Cursor Agent SDK

# Fix list: Cursor Agent SDK

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/cursor-agent-sdk, the October 2026 research run, assessed 27 October 2026. Grade B, 63.6 out of 100.

This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public.

For a coding agent working on Cursor Agent SDK: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published.

## 1. Reliability, 45 out of 100, up to 11 more on the total

Why it scored 45: Local-package reading, since the owner runs the framework, with Cursor's hosting as an option. Official npm package with Node 22.13 or newer stated in engines and in the docs, which also say never Bun (20). Closed source, so no public CI or test suite (0). No public issue tracker. github.com/cursor/cursor holds a README and SECURITY.md, and forum.cursor.com has no Agent SDK tag, so there's nothing to count open crashes against. The troubleshooting page lists known failures with fixes (5). A dated changelog of two entries, 0.2.29 and 0.3.0, both 23 October 2026, and 0.3.0 calls out the removed `persistent` option, but the 27 earlier @cursor/bdk versions and 123 @cursor/july versions have no notes, and one build ships under three package names (5). Version 0.3.0, but the changelog post of 20 October says generally available and the docs changelog labels 0.3.0 the GA release, which we read as declared stable (15). 20+0+5+5+15 = 45.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-reliability):

Hosted APIs, MCP servers, models and platforms.

- 20, a public status page with component history (Statuspage, Instatus, BetterStack or the vendor's own).
- 0 to 30, the incident record for the last 90 days on that page. 30 for a clean record or trivial incidents only, 20 for minor incidents only, 10 for one major outage (an hour or more of a core API down, or errors across the board), 0 for several. 5 when there's no history we could read, and the note says so.
- 15, rate limits documented with numbers.
- 15, documented 429 or overload handling (Retry-After, backoff guidance), and idempotency keys or safe-retry guidance where writes are involved.
- 10, an SLA published for any paid tier.
- 10, the surface agents use is generally available, not beta or preview.

Local packages, SDKs, frameworks and stdio MCP servers.

- 20, installs from an official package with supported runtimes stated.
- 25, a public CI and test suite, passing on the default branch.
- 0 to 25, open crash or regression issues relative to activity (25 for few and handled, 0 for many, old and unanswered).
- 15, semver discipline and breaking changes called out in a changelog.
- 15, version 1.0 or later, or declared stable.

Protocols are read from their reference implementations, the public facilitators or servers, spec stability and test vectors.

## 2. Payments & pricing, 30 out of 100, up to 8.8 more on the total

Why it scored 30: Framework reading of the published rubric. No payment protocol (0). Per-million-token prices for Cursor models, including the default Grok 4.5 with Fast, and plan prices are public, but no price for managed hosting and no statement of which plans cover the Agent SDK were found (15). The package is free and the Free plan needs no card, but its usage is only described as limited agent requests (15). A person signs in through a browser or creates an API key in the dashboard before a model turn runs (0). 0+15+15+0 = 30.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-payments):

The published rubric, also on the [x402 page](https://www.anchorterminal.com/x402/).

- 40, a machine payment protocol (x402, MPP or L402) on the tool's own endpoints. 10 to 30 when it covers only some endpoints or only goes through a third party, and the note says which.
- 20, per-call or per-unit pricing published without a login. 10 for public plan-only pricing, 0 for "contact sales" or prices behind a login.
- 20, a free tier or trial that doesn't need a card.
- 20, autonomous onboarding, meaning an agent can get access without a person signing up in a browser (keyless use, x402, a programmatic key API).

Payment platforms and agent wallets rarely charge for their own API over a machine protocol, so the first line has steps for them, and the highest one that applies counts. 40 when x402, MPP or L402 runs on all their own endpoints, 30 when it runs on part of their own API, 25 when their merchants can accept one, 20 for running a facilitator, 15 for paying as a buyer, and 0 when the only protocol is their own. Merchant acceptance sits above a facilitator because the platform's own customers can charge agents through it, while a facilitator settles for sellers who wire up the protocol themselves. The counter-argument (a facilitator does more for the protocol as a whole) has a point. Each note says which step applied.

Open-source software you run yourself is scored on its hosted or paid option if it has one. A free, self-hosted package with nothing to buy gets 20, 20 and 20 for the last three lines, and 0 to 40 for the first only if it ships a payment protocol.

## 3. Security & auth, 68 out of 100, up to 5.6 more on the total

Why it scored 68: Framework reading. 30 for what leaves the machine by default, 20 for approvals and sandboxing, 15 for prompt injection, 15 for audit and 20 for the security programme. Every model turn goes through Cursor's backend with a Cursor credential, even on a self-hosted server, and the data-use page says Cursor may store and train on prompts and code with Privacy Mode off. Managed hosting stores the repository and agent data whatever the Privacy Mode setting. OpenTelemetry export is off until an endpoint is set, with prompt text and tool payloads behind separate flags, and the npm update check has a documented switch. Install copies skills into ~/.cursor/skills (18). `needsApproval` parks a tool call for a person, dry runs stub writes and the harness tools take an allowlist, but `local.sandbox` is off by default, so shell commands auto-approve with the host's environment and file tools reach any path the process can. Hosted deployments enforce an egress allowlist (12). The docs tell you to turn the sandbox on for webhook payloads, diffs and chat, offer HMAC and shared-secret auth on webhook routes, and warn that anyone who can talk to the agent can drive Cursor account connectors. The scaffolded memory journal shares turns across sessions, with an instruction to treat it as untrusted (10). Sessions keep every message and tool call, hooks can mirror them for audit, and `ctx.approvedBy` records the approver (14). security.txt and SECURITY.md with a disclosure address and Cursor's GitHub advisories, no bug bounty found, and no advisory for this package in OSV on 27 October (14). SOC 2 and ISO 27001 aren't scored on the framework reading. 18+12+10+14+14 = 68.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-security):

- 0 to 30, the credential model. 30 for OAuth 2.1 with scopes, or scoped and revocable keys with rotation. 20 for plain revocable API keys. 10 for one all-powerful key. 10 off when a secret can travel in a URL query string as a documented option.
- 0 to 20, read-only or least-privilege modes, and confirmation or approval for destructive actions.
- 0 to 15, prompt-injection posture where the tool returns untrusted content (documented mitigations or guidance). A tool that returns no untrusted content gets 10.
- 0 to 15, audit logs or per-call visibility for the operator.
- 0 to 20, a security programme. security.txt or a disclosure policy, a bug bounty, SOC 2 or ISO 27001, advisories handled in public.

Models are read for retention, whether API data trains models (and whether that's off by default), zero-retention options and certifications. Frameworks for telemetry defaults, approval hooks, guardrails and sandboxing.

## 4. Agent ergonomics, 80 out of 100, up to 3.3 more on the total

Why it scored 80: Framework reading, scored on what it takes to run a tool-calling agent with MCP. One file per tool and one file per MCP connection, `tools` and `disallowedTools` filters with `effects: "read"`, and an allowlist for the harness's built-in tools (22). Concurrency caps with queued turns, `--fields`, page tokens and limits on list commands, and a per-session cost view. We found no turn cap or spend cap for a run (12). Built-in routes return `{ ok: false, error }` codes, direct tool calls have documented 400, 404, 409 and 500 codes, and the CLI has seven exit codes (18). Reads and idempotent calls retry up to three times, writes run once, MCP readOnlyHint and idempotentHint decide retries for connection tools, and dry-run sessions stub writes (20). `agent-sdk init` scaffolds a working project and every agent field is optional, but the framework is TypeScript only (8). 22+12+18+20+8 = 80.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-ergonomics):

- 0 to 25, context cost. For MCP, the number and size of the tool definitions (25 for ten or fewer compact tools, 15 for 11 to 30, 5 for more than 30, plus up to 10 back for toolsets, dynamic loading or read-only subsets). For APIs, whether responses can be sized (field selection, limits, summaries).
- 20, pagination, filtering and output-size controls.
- 20, actionable, documented error responses, codes and messages an agent can recover from.
- 20, idempotency or safe retries, and for MCP the `readOnlyHint` and `destructiveHint` annotations.
- 15, sensible defaults, few required parameters, and official SDKs in at least two languages.

Models are read for tool use, structured output, prompt caching, context length, batch and SDKs. Frameworks for how much code and how many defaults a tool-calling agent with MCP needs.

## 5. Schema & documentation, 82 out of 100, up to 2.9 more on the total

Why it scored 82: Framework reading. Typed TypeScript declarations ship in the package, tools take Zod or JSON Schema, `agent-sdk help --json` describes every command's flags and exit codes, and `GET /v1/info` and `GET /v1/tools/:name` return the discovered surface and each tool's schemas. No OpenAPI document for the HTTP routes (20). cursor.com/llms.txt lists every Agent SDK page with a Markdown twin, and the package ships the docs and an llms.txt (10). The overview has a table of when to use the Agent SDK, Automations or Grok Bot, and the hooks page has a table of what to use instead of a hook for each job (18). Typed config, `effect` as an enum or predicate, `timeArgs` checked against the input schema, and slug patterns stated (13). Examples on every page, with error tables for direct tool calls, HTTP 401, 403 and 409, the limit error's JSON and seven exit codes (15). A dated changelog, but only two entries, both from 23 October 2026 (6). 20+10+18+13+15+6 = 82.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-schema):

APIs and MCP servers.

- 25, a machine-readable contract (a public OpenAPI file or similar; for MCP, typed JSON Schema inputs on every tool).
- 10, llms.txt or Markdown docs served for agents.
- 0 to 20, descriptions that say what a tool is for, when to use it and when not to, read from the tool definitions in the source or the API reference.
- 0 to 15, typed inputs with enums, constraints and required fields, and no free-form JSON blobs.
- 0 to 15, examples and documented error responses.
- 15, versioning and a public changelog.

Models are read from the API reference, the OpenAPI file, llms.txt, the structured-output and tool-use docs and the model cards. Frameworks from docs a model can follow, typed interfaces, examples and the API reference.

## 6. Transparency & trust, 70 out of 100, up to 2.6 more on the total

Made of editorial 49, provenance 91.

Why it scored 70: Closed source. LICENSE.md reads All rights reserved, with use under Cursor's Terms of Service of 3 September 2026 (15). The data-use page says Privacy Mode keeps Customer Data out of training with zero-retention agreements at model providers, and that with it off Cursor may store and train on code and prompts. The deployment page says managed hosting stores the repository and agent data the same way under Privacy Mode and refuses Privacy Mode (Legacy). No retention periods are stated, and subprocessors are on a trust portal (14). No deprecation policy. The early-alpha migration guide, the `persistent` removal and the semantic-search removal are documented with error messages but only the changelog entry carries a date (6). OpenTelemetry and the update check are documented with switches. What the Cursor backend logs from turns is covered only by the data-use page (14). 15+14+6+14 = 49.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-transparency):

- 0 to 30, source availability and licence clarity. 30 for open source under an OSI licence, 15 for closed with clear terms, 0 for unclear terms.
- 0 to 30, data handling and retention statements that agree with each other (privacy policy, DPA, retention periods, subprocessors).
- 0 to 20, a deprecation policy or notices with dates.
- 0 to 20, telemetry disclosed with an opt-out (local software), or subprocessors and data locations disclosed (hosted).

The other half of Transparency and trust is the provenance score, computed from checked facts (below). The category score is the mean of the two.

Provenance checks not met in full (half of this category, computed from checked facts):

- Terms of service: read, states 6 of the 7 things a reader expects, and has 3 clauses that cost points (3.1 of 10)
- Privacy policy: read, states 7 of the 8 things a reader expects (9.3 of 10)

## 7. Maintenance & community, 74 out of 100, up to 2.3 more on the total

Why it scored 74: 0.3.0 on npm on 25 October 2026 (30). Five @grok/agent-sdk versions from 22 to 25 October and 29 @cursor/bdk versions since 22 September (20). Closed product. A public changelog with two entries and a forum with no Agent SDK tag, so we couldn't see replies to SDK reports (8). The npm package is current and official, TypeScript only (12). Dependencies are current releases, but there's no public CI and the same build goes out under three names with a postinstall script (4). 30+20+8+12+4 = 74.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-maintenance):

- 0 to 30, time since the last release, or the last published model or API change for a closed service. 30 within 30 days, 20 within 90, 10 within 180, 0 older.
- 20, at least three releases or dated changelog entries in the last 90 days.
- 0 to 25, responsiveness. Issues and pull requests answered on GitHub (the open issues and how recent the replies are). For closed services, a public changelog and a support or community channel that answers, 0 to 15.
- 15, presence in the official MCP registry under a verified namespace (MCP servers), or current official SDKs (APIs and models).
- 10, package health, current dependencies and CI.

Models are read for deprecation notice periods and model churn rather than release counts.

## What we couldn't check

What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it.

- Possible deduction left to landing. The changelog post of 20 October 2026 tells readers to run `npx @cursor/agent-sdk init`, and the npm registry answered 404 for @cursor/agent-sdk on 27 October. The docs install @grok/agent-sdk, which didn't exist until 22 October. Because the @cursor scope is Cursor's own, nobody else can publish the name, so we recorded it as a documentation error and didn't deduct
- The GA date differs between sources. The changelog post is dated 20 October 2026 and says generally available today, the docs changelog labels 0.3.0 the GA release on 23 October, and npm published 0.3.0 on 25 October
- firstReleased is 2026-07-15, the first @cursor/july version on npm, which the migration guide names as the early-alpha package. Whether that alpha was open to every Cursor account wasn't established
- unchecked: status.cursor.com, whose robots.txt answered with a Cloudflare challenge, so no component or incident history for Agent SDK hosting was read
- unchecked: Cursor's GitHub security advisories page, since GitHub's web pages and API weren't reachable from this session. OSV shows no advisory for the three package names
- Which Cursor plans include the Agent SDK and managed hosting, and what hosting costs, weren't found on the pricing page or in the docs
- npm weekly downloads weren't recorded, because the downloads API returned a stale range (25 June to 1 July 2026)
- The Cursor SDK (@cursor/sdk, from 26 April 2026), the harness this package runs on, isn't listed and wasn't graded here

## Weaknesses

- Closed source under Cursor's terms, with no public repository, CI or issue tracker
- Every model turn runs through Cursor's backend with a Cursor credential, including on a self-hosted server
- Local sandboxing is off by default, so shell commands auto-approve with the host's credentials
- The changelog has two entries, 0.2.29 and 0.3.0, both dated 23 October 2026. The 27 earlier @cursor/bdk versions have none
- The launch post's install command names a package that doesn't exist on npm

## What costs an agent a turn today

The notes we give agents before they call it. Each one is a workaround an agent shouldn't need.

- Install `@grok/agent-sdk`, not `@cursor/agent-sdk` from the launch post. That name answers 404 on npm
- Set `local.sandbox` before a turn reads webhook payloads, pull request diffs or chat. Without it shell commands auto-approve with the host's credentials
- Give every tool an `effect`. A tool without one counts as a write, runs once and is stubbed in dry runs
- Start `agent-sdk serve` with `--bearer-token` on any shared host. `--allow-anonymous` admits every caller as one principal
- Delete bot/hooks/memory.ts when sessions serve different people. The journal shares each turn's message and result with later turns

## When it's done

Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.