Can an AI agent use Arcade?
Arcade is an MCP gateway for running tools such as calculators and apps on behalf of users, with a REST API and a CLI.
Rank among the 246 products we drove
- Core work done
- tied #1
- 100%
- 8 of 8 core tasks
- Time to get in
- tied #42
- 4minutes
- Getting in on its own
- Steps on its own
- tied #35
- 3/ 5
- Find to Leave, no person
Yes. Our agent signed up for Arcade on its own in a browser with an email and password, got an API key and used its API, MCP server and CLI.
The record
Level confirmed
Not yet
Level from one drive; a second must agree.
- Got in by
- Sign-up in a browser4 minutes
- Got as far as
- Pay: reaches the checkoutA card form for $25 a month
Connects by
MCP, API, CLI
Agent payments
None found
On its own · 322 doc pages read
Docs it can act on: llms.txt, an OpenAPI and an MCP server.
On its own · 4 minutes
MCP sign-in, sign-up in a browser.
On its own · 8 of 8 core tasks
Did: search the tool catalog and fetch tool definitions in OpenAI and Anthropic formats; run a tool for a user and read the run back from the history.
Partly
Reaches the checkout. A card form for $25 a month.
Partly
Nothing to cancel. Its account stays.
- On its own
- Partly
To reach L4: Delete its account, so its keys and login stop working.
See how it ranks in the MCP gateways category.
Own Arcade? 7 things to fix. Get your findings, free
Checked 18 Nov 2026: MCP server asks for a sign-in, sign-up page, llms.txt and OpenAPI unchanged since the check before.
Find
How an agent connects
MCP, API, OpenAPI, CLI and llms.txt found.
MCP server
All testedOfficial server. Site MCP (4 read-only tools, no auth); per-project gateways with OAuth (CIMD/DCR, PKCE), 7-9 tools when driven. Sign‑in needed to see its tools.
10 tools mapped in our drive
API
58 of 60 testedREST at api.arcade.dev, OpenAPI 3.0.0.
60 operations mapped in our drive
OpenAPI
YesA machine-readable description of the API.
CLI
27 of 28 testedarcade-mcp 1.17.3 on PyPI: login, new, deploy, server, secret, evals, and more.
28 commands mapped in our drive
SDK
None foundllms.txt
YesAgent payments
None foundReached the card form and stopped: Growth, $25 a month. It checks out through a Stripe card form. No CAPTCHA at checkout.
Docs
In partCovers getting started, auth, error codes and code examples.
Describes each operation only in part: completeness 50 of 100.
6 of 8 documented ways to do the core work worked.
322 of 384 pages read
Get in
Getting in
Our agent tried 3 of 9 ways in; 3 worked without a person.
The way that worked
Sign-up in a browser4 minutes
- An inboxNeeded
- A browserNeeded
- A personNot needed
- Agent on a serverAPI, MCP, CLINoStops at sign-up: it works only in a browser.
- Assistant in its own browserLike ChatGPT agentPartlySigns up, confirms its email and does all of the core work, then stops: its account stays. It can't delete its account.
- Signed agentKnown to CloudflareYesThe sign-up page opens.
Every way in
| Way in | What happened | Why |
|---|---|---|
| Sign-up in a browser | Worked | |
| MCP sign-in | Worked | |
| Sign-up in a browser | Worked | Part of the work done |
| Runs locally, no account | Not tried | Not part of this drive |
| Continue with GitHub | Not tried | Not part of this drive |
| Continue with Google | Not tried | Not part of this drive |
| Continue with Microsoft | Not tried | Not part of this drive |
| Company single sign-on | Not tried | Needs a person's account |
| A key from a person | Not tried | The same as another way |
Use
What an agent can do with Arcade
All 8 core tasks done, and 15 of 20 others.
Core tasks
- done there
| Task | Result | Through |
|---|---|---|
| Run a tool action on behalf of a user | Done | API: done thereMCP: done there |
| Authorize a user's account for a tool (OAuth consent) | Done | API: done there |
| Browse and search the tool catalog | Done | API: done thereMCP: done there |
| Create a tool gateway that federates selected tools | Done | API: done thereMCP: done there |
| Change or delete a tool gateway | Done | API: done there |
| Build and deploy a custom tool server to Arcade | Done | API: done thereCLI: done there |
| Register a remote tool server in a project | Done | API: done there |
| Add a contextual access hook to enforce policy on tool calls | Done | API: done there |
Surface audit: 105 documented operations: 99 run, 3 blocked by sandbox or plan limits and 3 left out because they spend money or reach real people. Of the 99 run, 90 passed and 9 didn't; 25 behaved differently from the docs. By surface, 58 on the API, 27 in the CLI, 10 on the MCP server and 4 elsewhere, such as sign-in and webhooks.
Holds up
75of 100
- Reliability83
- Handles mistakes67
The mean of the two.
Docs match
68of 100
- Docs accuracy75
- Docs completeness50
Docs accuracy counts 70%, completeness 30%.
Surface tested
88of 100
90 of 102 passed: the 105 documented operations less the 3 that spend money or reach real people
What it can't do, and why
- Pay for the Growth plan Only a card form in the dashboard; no API or agent payment option.
- Delete the account No delete control or endpoint; the privacy policy says to email support.
- Schedule a tool run for later The API refuses it as removed, although the OpenAPI still lists it.
- Call an LLM through Arcade The documented chat endpoint answers 404.
- Deploy more than one hosted server at a time The free plan refuses a second deployment.
- Finish OAuth for prebuilt GitHub and Google tools Consent needs a person's GitHub or Google account.
15 of 20 other tasks done
- Review tool execution historyDone
- Check a tool's unmet authorization requirementsDone
- Get tool definitions formatted for an LLMDone
- Get a user's provider token to call a third-party API directlyDone
- Add a custom OAuth auth providerDone
- Store a secret for toolsDone
- Revoke a user's connected accountDone
- Set user verification for the auth flowDone
- Turn on tool recommendation for a gatewayDone
- Scaffold and run a tool server locallyDone
- Read a deployed server's logsDone
- Answer end users' capability requests in the private registryDone
- Read the audit logDone
- Count tool calls for usageDone
- Create an organization or projectDone
- Configure SSO for the Arcade organizationPaid plan
- Set a member's role (RBAC)Paid plan
- Schedule a tool execution for laterThe API refuses it as removed, although the OpenAPI still lists it.Blocked
- Run evaluations on toolsBlocked
- Call an LLM through ArcadeBlocked
- Connect an identity provider for gateway end usersNot tried
- Self-host Arcade in your own cloud with HelmNot tried
The plans they need
- Build and deploy a custom tool server to Arcadepaid (free plan allows one managed deployment at a time)
The drive used the free plan.
Tips for agent builders
- After sign-up, set up the authenticator-app second factor: the setup key is shown as text, so you can compute TOTP codes yourself.
- Prefer the MCP gateway at https://api.arcade.dev/mcp/arcade: it supports CIMD and PKCE and needs one consent click.
- Use the org-scoped paths /v1/orgs/{org}/projects/{project}/plugins and /hooks: the flat /v1/plugins and /v1/hooks in openapi.json answer 404.
- Run
arcade deploy -e src/<pkg>/server.pyfrom the package root, not from the src folder the scaffold's next steps point to. - Delete a user connection with its connection_id (ac_...), not the list item's id (at_...).
- Read tool_executions with the te_ id from the execute response; the tc_ execution_id is refused.
- When authorizing for a user_id, sign in to the consent browser as that same Arcade user: the default user verification refuses others with user_mismatch.
- Custom OAuth providers need authorize_request.params with {{client_id}} and {{redirect_uri}}, or the consent fails with unknown_client.
Pay
Paying
Ways to pay checked 7 Nov 2026.
- ResultPartly. Reaches the checkout. A card form for $25 a month.
- CheckoutReached the card form and stopped: Growth, $25 a month. It checks out through a Stripe card form. No CAPTCHA at checkout.
- TrialNo card needed to start.
Leave
Leaving
- ResultPartly. Nothing to cancel. Its account stays.
- Its accountNot on its own: blocked. The account stays. Manage my account and the dashboard have no delete control, and the privacy policy says to email support@arcade.dev; mailing a real support desk is outside our rules. Everything we made was removed and both API keys refused afterwards (401), but the login still works.
Daily checks
Checked every day
What our agent finds on the product's own site each day, without an account.
Sign-up page,
Loads on the newest check; 11 of 11 days
llms.txt,
Found on the newest check; 11 of 11 days
MCP server,
Asks for a sign-in on the newest check; 4 of 11 days
OpenAPI,
Found on the newest check; 11 of 11 days
Compare
How Arcade compares
In MCP gateways
None of the 3 products we drove in MCP gateways is higher; 2 others share L3. As of 7 Nov 2026.
| Product | Level | Find to leave |
|---|---|---|
| Arcade (this page) | L3 | Find: On its own, Get in: On its own, Use: On its own, Pay: Partly, Leave: Partly. |
| Nango | L3 | Find: On its own, Get in: On its own, Use: On its own, Pay: Partly, Leave: Partly. |
| Composio | L3 | Find: On its own, Get in: On its own, Use: On its own, Pay: Partly, Leave: Partly. |
For the team behind Arcade
Agents get stuck at Pay
7 things to fix for agents: 4 small, 2 medium and 1 large
One fix can cover several findings.
Agent blockers: 6 foundAgent surfaces: 27 issues found
Get your findings, freeWhat a drive covers
We mail a private link to an address at arcade.dev.
It opens the two findings that matter most in full, and every other by title. There is nothing to sign in to.
- FindClear
- Get inClear
- UseClear
- PayBlocks agents
- LeavePartly
Or claim Arcade to see where each issue is and ask for a free drive after a fix.
Want Arcade driven every month? See Watch
On our standard
Best or ahead on 8 of 34 criteria, behind on 6. Each against every product we measured on it.
- Agents pick the right toolAhead of 73 of 91 others, 18 the same.
- llms.txt that worksAhead of 185 of 239 others, 54 the same.
- A path written for agentsAhead of 140 of 187 others, 47 the same.
- Every way in works for an agentAhead of 134 of 193 others, 59 the same.
- Answers sized for an agentBehind 86 of 116 others, 13 the same, 17 lower.
- Export before leavingBehind 117 of 189 others, 72 the same.
- Errors a program can readBehind 99 of 145 others, 46 the same.
- Tool names that sort themselvesBehind 86 of 91 others, 5 the same.
All 34 criteria
- A way in with no personAhead of 137 of 243 others, 106 the same.
- Truthful annotationsAhead of 50 of 91 others, 41 the same.
- A tool list that fitsAhead of 48 of 91 others, 43 the same.
- Sign-in for agents acting for a personLevel with 83 of 158 others, 75 lower.
- All validation problems at onceAhead of 34 of 57 others, 12 the same, 11 higher.
- One way to page through listsIn the middle 49 of 114 others lower, 52 the same, 13 higher.
- Prices an agent can readLevel with 117 of 153 others, 36 lower.
- Agents may read docs, API and sign-upLevel with 129 of 160 others, 31 lower.
- Fast to the first callLevel with 118 of 135 others, 17 lower.
- Descriptions that say when to use the toolLevel with 56 of 91 others, 22 lower, 13 higher.
- The best way in is built for agentsIn the middle 79 of 196 others lower, 55 the same, 62 higher.
- Terms allow an agent's accountLevel with 179 of 189 others, 10 lower.
- Real status codesLevel with 161 of 168 others, 7 lower.
- Errors say what to do nextLevel with 107 of 145 others, 22 lower, 16 higher.
- Typed, tight input schemasLevel with 59 of 91 others, 17 lower, 15 higher.
- WebMCP tools on the web appLevel with 133 of 136 others, 3 higher.
- Structured resultsIn the middle 22 of 78 others lower, 32 the same, 24 higher.
- An agent payment rail on the coreLevel with 205 of 216 others, 11 higher.
- A request id on every answerIn the middle 42 of 143 others lower, 47 the same, 54 higher.
- Results sized for a context windowIn the middle 16 of 70 others lower, 27 the same, 27 higher.
- Docs an agent can navigateLevel with 108 of 187 others, 19 lower, 60 higher.
- Spend an agent can see and capLevel with 105 of 159 others, 54 higher.
- Pages fit one fetchLevel with 118 of 187 others, 69 higher.
- Limits known before they hitLevel with 47 of 90 others, 43 higher.
- Versions an agent can pinBehind 74 of 127 others, 43 the same, 10 lower.
- Delete by API, checkedBehind 52 of 97 others, 45 the same.
Where Arcade fits
- Tasks
- Wall
- Topics
- MCP serversFind step
- Docs and llms.txtFind step
- MCP sign-inGet in step
- Sign-upGet in step
- MCP and APIUse step
- Card checkoutPay step
- Delete the accountLeave step
History
Level by drive
One drive so far, on 7 Nov 2026.
Read more on this step
Pay is where our agent stopped at Arcade.
- Finding · 18 Nov 2026How far does an AI agent get in a card checkout?Our agent reached the checkout at 88 of the 190 products it set out to buy from, and the card form at 68.Read
- Guide · 18 Nov 2026Let AI agents pay you per call: x402 and MPPx402 or MPP on a paid endpoint, your prices where agents read them, and plans an agent can buy by API or MCP.Read
- Guide · 28 Oct 2026Tell an AI agent a feature needs a paid planHow to answer an AI agent at a paid-plan wall: 402 or 403, a Problem Details body with the required plan, and an upgrade link a person can open.Read
Every earlier drive is in the history. The level follows these rules. Data under CC BY 4.0.
Found an error? Tell us on this page or read how corrections work.
Questions people ask about Arcade
- Can an AI agent create an Arcade account on its own?
- Yes, in a browser, without a person. Done by the agent in its own browser at app.arcade.dev/register (account.arcade.dev): email, company, full name, nickname and password, no CAPTCHA (company and name asked; our agent answered). After email verification the dashboard requires a passkey or 2FA. The agent set up an authenticator app: the setup key is shown as text, so it computed the codes itself.
- Does Arcade have an MCP server, and can an agent connect to it?
- Yes. Arcade has an official MCP server. The server signs an agent in with OAuth. A new agent can get in on its own: in the drive of 7 Nov 2026, our agent signed up, verified and got its key without a person. Through the MCP server, our agent tried 3 tasks and did 3.
- How does an AI agent get an Arcade API key?
- In the drive of 7 Nov 2026, our agent got its Arcade API key in a browser, without a person. The agent made a project API key on the dashboard's API Keys page in its own browser (the documented route). MCP sign-in (CIMD, PKCE) and
arcade logingave OAuth tokens through consent in the agent's browser. Contrary to the docs,arcade loginstores a user token, not an API key. - Can an AI agent delete its Arcade account?
- Not on its own: blocked. The account stays. Manage my account and the dashboard have no delete control, and the privacy policy says to email support@arcade.dev; mailing a real support desk is outside our rules. Everything we made was removed and both API keys refused afterwards (401), but the login still works.
- Can an AI agent pay for Arcade?
- Partly. Our agent reached a card form for $25 a month. No card needed to start.
Cite this page
The page as a reference, in plain text, APA and BibTeX.
Cite this
AgentDrives, “Arcade: AI agent test drive”, 7 November 2026, https://agentdrives.com/tools/arcade. CC BY 4.0.
AgentDrives. (2026, November 7). Arcade: AI agent test drive. https://agentdrives.com/tools/arcade
@misc{agentdrives_arcade_2026,
author = {{AgentDrives}},
title = {Arcade: AI agent test drive},
year = {2026},
month = nov,
url = {https://agentdrives.com/tools/arcade},
note = {Agent test drive, CC BY 4.0},
}Something wrong on this page?
We check it against the drive and correct the page, with the date.