Agent readiness · magicbell.com
2026-08-12 10:16 UTC · formula v9.2 · 7.5s
There is no door built for a machine anywhere on your domain.
None of the nine known agent entry paths answered, including /agent-signup.md and /.well-known/agent-access.json.
A dashed column is a stage we could not measure, not a stage you failed.
- A · Discovery
- 4/5
- B · Agent entry
- 1/4
- C · Registration
- n/m
- D · Provisioning
- 2/3
- E · Integration
- 1/2
of the 14 points that apply to you and we could measure · 3 of 17 were not scored
You 8 · category median 9.5 · best novu.co 12 · 15 deterministic HTTP checks, published formula
Fix this first
This is the whole list: everything we could measure and you do not pass. Another 3 points sit behind checks we could not evaluate on your domain, and each of those lines below says what would make it measurable.
- 1llms.txt publishedminutes · +1
Publish /llms.txt: a markdown list linking your quickstart, API reference, pricing and package name. It is the cheapest file on this list.
- 2Agent entry pointan afternoon · +2
Publish /agent-signup.md: the path from nothing to a working key, written as steps for a machine rather than a tour for a person. Ours is at /agent-signup.md if you want a worked example, and it says plainly that we have nothing to sign up for.
- 3Programmatic key provisioningan afternoon · +1
Document how a key is created without opening a dashboard: management API, service account or CLI. If no such path exists, that gap is the finding, not the docs.
Cheap too, same week
- Machine-readable API description an afternoon · +1Publish an OpenAPI file at /openapi.json, or serve markdown when a client sends Accept: text/markdown. Either one stops agents guessing your API.
Then, when you have room
- MCP surface a project · +1Wrap your top three API calls in an MCP server and declare it at /.well-known/mcp.json.
Take it with you
6 of the checks an agent depends on, magicbell.com does not pass.
Send yourself the list above with the checks that cost the most, and a permanent link. Whoever owns the fix is usually not the person who ran the scan.
One email with the report, and one follow-up asking what you did with it. No newsletter, no sequence. Delete on request at hello@letagentsin.com.
None of this tells you whether an agent picked you
Everything above is a file an agent could read. Whether one actually chose you is a different measurement, and we have not run it on magicbell.com. What we have run it on is a rich text editor vendor: six agents, one brief, six isolated copies of a real codebase. All six picked the same competitor, and the vendor being studied was never named, not even on a rejection list, while the runs named and dismissed ten alternatives between them. In an earlier round, run before we isolated the copies, it was named twice and struck off both times in four words: Fully commercial, licence key required. That study is about them, not about you. It is here because it is the difference between a file and a decision.
Against notification infrastructure
You are 5 of 6 in a category where a developer is trying to send one event to email, push and Slack without writing three integrations.
- 1novu.co12/15
- 2knock.app11/15
- 3onesignal.com10/15
- 4courier.com9/15
- 5magicbell.comyou8/14
- 6pusher.com7/16
Where competitors pass and magicbell.com does not
- llms.txt publishedknock.app · courier.com · novu.co · onesignal.com
- Machine-readable API descriptionknock.app · courier.com · novu.co · onesignal.com
- MCP surfaceknock.app · courier.com · novu.co
- Agent entry pointnovu.co
Only domains we have actually scanned appear here, so the ranking is a floor: a competitor missing from the list has not been measured, not beaten.
By stage
Every check
Each line is one HTTP observation with a published rule, so you can rerun any of them yourself. Most are the same tomorrow if you are: the two that are not are the bot gate and the pricing page, which answer inconsistently on their own, which is why those run more than once and say so when the tries disagreed. PASS and PART are counted. UNMEASURED means we could not evaluate it, N/A means it does not apply to a product like yours, and neither is in the score or its denominator. See the formula
A · Discovery
- PASSAnswers an agent user-agentAnswered 200 to LetAgentsIn/1.0 (+https://letagentsin.com/methodology)
- FAILllms.txt publishedNo llms.txt at any of the 4 locations probedA curated map of your docs is the cheapest way to control what an agent reads first. Worth one point and not more: in eighteen isolated agent runs across four categories, not one cited llms.txt among its sources, and an independent ninety-day measurement published by Otterly in February 2026 found it served 84 requests against 62,100 AI-bot visits.
- PASSDocs readable without JavaScript6,187 characters of text without JS, on https://www.magicbell.com/docs/api/authentication/user rather than on https://www.magicbell.com/docs
- PASSOn-demand agents not blockedNo on-demand agent is blocked
- N/APaths robots.txt points at answerrobots.txt names no concrete path, only patterns or nothing, so there is no claim to check
- PASSNo punishing crawl delayNo Crawl-delay applies to the agents we check
B · Agent entry
- FAILAgent entry pointNone of the 9 known agent entry paths returns a file rather than your page shellA markdown file written for a machine turns a guessing game into a procedure it can follow.
- PASSOAuth dynamic client registrationregistration_endpoint published, and client_credentials is among the 13 advertised grants, so an unattended agent has a documented path to a token
- FAILMCP surfaceNo MCP surface: nothing answered at mcp.magicbell.com, mcp.magicbell.com/mcp, mcp.magicbell.com/v1/mcp, api.magicbell.com/mcp, /mcp or /api/mcp, and no file mentions MCPAn MCP server turns your API from something an agent reads about into something it can call.
C · Registration
- UNMEASUREDNo CAPTCHA in the signup HTMLUnmeasurable: we found no link to an account signup on the pages we read, while you publish prices at https://www.magicbell.com/pricing, so this is a gap in our reading rather than a finding about youLink signup from your home page or your pricing page and this becomes measurable.
- UNMEASUREDSignup reachable without a browserUnmeasurable: we found no link to an account signup on the pages we read, while you publish prices at https://www.magicbell.com/pricing, so this is a gap in our reading rather than a finding about youLink signup from your home page or your pricing page and this becomes measurable.
D · Provisioning
- PARTProgrammatic key provisioning1 of 7 provisioning phrases across the 4 documents we read: "management api"Documented key creation is the difference between a two-minute integration and a support ticket.
- PASSFree tier or no-card trial stated in textFree tier or no-card signals at https://www.magicbell.com/pricing: "$0"
E · Integration
- PASSTyped SDK on the registrymagicbell@4.5.1 ships types, matched from the registry by who publishes it rather than by a link on your site
- FAILMachine-readable API descriptionNo OpenAPI spec at the 5 usual paths, none declared by https://www.magicbell.com/docs, and no markdown negotiationAn OpenAPI file or markdown negotiation lets an agent read your API instead of guessing it.Point at your spec from your docs page with rel="service-desc" and an agent finds it without guessing.
What we discovered on the way
- Docs
- https://www.magicbell.com/docs · linked from the site
- Pricing
- https://www.magicbell.com/pricing · linked from the site
- Signup
- not found
- npm package
- magicbell (registry guess)
- GitHub
- not found
- Package licence
- SEE LICENSE IN LICENSE
- Crawl-delay
- none
- Content-Signal
- none
- Blocked on-demand agents
- none
Your package declares SEE LICENSE IN LICENSE rather than a plain licence identifier. We do not score this, because a commercial licence is a business model, not a defect. We report it because in six agent runs on a comparable decision, every model that hit a licence key requirement dropped that vendor in one line, without opening the product. What the runs showed