Head to head · Content management · October 2026 research run

Cosmic vs Ghost

Cosmic scores 72.5 (BB) on agent readiness against Ghost's 51.2 (D), and leads in 4 of 7 scored categories. Ghost leads on security & auth, payments & pricing and transparency & trust. Both do content management.

Best headless CMS and website publishing for AI agents · All 136 cms comparisons

Which one, for what

Cosmic BB

Best for Teams that want an agent to set up and fill a hosted CMS end to end, including signup.

Ahead on

  • Reliability, 90 against 83
  • Schema & documentation, 88 against 51
  • Agent ergonomics, 82 against 69

Also in its favour

  • Agent-ready, a grade of BB or better
  • A hosted endpoint, with nothing to install
  • Runs on your own machine
  • Free to start without a card
  • No incidents deducted, where Ghost loses 15 points for them

Watch for

No SOC 2, ISO 27001 or HIPAA certification, by the security page's own statement

Ghost D

Best for A publication, blog or newsletter where an agent drafts posts and a person or a higher role publishes.

Ahead on

  • Security & auth, 56 against 47
  • Payments & pricing, 50 against 45
  • Transparency & trust, 75 against 63

Also in its favour

  • Open source

Watch for

No OpenAPI or other machine-readable description of the Admin API was found in the docs or the repository

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.

Schema & documentation Cosmic +37

Cosmic 88

openapi.json at www.cosmicjs.com covers 21 operations on 12 paths, each with a summary, a description and documented 401, 402, 404, 429 and 500 responses, though the docs describe more endpoints than it covers (revisions, AI, search); MCP tools have typed inputs (20 of 25). llms.txt and every docs page as Markdown (10). Tool descriptions say when to use each tool and when not to, such as never calling delete speculatively (18 of 20). 11 enums in the spec, typed metafields and validation messages (11 of 15). Examples on every API page and an error model with a sample (14 of 15). API v3 paths and a dated changelog (15). 20+10+18+11+14+15 = 88.

Ghost 51

Per the 8 October check, not re-read on 4 November (docs host paused under the terms). No OpenAPI or similar file was found in the docs index or the repository (0 of 25). llms.txt with 177 links, llms-full.txt and a Markdown copy of every page (10). The reference states what each endpoint does and which fields are required, with little on when not to use one (10 of 20). No published schemas. Required fields are named (title on create, updated_at on edit), and post content is one Lexical JSON string (6 of 15). Request and response examples on each page and token generation in Bash, JavaScript, Ruby and Python. The errors page lists five status codes and the errorType and message fields, without a list per endpoint (10 of 15). Accept-Version header, a stability index per endpoint, a breaking changes page and release notes (15).

Agent ergonomics Cosmic +13

Cosmic 82

24 tools, in the 11 to 30 band (15), plus 5 because omitting the write key gives a read-only subset (20 of 25). limit, props, depth, cursor pagination with after, status and locale filters (20). Status codes with specific messages and fix-or-retry guidance per code (17 of 20). readOnlyHint and destructiveHint on the MCP tools; no idempotency keys for writes, though agent signup is idempotent (14 of 20). Few required parameters, a TypeScript SDK and a CLI, no SDK in a second language found (11 of 15). 20+20+17+14+11 = 82.

Ghost 69

Per the 8 October check, not re-read on 4 November (docs host paused under the terms). fields, include and formats size a response, and lists return 15 records by default (20 of 25). page, limit up to 100, order and a filter query language, with meta.pagination on every list (20). Errors are JSON with message and errorType, and the documented list is short (13 of 20). No idempotency keys. A PUT must carry the current updated_at and is refused on a collision, and a repeated POST creates a second post (8 of 20). A create needs only title. The one official client is JavaScript, @tryghost/admin-api (8 of 15).

Transparency & trust Ghost +12

Cosmic 63

A closed service; the MCP server is MIT (18 of 30). Terms and privacy policy of 13 January 2025 give general retention, and the security FAQ says data is kept indefinitely unless deletion is requested; storage in MongoDB Atlas is disclosed (15 of 30). Some fields are marked deprecated in v3 and the changelog is dated; no deprecation policy (8 of 20). The security page answers who the sub-processors are, but that answer wasn't readable in the page source (12 of 20). 18+15+8+12 = 53.

Ghost 75

MIT, copyright Ghost Foundation, with the source public (30). Per the 8 October check, not re-read on 4 November (ghost.org and docs.ghost.org paused under the terms): a self-hosted site keeps content on its owner's server. For Ghost(Pro) the privacy policy places servers in Amsterdam and a DPA lists subprocessors in an annex, while retention is stated only as kept where necessary and the terms make the customer responsible for backups (18 of 30). A table dates each major's end of life (5.x in January 2026, 6.x to be confirmed), endpoints carry a stability index and breaking changes are catalogued per major. No notice period in days was found (15 of 20). The configuration page documents Gravatar, structured data and IndexNow calls with a switch for each and useTinfoil for all. The default configuration also points updateCheck at updates.ghost.org, which that section doesn't list; the 4 November source sends only ghost_version in that request (13 of 20).

Security & auth Ghost +9

Cosmic 47

Separate read and write keys per bucket, rotatable independently (25), less 10 because the docs give read_key as a query parameter on GET (15 of 30). Read-only MCP access by leaving out the write key, four user roles on every plan, and a delete tool told not to run speculatively; no approval step (14 of 20). Objects hold text written by people and AI generation writes into the bucket; no prompt-injection guidance was found (3 of 15). An append-only audit log of security-relevant actions on Workspace or Pro plans (10 of 15). No SOC 2, ISO 27001 or HIPAA by the security page's own statement, no security.txt, reports to support email (5 of 20). 15+14+3+10+5 = 47.

Ghost 56

An Admin API key belongs to one custom integration, can be regenerated, and signs tokens that last at most 5 minutes and travel in the Authorization header. Integrations have one fixed permission set and no scopes, so we scored plain revocable keys plus 2 for role-bound staff tokens (22 of 30). The Content API key is sent as a query parameter. It reads published content only and isn't the graded credential, so no deduction. A Contributor's staff token can add and edit drafts but not change status, per the post model. No read-only Admin key and no approval step for a publish or delete (12 of 20). Posts, members and comments hold text written by other people, and no prompt-injection guidance was found (3 of 15). An actions log is readable at GET /actions. No per-call log was found (8 of 15). A disclosure policy with response times and security@ghost.org per the 8 October check, and advisories published on GitHub with credits (OSV holds 46 published from 4 November 2025 to 27 October 2026). No bug bounty or certification was found, and ghost.org/.well-known/security.txt returned 404 (11 of 20).

Reliability Cosmic +7

Cosmic 90

Graded on the hosted service with the hosted lines. Statuspage at cosmicstatus.com with API and Web Dashboard components (20). The history page shows no incidents in September, October or November 2026, and the newest incident in the feed is from 26 May 2025. The MCP server isn't a component, so 25 of 30. 100 requests a second with a burst of 200, timeouts and size limits are documented (15). 429 is documented and the error page says which codes to fix and not retry; no Retry-After guidance was found. Agent signup is idempotent (12 of 15). The security page says Enterprise has an SLA from 99.95 per cent uptime; the SLA page itself wasn't read (8 of 10). Hosted MCP is announced as available now, with no beta label (10). 20+25+15+12+8+10 = 90.

Ghost 83

Read with the local-software lines, since the graded surface is the open-source release its owner hosts. Ghost(Pro)'s status page isn't scored here. Official ghost package on npm, with Node.js ^22.23.1 or ^24.20.0 in engines in the 4 November source, and a supported stack named in the hosting guide (20). A public CI workflow (ci.yml) runs on pushes to main and runs the unit, type and end-to-end tests; GitHub's pages answer 403 in this session, so today's result is unread (the badge read failing on 8 October) (15 of 25). The repository showed 33 open issues and 140 open pull requests against weekly releases on 8 October; not re-read (20 of 25). Semver version numbers, release notes per version and a breaking changes page per major, with the 6.0 removal of limit=all documented there (13 of 15). 6.x is the current stable line (15).

Payments & pricing Ghost +5

Cosmic 45

No x402, MPP or L402 (0). Plan prices and quotas are public, with add-ons; no per-call price (10). Free plan with no card (20). Agent signup returns working bucket keys with no human signup; the human claims the project by emailed code within 14 days, so partial (15 of 20). 0+10+20+15 = 45.

Ghost 50

Per the 8 October check, not re-read on 4 November (ghost.org paused under the terms). Scored with the self-hosted rule, taking prices from Ghost(Pro) beside the free software. No x402, MPP or L402 found (0). The software is free under MIT, and Ghost(Pro) lists $18, $29 and $199 a month billed yearly at up to 1,000 members (20). Self-hosting needs no card or account. Whether the Ghost(Pro) trial asks for a card wasn't readable (20). The docs create Admin API keys in the Ghost Admin screens, so a person sets up the integration. Install is scriptable, so half (10).

Score by category

CategoryWeight this runCosmicGhostEdge
Reliability16%209083Cosmic +7
Performance10%pendingpendingpendingnot scored in this run
Schema & documentation13%16.28851Cosmic +37
Agent ergonomics13%16.28269Cosmic +13
Security & auth14%17.54756Ghost +9
Payments & pricing10%12.54550Ghost +5
Task success10%pendingpendingpendingnot scored in this run
Maintenance & community7%8.88685Cosmic +1
Transparency & trust7%8.86375Ghost +12
Negative events≤150-15
TotalBB 72.5/100D 51.2/100

Facts side by side

FactCosmicGhost
KindHTTP APIHTTP API
VendorCosmic Inc.Ghost Foundation
Hosted endpointhttps://mcp.cosmicjs.com/v1/buckets/{bucket_slug}no (local only)
TransportsHTTP, Streamable HTTP, stdioHTTP
AuthAPI keyAPI key
PricingFreemiumFreemium
x402nono
LicencenoneMIT, copyright Ghost Foundation. Ghost(Pro) is a paid hosting service under the Ghost Foundation's terms
Tools exposed24none
Read-only variant documentedyesno
llms.txtyesyes
MCP registrycom.cosmicjs/mcpnot listed
Last release2026-11-022026-11-03
Terms last updatedno date givenno date given
Privacy policy last updatedno date givenno date given
Customer content may train modelsnot found in the textnot found in the text
Terms restrict automated accessyesyes
Terms restrict benchmarkingnot found in the textnot found in the text
Terms or service can change without noticenot found in the textnot found in the text
Arbitration or class-action waiveryesnot found in the text
Popularity295 npm/wk56k stars, 24k npm/wk

How the scores have moved

Verdicts

Cosmic

Cosmic gives agents 24 annotated MCP tools with when-to-use descriptions, a read-only mode by key, documented limits and an agent signup API that returns working keys before a human claims the project. It holds no SOC 2 or ISO certification by its own statement, and read keys can travel in the URL.

Ghost

A create needs only a title, updates are checked against updated_at so a stale write is refused, and a Contributor's staff token can draft without publishing. The Admin API has no OpenAPI file and one fixed permission set per integration, and OSV lists 46 Ghost advisories in the year to 27 October 2026, 15 high or critical.

Before you call either

Cosmic

  1. Send only the read key in the Bearer token unless the task needs writes
  2. Use cosmic_agent_signup when the user has no Cosmic account, and hand the human the claim URL within 14 days
  3. Fix and don't retry on 400; wait and retry on 429
  4. Use props, depth=0 and cursor pagination with after to stay under the 6 MB response limit
  5. Confirm with the user before cosmic_objects_delete, which deletes permanently

Ghost

  1. Sign a fresh HS256 token for each batch of calls. Decode the hex secret to bytes, set kid to the key id, aud to /admin/ and exp at most 5 minutes ahead
  2. Set status to draft on every create unless told to publish, and publish later with a PUT that sets status to published
  3. GET the post before each PUT and send its updated_at back. Tags and authors in a PUT replace the existing lists
  4. Send content as a Lexical JSON string, or add ?source=html and send html. The HTML conversion is lossy unless wrapped in an HTML card
  5. Page through lists with limit up to 100 and page. Since Ghost 6.0 limit=all returns 100 items without an error

Questions

Which is better for AI agents, Cosmic or Ghost?

Cosmic scores 72.5 (BB) on agent readiness against Ghost's 51.2 (D), and leads in 4 of 7 scored categories. Ghost leads on security & auth, payments & pricing and transparency & trust.

Do Cosmic and Ghost need an API key?

Both need an API key.

Can an agent call Cosmic and Ghost without installing anything?

Cosmic has a hosted endpoint at https://mcp.cosmicjs.com/v1/buckets/{bucket_slug}. No hosted endpoint is listed for Ghost.

Are Cosmic and Ghost open source?

No open-source release is listed for Cosmic. Ghost is open source (MIT, copyright Ghost Foundation. Ghost(Pro) is a paid hosting service under the Ghost Foundation's terms).

Other comparisons with Cosmic or Ghost

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.