guilds.run vs AI SDK

guilds.run 21

No tagged release, version or dated changelog entry. Main, which the install clones and SECURITY.md names as the supported version, last changed on 3 November 2026 (1,072 of the commits since 14 October by one author), but commits with no version or tag aren't releases in this run (hyperqwen, brave-search-mcp), and CONTRIBUTING.md itself defines a release as a vMAJOR.MINOR.PATCH tag, so nothing has been released (0, from 20). No releases or dated changelog entries in the 90 days from 6 August (0). Issues and pull requests per the 14 October check (eight, all closed, the three issues within an hour), not re-read on 4 November (15 of 25). No published package (0). CI on every push with a frozen pnpm lockfile, and still no Dependabot configuration in the repository (6 of 10). 0 + 0 + 15 + 0 + 6 = 21.

AI SDK 97

7.0.163 on 2026-11-05 (30). 226 stable npm releases of ai across the 7.x, 6.x and 5.x lines in the 90 days to 6 November, 70 of them since 6 October (20). 47 of the 50 newest issues commented and 32 closed, with triage automation and backport workflows (22). ai and the provider packages are current, and 6.x and 5.x still get patches (15). Dependabot configured and CI passing on main (10). 30+20+22+15+10 = 97.

guilds.run 43

Read with the local-software lines, because the owner runs it. Docker with Compose is the stated requirement, and the install is a git clone of main plus ./setup.sh, which builds the images. No package, tagged release or published image exists, since the image workflow runs only on a v* tag and the remote still has none on 4 November 2026 (10 of 20, the earlier partial kept). The verify workflow runs scripts/check.sh on every push, with typecheck, lint, orchestrator, UI and web tests and a smoke test of the Compose stack, and the clone has 356 test files. GitHub's Actions pages and API answered 403 on 4 November, so the pass state on main is unread (15 of 25, down from 25 on the 14 October reading of 13 passing runs). Open issues per the 14 October check, not re-read on 4 November (18 of 25). CHANGELOG.md, added on 15 October, marks each change no impact, rebuild or breaking and CONTRIBUTING.md defines a semver tag policy, but no version has been cut (0, the earlier value kept). Not versioned or declared stable (0). 10 + 15 + 18 + 0 + 0 = 43.

AI SDK 93

Official npm packages, with ai 7.x requiring Node 22 or newer (20). A public CI workflow with lint, type checks, unit tests, Node-version tests, MCP tests on Windows and e2e tests runs on every push to main, and the newest completed CI runs on main passed on 6 November (25). 1,517 open issues and pull requests on 6 November from GitHub's API, with 1,059 issues opened in the 90 days to 3 November; 47 of the 50 newest had a comment and 32 were closed (18). A written versioning page with breaking changes only in majors, experimental_ prefixes for unstable APIs and a migration guide per major (15). 7.0.163 (15). 20+25+18+15+15 = 93, up from 73 on 6 October, when GitHub couldn't be read.

guilds.run 58

Framework reading. No OpenAPI document for the HTTP API. The recipe format guilds.recipe v2 rejects unknown keys, and connector manifests are validated with TypeBox at load (12 of 25). No llms.txt on docs.guilds.run (404 on 4 November). The docs are Markdown in the repository with an AGENTS.md (5 of 10). Pages state what the code does today, which page to read for which job, and the trust boundary in numbered rules (15 of 20). Tool input schemas, recipe caps and typed manifest fields (11 of 15). The getting-started walk says what you should see at each step, and the Operate page explains each startup error message (11 of 15). CHANGELOG.md (since 15 October) holds only an Unreleased section with no version or date, and the security register keeps dated history (4 of 15, the earlier partial kept). 12 + 5 + 15 + 11 + 11 + 4 = 58.

AI SDK 100

Typed TypeScript API with a reference section for every function (25). llms.txt and a Markdown copy of each page (10). The docs say when to use each piece, such as a table of WorkflowAgent against ToolLoopAgent, a when-to-use section for subagents and harnesses, and HTTP over stdio for MCP in production (20). Tools take Zod or JSON Schema inputs (15). 38 error types documented in the reference, with error-handling guides and examples throughout (15). Versioning page and a changesets CHANGELOG per package (15).

guilds.run 60

AGPL-3.0-only, OSI approved (30). No privacy policy or terms. The footer still links to both and they returned 404 on 4 November 2026. The docs say where data lives (PostgreSQL and STATE_ROOT on your host), that model keys go only to their provider, and that stored credentials are encrypted under one master key in .env (12 of 30). No deprecation policy or dated notices. CHANGELOG.md marks breaking changes with steps, undated, and the security register is dated (3 of 20, the earlier partial kept). Sentry is opt-in and documented on the Configure page, and the models.dev catalogue fetch and the connector registry address are documented with their switches (18 of 20). 30 + 12 + 3 + 18 = 63.

AI SDK 89

Apache 2.0 (30). The library stores nothing. The AI Gateway's ZDR page says it doesn't retain prompts or outputs and per-request ZDR is available on Pro and Enterprise, and Vercel's privacy notice is dated 1 June 2026, but we didn't read a DPA or subprocessor list (22). Deprecated APIs are removed only at majors with a migration guide, and 6.x and 5.x keep getting releases, without a written support window (14). Telemetry is opt-in and documented, with switches for inputs and outputs (20).

guilds.run 67

Framework reading, scored on what it takes to run a tool-calling team with MCP. No code is needed. Three commands and a model key start a guild, and each agent is offered only the tools turned on for it (18 of 25). max_tool_calls defaults to 50 per run, max_input_tokens trims old tool results and hands off to a new run, and remaining_tool_calls is appended to each result after the platform cap (15 of 20). Runs end with named codes such as generation.max_tool_calls and workspace.full, and interrupted tool calls get an error result so the next model request is valid (13 of 20). readOnlyHint and destructiveHint are stored for each MCP tool and writing tools start Off. Startup does not replay a model call (14 of 20). Sensible defaults and a guided first-run dialog, but it is an application with an HTTP API and no SDK in any language (7 of 15). Confirmed on 4 November in the clone (max_tool_calls default 50, 200 for a new coding agent).

AI SDK 95

An MCP client in a few lines with createMCPClient and client.tools(), and tool search with deferLoading keeps unused tool definitions out of the prompt (25). stopWhen with a default of 20 steps, and prepareStep to trim messages, switch models or limit active tools per step (20). 38 typed errors with isInstance guards, and MCP results with isError go back to the model (20). maxRetries on model calls, opt-in retries for transient MCP failures only, and WorkflowAgent for durable steps with automatic retries (20). Few required parameters, but TypeScript only, and a string model ID needs an AI Gateway key unless a default provider is set (10).

guilds.run 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. Model keys go only to their provider and Sentry starts only when SENTRY_DSN is set, with user info, bodies and model inputs and outputs excluded. Graft left the runtime image on 13 October (SEC-9), so its version lookup is gone, but the box fetches the models.dev catalogue on first read and hourly (switch MODELS_DEV_URL, documented) and the OAuth pages load Google Fonts (27). Each machine is an OpenSandbox container with dropped capabilities and no_new_privileges, writing MCP tools start Off and Studio changes wait for the operator. Agents on one machine share it and run commands as root with open egress (SEC-6, SEC-24), no per-call approval gate for tool calls was found, and the default AUTH_MODE compat-cookie has no login (10). The coding prompt says repository content is data, Studio wraps evidence as untrusted, and known secret values are redacted from transcripts. SEC-8, injection into coding agents, and SEC-47, GitHub content steering writes, are open (8). Every run keeps the prompt, each tool call and result, and versioned memory with who changed it (14). SECURITY.md with GitHub private reporting and no acknowledgement time, and a public register of 58 issues on 4 November with exposure and residual risk, 28 still open (18 open, 10 open and accepted). No security.txt (404), bounty or published advisory (9). 27 + 10 + 8 + 14 + 9 = 68. Sixteen register fixes dated 17 October to 3 November are on main with no release and no advisory; no term, because packages is empty and there is no hosted product, as ayrshare's plugin fix was read.

AI SDK 90

No telemetry until an OpenTelemetry integration is registered. String model IDs route to Vercel's AI Gateway by default, which is documented and needs a Vercel key, so it isn't silent (25). Per-tool approval, HMAC-signed approvals against forged client history, and Open Policy Agent rules in @ai-sdk/policy-opa, though sandboxes are experimental (20). ToolLoopAgent rejects system messages in prompts by default as an injection risk, URL fetches are checked against private addresses, and MCP HTTP transport refuses redirects by default, but there's no guardrail primitive (10). OpenTelemetry spans for every model and tool call to any backend, opt-in (15). HackerOne programme, a valid security.txt, SECURITY.md and two low-severity advisories published with fixed versions (20). Framework reading, so SOC 2 isn't scored.