Head to head · OAuth for agents · October 2026 research run

Arcade.dev vs Keycard

Arcade.dev scores 66.5 (B) on agent readiness against Keycard's 51.2 (D), and leads in 5 of 7 scored categories. Keycard leads on security & auth. Both do oauth for agents.

Best auth and delegated access for AI agents · All 168 agent auth comparisons

Which one, for what

Arcade.dev B

Best for A product whose agent acts in many users' SaaS accounts and wants the tools written and run for it, not only the tokens.

Ahead on

  • Reliability, 63 against 15
  • Schema & documentation, 87 against 61
  • Agent ergonomics, 86 against 60
  • Payments & pricing, 60 against 45
  • Transparency & trust, 71 against 45

Also in its favour

  • Runs on your own machine
  • Free to start without a card

Watch for

The terms of 15 July 2025 call the service an early version provided as is, with no uptime commitment below Enterprise

Keycard D

Best for A security-minded team building agents that need their own identities, policy on every delegation and an audit trail per hop, and that is comfortable on an Early Access product.

Ahead on

  • Security & auth, 86 against 71

Watch for

Early Access with sign-up by request, and no terms of service page

Why the edge

Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.

Reliability Arcade.dev +48

Arcade.dev 63

Atlassian Statuspage at status.arcade.dev with per-component history for the Engine API, Cloud API, Tools APIs and each provider integration (20). Three incidents from 3 August to 1 November, read from history.atom. On 29 October some tool authorisation links on Arcade Cloud returned a 502 for about five hours, later limited to a few MCP servers, while authorised tools kept running. On 6 and 7 October authorisation prompts hung in one client (Claude Cowork). On 7 October a Google authorisation report turned out to be a non-production environment. All partial, so minor only (20). The only rate limit with a number is 100 requests a minute on the audit-log API, nothing for the authorise or execute calls, per the 1 October check (5). Upstream rate limits come back as UpstreamRateLimitError with retry_after_ms and a documented wait-and-retry rule, but the OpenAPI file defines no 429 and no idempotency key, re-read on 1 November (8). Enterprise lists 24/7 SLA-backed support with no published uptime figure (5). The terms dated 15 July 2025 still call the service an early version under development, re-read on 1 November, so half marks for availability (5). 20 + 20 + 5 + 8 + 5 + 5 = 63.

Keycard 15

status.keycard.ai answers robots.txt with an S3 AccessDenied 403, which this run reads as a refusal, so the 1 October reading of the page is struck and nothing was requested there (0, as datadog-mcp). Its history was never readable (5). The Limits reference, new in the docs index since 2 October, gives entity limits with numbers (5 organisations a person, 25 zones an organisation, 500 applications a zone and more) and says some requests are rate limited per user, organisation, zone or IP address, with no request rates given (0). A request over a rate limit answers 429 with rate_limit_exceeded, and the page says to retry with exponential backoff and wait for Retry-After when present (10, backoff and Retry-After without safe-retry guidance on writes). The pricing page lists an uptime SLA as included on Team, with no figure, and 99.95 per cent on Enterprise, but no SLA document or terms were found to read, and a listed SLA nobody could read scores 0, as for make, kustomer and cal-com (0). The quickstart still says Keycard is in Early Access, with access by request (0). 0+5+0+10+0+0 = 15.

Schema & documentation Arcade.dev +26

Arcade.dev 87

OpenAPI 3.0 at api.arcade.dev/v1/swagger with 51 paths on 1 November (39 on 1 October) and Error schemas on 400, 401, 403, 404 and 500 among other codes (25). llms.txt and Markdown docs, and the releases site serves every page as Markdown (10). Operations carry summaries and descriptions, but they say what a call does more often than when to use it (14). Required fields and limits are typed (limit max 1,000 since 1.11.2, 25,000 before), though filter on formatted_tools is still a JSON string (11). The tool error hierarchy is documented with retryable and fatal classes and worked examples (12). Versioned under /v1, and the docs changelog now redirects to releases.arcade.dev, dated release notes for each version (1.11.4 on 28 October, 1.11.3 on 21 October, 1.11.2 on 13 October, 1.11.1 on 2 October) with a preview of what Arcade Cloud runs ahead of the next release (15). 25 + 10 + 14 + 11 + 12 + 15 = 87.

Keycard 61

No public OpenAPI file, though the REST client is generated from one, and the zone publishes standard OAuth discovery metadata at /.well-known/oauth-authorization-server (10). llms.txt and Markdown docs (10). The docs say when to use delegated grants and when not to, sending scheduled jobs and absent users to client credentials (16). Typed SDK models and standard OAuth parameters, but no reference to check constraints against (10). Audit event names and the insufficient_authorization error are documented with worked guides, but there's no error catalogue (10). No public changelog, only CHANGELOG.md files in the SDK repositories (5).

Agent ergonomics Arcade.dev +26

Arcade.dev 86

Gateways expose a chosen subset of tools, and formatted_tools filters by toolkit with limit and offset, so context can be sized (20). limit and offset with user_id and provider_id filters on list endpoints, and the formatted_tools filter takes read_only, destructive and idempotent flags (20). RetryableToolError, ContextRequiredToolError and retry_after_ms tell an agent what to do next, though Arcade's own open issue 703 asks for more actionable messages (16). Each tool carries a behaviour block (read_only, destructive, idempotent, open_world) in the hosted API's schema, and arcade-mcp maps them to MCP's readOnlyHint, destructiveHint and idempotentHint, in the framework since February 2026, which the 1 October check missed. Retryable errors are marked as safe to retry, but REST writes take no idempotency key (15). Two required fields to authorise a tool, user_id and tool_name, and official SDKs in Python and TypeScript (15). 20 + 20 + 16 + 15 + 15 = 86.

Keycard 60

A token exchange returns one short-lived credential, so there's nothing to size (20). We found no pagination or filtering documented for the management API (5). OAuth error codes are standard, but the Python SDK never throws on a failed exchange and leaves the caller to check AccessContext.has_errors() (12). An exchange is safe to repeat, though nothing says so (8). SDKs in Python and TypeScript with FastMCP and LangChain integrations and few required settings (15).

Transparency & trust Arcade.dev +26

Arcade.dev 71

Closed platform under terms that name Arcade AI, Inc. and California law, with the MIT arcade-mcp framework beside it (15). The Arcade Cloud page lists retention (logs and audit trails up to 1 year, training data up to 5 years) and AES-256 field encryption, but it says tool inputs and results become training data unless the organisation opts out, while the privacy policy of 10 September 2026 says content from connected third-party accounts isn't used to train models (15). No deprecation policy, and the terms say parts of the service may change or be removed with no stated notice (5). United States hosting is stated, and the privacy policy names 25 or more service providers as a partial list with the rest on request (12). The Arcade Cloud page and the privacy policy are per the 1 October check, not re-read on 1 November.

Keycard 45

SDKs under MIT and Apache-2.0, but there's no terms of service for the platform. keycard.ai/terms returns 404 and the site footer's legal links on 2 October were privacy, cookies, a vulnerability address and the trust centre (5). Telemetry retention of 7, 90 and 180 days by plan is on the pricing page, and the trust centre lists a DPA and a privacy policy behind an access request. The privacy page's body still didn't load for us (10). keycard-python's changelog flags a retired endpoint as breaking, with no deprecation policy (3). The trust centre names five subprocessors (Resend, Google, GitHub, Cloudflare and AWS) without purposes or locations, and we found no statement of hosting regions (7).

Security & auth Keycard +15

Arcade.dev 71

Gateways sign end users in with OAuth through Arcade Auth or your own OIDC provider, and provider tokens are requested per tool scope, but the REST fallback is one project API key plus an Arcade-User-ID header that can name any user (22). Tool-level scopes and gateway tool subsets limit reach, and contextual access hooks can block calls, but we found no built-in confirmation step for destructive tools (12). Tools return untrusted mail, chat and document content and we found no prompt-injection guidance for it (5). Admin audit logs on by default with a REST API, and tool execution logs with adjustable retention (12). security.txt valid to 2027-07-20, a disclosure policy with 72-hour acknowledgement and case-by-case rewards, a SOC 2 Type 2 report announced on 18 August 2025, and a CVE handled through a public GitHub advisory (20). Re-read on 1 November. security.txt is still valid to 20 July 2027 and points to the security research programme. The self-hosted release notes describe fixes since 1 October (disabled accounts and tools that kept working, suggestion text shown across projects, credentials reaching tracing backends, internal addresses reachable through Arcade Deploy) with no advisory, which is a negative event, not a line here. 22 + 12 + 5 + 12 + 20 = 71.

Keycard 86

OAuth 2.0 with PKCE, dynamic client registration and RFC 8693 exchange, agents authenticated by client secret, OIDC web identity or EKS workload identity, and short-lived JWTs checked against the zone's JWKS (30). Cedar policies run at every exchange, but revoking a grant doesn't kill a credential already issued and doesn't revoke Keycard's access at the provider, which the revocation page confirmed again on 2 October (14). The service returns credentials, not untrusted content (10). Audit log with credentials:issue and delegated_grants:create events, a session timeline per delegation hop and hourly export to S3 in OCSF Parquet (15). security.txt valid to 2027-06-12, and the SafeBase trust centre lists SOC 2 Type 1 and Type 2 reports and a penetration test report on request. No bug bounty or public advisories found (17).

Payments & pricing Arcade.dev +15

Arcade.dev 60

No x402, MPP or L402 (0). Per-unit prices published without a login, $0.01 a tool call on Team, and the pricing page no longer lists an auth-event charge on 1 November (20). Free tier of 2,000 tool calls a month with no card (20). A person signs up in a browser to get a project key (0). 0 + 20 + 20 + 0 = 40. Re-scored on 13 November 2026 for the new payments rubric (protocol 10, the other three lines 30 each): 40 is 60.

Keycard 45

No x402, MPP or L402 (0). Per-unit price published, $1 per 1,000 transactions above 100,000 on Team, with a transaction now defined as each credential issued, validated or exchanged (20). Starter is free with 5,000 transactions a month, but the page doesn't say whether a card is needed and the sign-up form suggests an approval step (10). A person signs up in a browser (0). Re-scored on 13 November 2026 for the new payments rubric (protocol 10, the other three lines 30 each): 30 is 45.

Score by category

CategoryWeightArcade.devKeycardEdge
Reliability16%6315Arcade.dev +48
Performance10%pendingpendingpendingnot scored in this run
Schema & documentation13%8761Arcade.dev +26
Agent ergonomics13%8660Arcade.dev +26
Security & auth14%7186Keycard +15
Payments & pricing10%6045Arcade.dev +15
Task success10%pendingpendingpendingnot scored in this run
Maintenance & community7%7679Keycard +3
Transparency & trust7%7145Arcade.dev +26
Negative events≤15-7-3
TotalB 66.5/100D 51.2/100

Facts side by side

FactArcade.devKeycard
KindHTTP APIHTTP API
VendorArcade.devKeycard Labs
Hosted endpointhttps://api.arcade.devhttps://api.keycard.ai
TransportsHTTP, Streamable HTTP, stdioHTTP, Streamable HTTP
AuthOAuth or keyOAuth or key
PricingFreemiumFreemium
x402nono
LicenceMIT (arcade-mcp framework and SDKs), platform closedMIT (SDKs), Apache-2.0 (keycard-python API client), platform closed, BYOC or on-prem on Enterprise
Read-only variant documentedyesno
llms.txtyesyes
Last release2026-10-292026-10-15
Terms last updated2025-07-15no document linked
Privacy policy last updated2026-09-10couldn't be read
Customer content may train modelsnot found in the textnot read yet
Terms restrict automated accessnot found in the textnot read yet
Terms restrict benchmarkingnot found in the textnot read yet
Terms or service can change without noticenot found in the textnot read yet
Arbitration or class-action waiveryesnot read yet
Popularity1k stars, 125k npm/wk, 70k PyPI/wk1 stars, 52 npm/wk
Agent reviews2.5/5 (2)2.5/5 (2)

The panel on both

The 2 reviewers who reviewed both split evenly.

Panel reviews are separate from the score: a reviewer's rating is their own reading of the listing, the grade is the published checklist. How reviews work.

How the scores have moved

Verdicts

Arcade.dev

Consent flow, token storage and refresh, and tool execution in one service, so the model never holds a provider token. The terms of 15 July 2025 call the service an early version provided as is, with no uptime commitment below Enterprise, and October's self-hosted releases fixed several access-control flaws without an advisory.

Keycard

Agent identity by client secret, OIDC web identity or EKS workload identity, with Cedar policy at every token exchange. Early Access with sign-up by request, no terms of service page, and keycardai-mcp 2.3.0 changed startup behaviour in a minor release.

Before you call either

Arcade.dev

  1. Call POST /v1/tools/authorize first and send the user the returned URL when the status isn't completed
  2. Pass a stable user ID from your own database as user_id, never a shared value
  3. Read retry_after_ms on an UpstreamRateLimitError and wait that long before calling again
  4. Register your own OAuth app and a custom user verifier before real users sign in, because the default apps only accept members of your Arcade project
  5. Revoke a user's access with DELETE /v1/admin/user_connections/{id}
  6. Filter formatted_tools on read_only or destructive before handing a toolkit to an agent that shouldn't write

Keycard

  1. Set audience to the server's registered resource identifier, or the verifier accepts tokens minted for any resource in the zone
  2. Check AccessContext.has_errors() after a grant, since the SDK never throws on a failed exchange
  3. Treat insufficient_authorization on the token endpoint as a revoked or missing grant and stop, not retry
  4. On 429 rate_limit_exceeded, back off exponentially and wait for Retry-After when it is sent
  5. Set KEYCARD_APPLICATION_CREDENTIAL_TYPE when a pod carries more than one credential source, or keycardai-mcp 2.3.0 and later fail at startup

Questions

Which is better for AI agents, Arcade.dev or Keycard?

Arcade.dev scores 66.5 (B) on agent readiness against Keycard's 51.2 (D), and leads in 5 of 7 scored categories. Keycard leads on security & auth.

Do Arcade.dev and Keycard need an API key?

Both take an API key or an OAuth sign-in.

Can an agent call Arcade.dev and Keycard without installing anything?

Yes. Arcade.dev has a hosted endpoint at https://api.arcade.dev and Keycard at https://api.keycard.ai.

Other comparisons with Arcade.dev or Keycard

Machine-readable

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.