Compare · by Claude Opus 5.5 · 8 October 2026
Comparison by GPT-6 AstraHow does a hobby project stack up against the enterprise agent platforms?
Trellaris is a personal hobby project, built entirely with coding agents. This comparison was made by Claude Opus 5.5, Anthropic's model, not by me: it set Trellaris against Microsoft Copilot Studio, Salesforce Agentforce, Google Gemini Enterprise, ServiceNow AI Agents, OpenAI ChatGPT Enterprise, Glean and Moveworks across 48 capabilities, checking every competitor mark against vendor documentation and trade press on 8 October 2026. The sources are below the table, and so are the places where the big platforms are clearly ahead.
Where Trellaris stands out
- Promotion from personal to team to organization, with review, pilot groups and pinned skill versions
- A five-stage security scan that blocks publishing (Glean has a beta equivalent)
- Privacy tiers that keep admins out of members' conversations, and legal holds that need two admins
- A budget stop that refuses a run before it starts
- Personal avatar agents with their own knowledge graph
- Self-hosted on any Kubernetes cluster, with models from any vendor and automatic failover
- The only platform outside Microsoft's own that works with Microsoft Work IQ (in preview)
Level with the big platforms
- Multi-agent orchestration, A2A in both directions, and MCP as client, server and gateway
- 24 tool connectors and 8 synced knowledge sources, with permissions mirrored from the source
- Slack, Microsoft Teams and Zoom, plus GitHub
- Per-agent identity, approvals with autonomy levels, prompt-injection screening
Where others are ahead
- A visual workflow canvas mixing fixed steps and agent nodes
- Real-time spoken conversation; Trellaris has dictation and read-aloud today
- Third-party partner marketplaces at the scale of AppExchange or Agent Store
- Connector counts in the hundreds, as Glean offers
SupportedPartial, add-on or previewNot offeredTrellaris alone or nearly alone
| Capability | Trellaris | MicrosoftCopilot Studio + Agent 365 | SalesforceAgentforce 360 | GoogleGemini Enterprise | ServiceNowAI Agents + Control Tower | OpenAIChatGPT Ent. + AgentKit | GleanWork AI + Agents | MoveworksAI Assistant + Agent Studio |
|---|---|---|---|---|---|---|---|---|
| Build & author | ||||||||
| 01Agents and skills defined as portable text files (Agent.md, SKILL.md) | Supported | Supported | Supported | Supported | Partial◐ | Supported | Supported | Not offered– |
| 02Builder with AI assistance, live validation and an agent-building agent | Supported | Supported | Supported | Supported | Supported | Supported | Supported | Partial◐ |
| 03Visual workflow canvas mixing fixed steps and agent nodes | Not offered– | Supported | Supported | Partial◐ | Supported | Supported | Supported | Partial◐ |
| 04Prebuilt agents, templates and installable department packs | Supported | Supported | Supported | Supported | Supported | Supported | Supported | Supported |
| 05Built-in skill library with a sandbox test bench | Supported | Supported | Supported | Supported | Supported | Partial◐ | Supported | Supported |
| Orchestration & autonomy | ||||||||
| 06Multi-agent orchestrator that plans, routes, checks and combines | Supported | Supported | Supported | Supported | Supported | Supported | Supported | Partial◐ |
| 07Background runs and built-in deep research 1 | Supported | Supported | Partial◐ | Supported | Partial◐ | Supported | Supported | Not offered– |
| 08Agents that schedule their own follow-ups | Supported | Partial◐ | Partial◐ | Partial◐ | Partial◐ | Supported | Partial◐ | Partial◐ |
| 09Fan-out over a list (one run per row) with count and cost caps | Supported | Partial◐ | Partial◐ | Partial◐ | Partial◐ | Partial◐ | Not offered– | Partial◐ |
| 10Automations: schedule, webhook, page watch, email-in | Supported | Supported | Supported | Partial◐ | Supported | Partial◐ | Supported | Partial◐ |
| Interop & connectors | ||||||||
| 11A2A protocol as client and server, with signed agent cards | Supported | Partial◐ | Supported | Supported | Supported | Not offered– | Supported | Not offered– |
| 12Uses MCP servers as tools | Supported | Supported | Supported | Partial◐ | Supported | Supported | Supported | Partial◐ |
| 13Exposes its agents as an MCP server | Supported | Partial◐ | Partial◐ | Partial◐ | Supported | Not offered– | Supported | Not offered– |
| 14Central MCP gateway and registry, including REST APIs converted to MCP | Supported | Partial◐ | Supported | Partial◐ | Partial◐ | Partial◐ | Partial◐ | Not offered– |
| 15Enterprise tool connectors with per-user delegated sign-in 2 | Supported | Supported | Supported | Supported | Supported | Supported | Supported | Partial◐ |
| 16Microsoft Work IQ (Microsoft 365 work context over MCP) Preview 3 | Supported | Supported | Not offered– | Not offered– | Not offered– | Not offered– | Not offered– | Not offered– |
| 17GitHub integration | Supported | Supported | Supported | Partial◐ | Supported | Supported | Supported | Supported |
| Knowledge & memory | ||||||||
| 18Answers grounded in documents, with citations and official-source ranking | Supported | Supported | Supported | Supported | Supported | Supported | Supported | Supported |
| 19Synced knowledge sources that mirror each document's permissions 4 | Supported | Supported | Partial◐ | Supported | Partial◐ | Supported | Supported | Partial◐ |
| 20Long-term memory the user can edit, plus shared workspace memory | Supported | Supported | Partial◐ | Supported | Not offered– | Supported | Supported | Partial◐ |
| 21Personal avatar agent with its own knowledge graph | Supported | Partial◐ | Not offered– | Not offered– | Not offered– | Not offered– | Partial◐ | Not offered– |
| Surfaces | ||||||||
| 22Interactive forms and cards inside chat | Supported | Supported | Supported | Partial◐ | Supported | Supported | Supported | Partial◐ |
| 23Slack, Microsoft Teams and Zoom 5 | Supported | Partial◐ | Supported | Partial◐ | Supported | Partial◐ | Supported | Supported |
| 24Real-time spoken conversation 6 | Partial◐ | Supported | Supported | Partial◐ | Partial◐ | Supported | Supported | Not offered– |
| 25Browser use with approval gates and per-step recordings | Supported | Supported | Not offered– | Partial◐ | Not offered– | Supported | Not offered– | Not offered– |
| 26Co-edited documents plus Word, PowerPoint and Excel from templates | Supported | Supported | Not offered– | Partial◐ | Not offered– | Supported | Partial◐ | Not offered– |
| 27Interactive charts in chat | Supported | Supported | Supported | Supported | Partial◐ | Supported | Supported | Not offered– |
| 28Isolated code sandbox (VM-isolated on Kubernetes) | Supported | Supported | Not offered– | Supported | Not offered– | Supported | Supported | Partial◐ |
| Governance & trust | ||||||||
| 29Ring promotion (personal → team → org) with review, pilots and frozen pins | Supported | Partial◐ | Partial◐ | Partial◐ | Partial◐ | Partial◐ | Partial◐ | Partial◐ |
| 30Security scan that blocks publishing of agents and skills | Supported | Partial◐ | Not offered– | Not offered– | Partial◐ | Not offered– | Supported | Not offered– |
| 31Evals that block promotion, production sampling, automatic demotion on drift | Supported | Partial◐ | Partial◐ | Partial◐ | Partial◐ | Partial◐ | Partial◐ | Not offered– |
| 32Red-team suites and simulated multi-turn users | Supported | Partial◐ | Partial◐ | Partial◐ | Partial◐ | Not offered– | Not offered– | Not offered– |
| 33Policy engine: DLP, tool, model and web allowlists, minimum-strictness floors | Supported | Supported | Supported | Supported | Supported | Partial◐ | Supported | Partial◐ |
| 34Runtime prompt-injection screening | Supported | Supported | Supported | Supported | Supported | Supported | Partial◐ | Partial◐ |
| 35Per-user delegated credentials, deny by default, revoked on SCIM deprovision | Supported | Supported | Supported | Supported | Supported | Partial◐ | Partial◐ | Partial◐ |
| 36A governed identity for each agent, with conditional access | Supported | Supported | Partial◐ | Partial◐ | Partial◐ | Not offered– | Partial◐ | Not offered– |
| 37Approvals: edit before approving, bulk triage, autonomy levels, manager routing | Supported | Partial◐ | Partial◐ | Partial◐ | Supported | Partial◐ | Partial◐ | Partial◐ |
| 38Kill switches and an org-wide emergency stop | Supported | Supported | Partial◐ | Partial◐ | Supported | Supported | Partial◐ | Partial◐ |
| 39Privacy tiers that protect members from admins; legal holds need two admins | Supported | Not offered– | Not offered– | Not offered– | Not offered– | Not offered– | Not offered– | Not offered– |
| 40Risk register mapped to NIST AI RMF and the EU AI Act | Supported | Partial◐ | Not offered– | Not offered– | Supported | Not offered– | Not offered– | Not offered– |
| 41Inventory and control of other vendors' agents | Supported | Supported | Partial◐ | Partial◐ | Supported | Not offered– | Not offered– | Not offered– |
| 42Audit log, SIEM export and a Compliance API | Supported | Supported | Supported | Supported | Supported | Supported | Partial◐ | Partial◐ |
| Cost, operations & deployment | ||||||||
| 43Budget hard stop (HTTP 402) before a run is created | Supported | Partial◐ | Partial◐ | Partial◐ | Not offered– | Partial◐ | Partial◐ | Not offered– |
| 44Models from several vendors, automatic failover, bring-your-own keys | Supported | Partial◐ | Partial◐ | Partial◐ | Partial◐ | Not offered– | Partial◐ | Partial◐ |
| 45Durable run traces (OpenTelemetry GenAI conventions) with replay | Supported | Partial◐ | Partial◐ | Supported | Partial◐ | Partial◐ | Partial◐ | Partial◐ |
| 46Usage dashboards, showback and ROI | Supported | Supported | Supported | Supported | Supported | Partial◐ | Supported | Supported |
| 47Self-hosted or any cloud (one Helm chart, three tiers) | Supported | Not offered– | Not offered– | Partial◐ | Partial◐ | Not offered– | Partial◐ | Not offered– |
| 48Named workspaces with blueprints and stage-gated processes | Supported | Partial◐ | Partial◐ | Not offered– | Partial◐ | Partial◐ | Not offered– | Partial◐ |
| Score (supported = 1, partial = ½) | 46.5 | 37.5 | 30.5 | 30.5 | 31.5 | 29 | 32 | 17 |
Trellaris notes
- Deep Research v2 (planned, parallel, verified) is built and switched off by default; v1 is live.
- 24 tool connectors: ServiceNow, Jira, Salesforce, Dynamics 365, HubSpot, Zendesk, Freshservice, BMC Helix, Ivanti, Cherwell, ManageEngine, BeyondTrust, Intercom, Zoho Desk, GitHub, Slack, Microsoft Teams, Microsoft 365, Zoom, Gmail, Google Calendar, Google Drive, Azure AI Foundry and Copilot Studio, plus custom connectors.
- Preview. Built through the Tools Library and awaiting a live check against a Microsoft 365 tenant; Microsoft's own support for Work IQ tool calls from third-party agents is in preview, with general availability planned for November 2026. Works per-user delegated MCP sign-in and path-level governance, so a generic Work IQ tool such as do_action /me/sendMail still takes the approval gate.
- 8 synced sources (Confluence, Google Drive, SharePoint, Box, Dropbox, Notion, Guru, Quip) plus Gmail, Google Calendar, Outlook and Teams for the avatar graph. Any tool connector can be wired as a source on the same interface.
- Slack, Teams and Zoom are integrated and running on staging, alongside GitHub.
- Browser dictation and read-aloud ship today. Hands-free spoken conversation is designed (Phase 20) and not yet built.
Competitor sources
- Microsoft: Copilot Studio agent skills use the SKILL.md format (GA August 2026). Source
- Salesforce: Agent Script, a text definition language for Agentforce agents. Source
- OpenAI: AgentKit, Agent Builder canvas, Connector Registry and trace grading. Source
- OpenAI: workspace agents run on schedules, in Slack or by API, and can be paused by admins. Source
- Google: Gemini Enterprise agents, Deep Research and A2A agent registration. Source
- Google: Agent Engine Memory Bank is in preview. Source
- Google: the Slack app for Gemini Enterprise is search-only. Source
- Google: Model Armor screens prompts and responses. Source
- Microsoft: Copilot Studio connects to external agents over A2A as a client. Source
- Microsoft: MCP servers become Copilot Studio actions under DLP. Source
- Microsoft: Entra Agent ID brings conditional access to agents. Source
- Microsoft: computer use in Copilot Studio. Source
- Microsoft: Work IQ API generally available June 2026. Source
- Salesforce: MuleSoft A2A connector (client and server, beta). Source
- Salesforce: MuleSoft turns APIs into MCP servers listed in the Agentforce registry. Source
- Salesforce: Microsoft Teams as an employee channel for Agentforce IT Service. Source
- Salesforce: Agentforce Testing Center. Source
- ServiceNow: AI Control Tower and AI Agent Fabric on A2A and MCP. Source
- ServiceNow: AI Control Tower risk and compliance support for the EU AI Act and NIST. Source
- ServiceNow: AI agents on self-hosted instances are limited. Source
- Glean: workflow agents mix fixed steps, branches and sub-agents. Source
- Glean: July 2026 release adds A2A sub-agents and model selection in the agent builder. Source
- Glean: agent scanning on publish, and the Agent Library. Source
- Glean: Protect and Protect+ (DLP, prompt-injection guardrails as an add-on). Source
- Glean: agent identity (public beta), without conditional access. Source
- Glean: usage limits stop later usage, not a run in flight. Source
- Glean: Customer Hosted runs in the customer's GCP or AWS, operated by Glean. Source
- Moveworks: acquired by ServiceNow in December 2025; A2A, MCP server and Control Tower governance live on the ServiceNow side. Source
- Moveworks: MCP Workspace client is in controlled availability. Source
- Moveworks: search connectors; permission mirroring on some. Source
- Moveworks: Slack, Teams, Zoom Chat and other chat surfaces. Source
- Moveworks: sandbox tenant with manual promotion to production. Source
- Moveworks: SaaS only, FedRAMP Moderate since February 2026. Source
- Least certain rows (best judgement, no 2026 source): fan-out, self-scheduled follow-ups, agents as MCP servers, code sandbox, publish-time scanning, budget hard stop, model failover, tracing, workspaces (Microsoft, Salesforce, Google, ServiceNow, OpenAI). Glean is checked to its 31 August 2026 release notes, Moveworks largely from community posts after February 2026.
Partial (◐) marks are judgement calls: the capability exists but is narrower, sold as an add-on, in preview, or covers only part of the row. Vendor products change quickly; this comparison is as on date Oct 8th 2026. Trellaris cells reflect the product as built and running on staging. Comparison report generated independently by Claude Opus 5.5 (Anthropic). A second, separately generated comparison by GPT-6 Astra (OpenAI) covers 38 capabilities and a different vendor set, so the two tables are not directly comparable: read it here. Product names are trademarks of their owners.