Gumloop vs StackAI (2026)

Stands at F. Connectors are the primary configuration surface rather than an afterthought, with the vendor stating outright that what you connect determines what the agent can access, and the picker separating Gumloop-managed connectors from custom MCP servers the customer has added. That second tab is the important one for breadth: anything with an MCP server becomes a connector without waiting for Gumloop to build it, and the vendor's own framing is that an MCP integration can be built from a natural-language prompt. Tool calling is governed rather than open: a Tool Discovery toggle controls whether the agent finds tools itself, and an integration can be restricted to a subset of its tools, so breadth and restraint are configured separately. Connectors authenticate by API key as well as OAuth, which matters for internal systems that never had an OAuth app. Skills can call connected integrations directly from Python in the sandbox through the gumcp-client library, so integrations are reachable from code as well as from tool calls.

Stands at F on an enumerated catalogue of roughly ninety individually documented app nodes, each with its own actions, inputs and outputs, which makes this checkable rather than a headline number. Breadth across classes is emphatic and skews enterprise in a way that distinguishes it from the prosumer catalogues elsewhere in this lane: SAP, NetSuite, Oracle, Workday, ServiceNow, Snowflake, Databricks and Egnyte are not integrations a no-code tool aimed at individuals carries. Eight database and warehouse connectors with natural-language-to-SQL querying is a second distinguishing cluster. The MCP node means anything exposed as an MCP server becomes callable too, so the ceiling is open. Recorded per section 7: the Zapier and Make nodes are StackAI reaching out to those platforms, which counts as its own integration breadth here and is not evidence of StackAI's extensibility, graded separately on Ext.

Stands at F, and the distinguishing feature is that orchestration runs in BOTH DIRECTIONS between deterministic workflows and reasoning agents. Subagents let an agent delegate to other agents rather than doing everything in one conversation, which is documented multi-agent composition. In the other direction an Agent node drops into a workflow, so a deterministic pipeline can enrich data, hand judgment to an agent, and route on the result, and Loop Mode runs an agent across a batch with the vendor noting the sharp edge that each iteration starts a new conversation unless conversation IDs are tracked explicitly. That candour about a real footgun is a good sign. Underneath, the canvas provides branching, conditional logic, loops and subflows, and subflows are reusable components callable from any other flow and testable independently, which is the closest thing in this lane to a function abstraction.

Stands at F. Both halves are documented. Multi-step control flow is complete: AI Routing for model-decided branching, If/Else for deterministic branching, Loop Subflow for iteration, Delay for pacing, and Code and Python nodes as escape hatches. Multi-agent is genuine and works two ways, which is the part worth carrying: Subflow Tools let an AI Agent node call another flow AS A TOOL, so a supervising agent selects among specialist subflows at runtime, and the StackAI Project Node lets one project invoke another as a unit. A dedicated Orchestrating AI Agents guide sits in the tips tree, and Handling Errors and Fallback plus Skip and Replace Node give the reliability affordances that separate a production orchestrator from a demo canvas.

Stands at F and is among the strongest Trig cells reviewed, on one property nothing else in this lane has: THE AGENT CAN AUTHOR ITS OWN TRIGGERS. An AI Managed toggle lets an agent create, edit and manage its own triggers and schedules during a conversation, and a Create With AI path has the agent write a custom polling trigger class that queries each connected service for real tool names and schemas rather than guessing, validates the code in an isolated sandbox against live data, and records a baseline on first run so later polls detect only genuine changes. That covers combinations no pre-built trigger handles. Schedules are described in plain language and converted by an AI scheduler, so no cron is written. All three classes are otherwise well covered: schedules, per-integration events, and per-agent webhook URLs. Two operational details worth carrying: repeated trigger failures auto-disable the trigger with a notification, and Outlook triggers can fire relative to a calendar event, before, at start, or after it ends.

P>F. Channel coverage is the strongest part and is documented page by page rather than claimed: eight distinct end-user surfaces, which is wider than anything else reviewed in this lane. Events are covered by a dedicated Trigger node plus app-level triggers such as inbound email and inbound webhooks. Schedules are the thinnest of the three classes and the reason confidence is medium rather than high: recurring execution is referenced in two separate documentation pages as an established pattern, but no dedicated scheduling page appears in the complete index, so the configuration surface was not read. Fetching the Trigger Node page would settle it and is the single call that would take this to high confidence. Graded F because all three classes are documented as shipped rather than because the schedule surface is fully described.

P>F, the third correction on this record, and it rests on a product surface the record never mentioned: COMPANY BRAIN. Sources are attached whole or drilled into at file and folder level, kept synced, and searched automatically when a question touches internal knowledge, with the agent able to open a full document for more context. The chat surfaces this as Searching Company Brain, so retrieval is a visible first-class step rather than a prompt stuffing. Sync coverage keeps widening, with call recaps and transcripts added recently. That is a maintained retrieval structure with a lifecycle, the same bar that took zapier's knowledge sources to F this session. WORTH RECORDING FOR THE LANE: Brain appears in the product navigation beside Skills, and on last turn I flagged it as possibly the memory feature. It is not. Brain is the KNOWLEDGE surface, which means the Mem grade on this record rests on Skills alone and not on Brain, and the two cells are not double-counting one fact.

Stands at F. Knowledge bases are durable managed artifacts with their own creation, usage, node and feature documentation, plus a REST endpoint and, importantly, their own PERMISSION model, which is rare: access to a knowledge base is restricted independently of access to the workflow that uses it. For a regulated buyer that is the control that makes a shared agent platform viable across departments. Relevant to the open Knowledge convention: this is decisively a maintained retrieval structure rather than per-request assembly, and the vendor documents two distinct ways an agent consumes it. A Dynamic Vector Store covers the runtime-constructed case, and Search Connected Apps grounds on live systems without indexing them first, so both patterns are available and separable. Customer-managed vector stores through Pinecone and Weaviate mean the index can live outside the platform entirely.

P>F, and this needs care because THE VENDOR ITSELF DISCLAIMS THE WORD MEMORY. The docs say twice that skills can feel like memory but technically are not remembering your conversation, and that the agent must reload the skill in future conversations. That disclaimer is about conversation transcripts, and it is quoted in the basis so nobody reading this cell thinks I confused the two. What the axis measures is whether state accumulates and whether the accumulation survives, and on that test skills pass clearly. The agent edits and creates its own skills, enabled by default, so a correction in one conversation is written back and applied in the next. The vendor documents the trajectory explicitly as a compound effect running from a basic template on day one to a full playbook by month six, with file-level version history showing which agent made each change. That is durable learned state surviving across sessions, and it is more than mindstudio's cross-run change detection, which I graded P precisely because nothing learned from prior runs. CONFIDENCE MEDIUM for two honest reasons. The vendor's own framing cuts against the grade even though the capability meets the bar, so a later re-level could reasonably disagree. And the product navigation shows a BRAIN surface alongside Skills which I did not verify; if Brain is conversational memory it would only strengthen this cell, but it is not relied on here.

Stands at P, and it is now the ONLY cell below F on this record, so the reasoning matters. The Shared Memory node is explicitly a windowed context-passing mechanism, not a memory store: the builder chooses how many past interactions to forward and the documentation warns that passing too many will overwhelm the receiving model's context window. That is the same token-buffer shape as dify, and it is bounded by the context window rather than by a persistence layer. Nothing in the complete documentation index describes memory that survives a session, accumulates over time, or is learned from prior runs, which is the F bar met by cosine, greptile and cognition. Two things were deliberately NOT counted here, to avoid one fact doing work on three axes: the browser sandbox persists cookies and session state across runs, which is graded on Comp, and Canvas keeps per-conversation version history, which is a document feature. The Dynamic Vector Store is a retrieval structure and is graded on Know.

Stands at F, and the June re-grade that put it here is confirmed on wider first-party evidence than the changelog entry it originally cited. Three distinct mechanisms. A runtime approval queue: agents raise pending approval and input requests, and a recent change widened who can reject them from only the named recipient to anyone who can chat with the agent, which is a small detail that only exists once approvals are genuinely load-bearing. A policy gate: enterprise controls require human approval before EXPENSIVE actions, with per-agent, per-team and org-wide budgets, which is an unusual axis to gate on and complements permission-based gating rather than duplicating it. And a permission gate: per-tool authorization, restriction of which tools an integration may expose, and connectors blocked by organization policy showing as restricted with request-access on the spot. Skill creation itself can be restricted by custom role. Together these cover approve, constrain and escalate, which is the full shape this axis looks for.

Stands at F. Human in the Loop is a documented first-class node rather than a posture, and the vendor treats it as foundational enough to build the first of its five learning challenges around it, which is a good signal that it is a load-bearing feature rather than a checkbox. The approval surface is the vendor's own and is delivered where the approver already is: the product page shows an approval prompt in Slack with explicit approve and disapprove controls. That clears the 30 August ruling, since Slack is the delivery channel while the gate itself belongs to StackAI. Workflow Notifications and the production workflow lock in Project Controls are separable second and third oversight mechanisms, the latter unusual because it constrains what a builder can change rather than what an agent can do.

Stands at F and is a strong one. The attestation half is met with SOC 2 Type II and GDPR and a trust centre. The control half is unusually deep because Gumstack was built for exactly this: role- and attribute-based access, PER-TOOL authorization rather than per-integration, every call tied to a specific user, agent or service principal, and central secret management with the vendor making a point that keys never sit in plaintext on a laptop. A LIVE INVENTORY OF EVERY MCP SERVER IN USE is the rarest item here and the one worth carrying to comparison pages, because shadow MCP servers are the current version of shadow IT and almost nothing else in this lane claims to enumerate them. Governance extends past security into spend, with per-agent, per-team and org-wide budgets and quota controls. Identity integrates with the customer's own provider, extending their system of record to agents. Recorded per the 30 August ruling: VPC deployment is a delivery property and is carried on Dep, not credited here.

Stands at F and is among the two or three strongest Sec cells in the index. The conjunction is met several times over, and unusually the control surface is documented page by page rather than asserted as a bullet list: RBAC, workspace and folder isolation, feature-level admin enforcement, per-connection and per-knowledge-base permissions, MFA, and full SCIM provisioning with separate Okta and Entra guides. Encrypted environment variables resolved per stage is a maturity signal most no-code platforms lack. Graded strictly on the control surface: on-premise deployment and government cloud are carried on Dep and deliberately excluded here, per the 30 August ruling that sovereign delivery never reaches Security. Disclosure quality is high by index standards, with a trust centre, a SOC 2 report request route, model-provider DPAs published as signed PDFs and a BAA route, though no audit firm is named.

Stands at F and the scope is wider than the axis usually sees. Gumstack traces every tool call through a single logging and analytics layer, ties each call to a specific user, agent or service principal, and captures audit trails showing where data is flowing. That is the why rather than the what. Two properties are genuinely distinctive. It reaches AGENTS OUTSIDE GUMLOOP, giving one audit surface across third-party AI tools as well as the vendor's own, which is a materially different proposition from logging your own runs. And cost is a first-class dimension alongside actions, with real-time organization-wide credit tracking and per-agent and per-team attribution, so the question who spent what is answerable rather than inferred. Agents can pull credit usage and run history reports themselves through data export tools. Skill-level observability sits underneath, with file-level version history naming which agent made each edit and an activity timeline of views, runs and edits.

Stands at F but at MEDIUM confidence, and the gap is named rather than glossed. A dedicated Observability section with three pages, a REST Analytics endpoint, audit logs on the product page and version history through the development lifecycle is comfortably more than reporting, and the Evaluator adds a quality dimension most vendors have nothing equivalent to. What was NOT confirmed is the per-run trace: the Manager page was identified from the complete documentation index but not read, so whether a customer can reconstruct why a specific run took the path it did is inferred rather than verified. One detail cuts slightly against F and is recorded honestly: a guide titled Adding Advanced Logging for Analytics suggests some richer telemetry is assembled by the builder rather than supplied. Fetching the Manager and Analytics pages would settle this to high confidence either way, and it is the main outstanding item on this record.

Stands at F, and residency is answered properly rather than gestured at. The vendor offers managed SaaS or deployment into the customer's own cloud as a VPC IN THE REGION OF THEIR CHOICE, with the data staying in infrastructure the customer controls, and states that for private infrastructure it deploys and operates the full stack inside that cloud. Region selection plus VPC is exactly what the section 7 rule names as residency rather than legal cover, and the operated-by-vendor framing means the customer is not taking on the upgrade and patching burden that self-hosting normally implies, which is the trade-off against n8n's open-source route. Gumstack can likewise run on Gumloop infrastructure or inside the customer's VPC, so the governance layer does not force data back out. A second fact reinforces it: inference for open-weight models runs through US-based providers under Zero Data Retention, so the model layer does not quietly relocate data either. Useful contrast for comparison pages: this is managed-cloud plus VPC, not open-source self-hosting.

Stands at F, and the strongest single piece of evidence is not a claim but an artifact: StackAI ships a dedicated enterprise CLI whose entire purpose is deploying and managing the platform on customer infrastructure, with its own documentation set, engineering standards, Kubernetes and Docker migration guides and a CVE upgrade runbook. A vendor that maintains a deployment CLI with release engineering docs is not offering on-premise as a sales concession. Government deployment is a separate documented path including Azure OpenAI in Azure Government, which is a genuine sovereignty story rather than legal cover. Per the 30 August ruling all of this lives here and only here; the Security cell was written on the control surface alone. Local LLM hosting reinforces it, since an air-gapped-leaning customer can keep both the platform and the inference inside its own boundary.

Stands at F and is a clean case under the 31 August wider bar, adopted by several routes at once. A template library ships prebuilt flows, and the vendor's own onboarding advice is to start from a template rather than a blank canvas. Subflows are reusable components shareable across a team, so an organisation accumulates its own adoptable assets. The strongest route is the skills library: skills are packaged folders with instructions, scripts, references and assets, downloadable and uploadable as files, shareable with four distinct roles including a USE ONLY role that lets a colleague run a skill without seeing its contents, which is a genuinely thoughtful packaging primitive for proprietary process. Admins can maintain an organisation-wide skill library kept in sync with a GitHub repository, and the vendor points users at public example skill repositories, noting compatibility with the Claude Skills packaging convention. Two built-in system skills ship with every team.

Stands at F and the basis is now specific where the April one named nothing. Three separable layers of prebuilt material, which is more than most: a browsable Templates catalogue reachable from both the product and the documentation, named agent templates such as Content Writer that the documentation directs users to open and adapt, and Skills, which are reusable instruction packs agents load on demand and which function as prebuilt behaviour rather than prebuilt structure. A Prompt Library and a Common Architectures page add reusable material at the prompt and pattern level. Clears the browsable-catalogue bar carried from Coding agent. Recorded honestly: the catalogue page itself was not enumerated this pass, so its depth is unmeasured, though the vendor names specific templates by name in the documentation which is stronger than a bare claim of a library.

Stands at F. Model choice is a per-agent setting at the top of configuration, with three presets for people who do not want to choose and the full roster for those who do, and the vendor commits to adding new models within a day of public release. Two details are unusual enough to carry. First, TOOL CALLING IS THE SELECTION CRITERION: every model in the agent picker supports tool calling, and models that do not are confined to workflow AI nodes, so the picker cannot be used to pick a model that will silently fail as an agent. Several open-weight models run the other way and are agent-exclusive. Second, open-source models are routed through US-based providers under Zero Data Retention with a US-provider hosted badge in the picker, and the vendor states data is neither used for training nor stored after inference, which turns model choice into a decision the buyer can see at selection time. Enterprise governance adds control over which models teams may use, so this is also an administrative surface.

Stands at F, and the distinguishing feature is not the breadth of model choice but the GOVERNANCE over it, which nothing else reviewed in this lane ships. LLM Provider Governance lets an administrator control which models the organisation may use and where information is sent and stored, so model flexibility is bounded by policy rather than left to whoever builds the workflow. For a regulated buyer that is the difference between model choice being an asset and being a compliance risk. Underneath it the ordinary requirements are met: multiple providers, per-node selection, models hosted in the customer's own Azure or Bedrock account, and locally hosted models. Not to be confused with the Ext credit for publishing an MCP server, which is the opposite direction of travel and is graded there.

Stands at F but at medium confidence, and the gap is named rather than papered over. What is solidly documented is inbound and protocol extensibility: each agent gets a unique webhook URL so any external system can invoke it with a payload, custom MCP servers are added as connectors, enterprise entitlements include MCP server hosting, and Gumstack maintains a live inventory of MCP servers in use. Skills are portable artifacts, downloadable and uploadable as .zip or .skill packages, with organization libraries syncable from a GitHub repository, and a code sandbox runs Python and JavaScript with an integration client library. WHAT I DID NOT VERIFY is a documented public REST API or SDK for driving the platform programmatically, which is the primary route under Mike's 30 August bar; the webhook surface carries the grade instead, with MCP as corroboration exactly as that ruling specifies. A docs page for a Gumloop API would take this to high confidence, and its absence would be worth knowing too.

Stands at F and clears Mike's 30 August Ext bar on every available route rather than just one. A documented REST API with named endpoints makes the platform callable from outside, which is the bar itself. StackAI also PUBLISHES ITS OWN MCP SERVER, which under the section 7 axis rule credits Ext and refuses Model, because the customer picks the assistant that reads StackAI. The enterprise CLI is a third surface, and project export and import means an agent definition is a portable artifact rather than something locked in a tenant. The Custom API node is the inbound counterpart, letting a workflow call anything the catalogue does not cover. This is a wider extensibility surface than most no-code platforms in this lane offer, and it is the reason the platform can sit underneath another product rather than only in front of a user.

Stands at P. The iteration surface is real and in places better than several vendors graded higher: skills validate on save with warnings for malformed frontmatter and oversized instructions, a Local Changes diff view shows unsaved edits across files before committing, an Edits tab gives file-level version history with rollback, subflows can be tested independently before composition, and AI-created triggers are validated in an isolated sandbox against live data before being saved. The vendor is also unusually frank about failure modes, documenting concurrent-edit races and the fact that skill changes are lost if a conversation ends abnormally, which is the kind of thing that helps a builder debug. What is absent is measurement. No evaluation product, no scoring, no test-case generation, no regression suite, and nothing that compares one agent version against another, which is the F bar that n8n, lyzr and stackai meet. Gumloop is strong on validation and iteration and silent on evaluation, and the P should be read that way rather than as a weak debugging story.

P>F. The Evaluator is a shipped customer-facing evaluation product, not the vendor testing its own work, which is the distinction the axis turns on. LLM-as-a-judge scoring of the customer's own agents clears the bar set by goose and openhands in Coding agent and matches the shape that took lyzr to F earlier in this lane. The surrounding surface is stronger than the Evaluator alone: an Agentic Development Lifecycle with version control, pull requests and prompt diffs, Project Controls that track versions during development and LOCK a workflow once in production, and documented troubleshooting and error-fallback guides. The production lock is the detail worth carrying, since it is a governance control on change rather than a testing feature and few vendors in this lane ship one. Recorded honestly: the Evaluator page itself was identified from the complete documentation index and its description rather than fetched in full, so the scoring mechanism's depth is not measured.

N>F, and this is the clearest positive case for the axis after stackai. The vendor's own extension listing states the purpose in the exact terms this axis measures: automate tasks on websites that DON'T OFFER API ACCESS, going beyond traditional API limitations, and if you can do it in your browser you can automate it. Record-and-replay of real clicks, typing and form entry is the deterministic replay mode credited on stackai's browser navigation, and a Web Replay is referenceable from an ordinary Gumloop automation, so browser operation composes into workflows rather than sitting apart as a toy. WHY THIS IS NOT THE MINDSTUDIO REFUSAL. I refused mindstudio's Chrome extension because it supplies page content to the agent, so the agent reads what a human is looking at and never drives anything. Gumloop's records and replays the user's own actions on the page and fills forms. One observes, the other operates, and that is the whole distinction. CONFIDENCE MEDIUM AND THE REASONS ARE RECORDED. The strongest evidence is a store listing rather than a documentation page, the listing calls itself an early release, and a vendor forum thread indicates the related web automation node is access-gated, so the tier at which this is purchasable was not established. Under the 31 August tier line it is graded because it is purchasable at some tier, with the gating noted. The docs page for the web automation node would take this to high confidence.

N>F, a full point, and the most consequential single correction of this lane so far. THE REASONING IS THE THIRTEEN JUNE COMP ERRORS RUN IN REVERSE, so this record is worth keeping as the worked example of the distinction. The StackAI Computer provider ships three actions and only one of them is Comp. THE TERMINAL TOOL IS EXPLICITLY NOT CREDITED HERE: running shell commands in an isolated sandbox is code execution, which is exactly what blink-new, codebuff, compyle, cosine, cubic and eight others were wrongly graded F for. Canvas is a document workspace and is likewise refused. What earns F is browser navigation, and it earns it decisively: the agent drives a real browser through authenticated web applications whose session state persists between runs, which is the definition of operating software with no programmatic interface. Two modes exist, a deterministic replay of a recorded sequence and an agentic mode for unfamiliar or dynamic tasks, and a separate HyperBrowser node adds a third route. A live stream URL for watching execution and a step-by-step result payload are unusual and make the capability inspectable rather than opaque.