, $0.10 per auth event |
| Enterprise | Custom, with SLA guarantees and on-prem deployment | Custom annual bundles with discounted rates |
What a month of tool executions costs on the self-serve plans at list price, before Arcade's auth events:
| Tool calls per month | MewCP | Arcade Team |
|---|---|---|
| 2,000 | $0 | $0 |
| 20,000 | $0 | about $225 |
| 100,000 | $9 | about $1,025 |
| 300,000 | $24 | about $3,025 |
Enterprise contracts on both sides are negotiated, and Arcade's bundles bring per-call rates down at volume. These figures count tool executions only. Discovery calls, auth events and add-ons follow each vendor's own rules, so check your traffic against MewCP's pricing and Arcade's pricing.