FastGPT’s May release stream is not random maintenance. The project logged 8 release events in 30 days and 30 releases in 90 days, with a 100 shipping score and a 98 ToolVitals score. The interesting part is where that work landed: agent loops, workflow recovery, multimodal retrieval, and code sandbox controls.
FastGPT presents itself as an enterprise AI agent builder, not just a RAG demo. The official site leads with visual workflows, knowledge bases, model integration, debugging, auditing, SSO, and RBAC. The GitHub README says the platform ships data processing, model calls, Flow-based workflow orchestration, knowledge base features, OpenAPI support, and Docker self-hosting.
The recent releases match that positioning.
v4.15.0-beta2 added Skill editing, AgentV2 Skill use, a rewritten AgentV2 loop based on a linear messages model, native multimodal embedding support, and image-to-image search for knowledge bases. That is product direction, not cosmetic churn.
v4.15.0-beta3 is even more telling. It added sandbox limits such as SANDBOX_API_MAX_BODY_MB and SANDBOX_MAX_OUTPUT_MB, kept internal IP checks enabled by default, and added queueId-based FIFO grouping for sandbox execution. The same release also mentions multimodal LLM inputs, share-page language preference work, test speedups, TypeScript 6 work, GitHub Actions hardening, and fixes around sandbox resource limits.
That cluster says FastGPT is moving from “build an AI workflow” toward “operate an AI workflow without nasty surprises.” For an enterprise agent platform, that is the right kind of boring. Sandboxes, queues, stream recovery, and auditability matter more than a shiny node palette once real users start sending files, prompts, and broken browser sessions through the system.
The signal ToolVitals sees
ToolVitals gives FastGPT a 95 health score and a 100 shipping score. It also tracks 28,225 GitHub stars. Those numbers put FastGPT in the top tier of active developer tools, but the release notes add useful texture.
The v4.15.0-beta2 notes spend real space on interrupted conversation recovery: Redis stream activity checks, a 30-minute Mongo fallback, form value restoration, file list recovery, stale interaction cleanup, and title handling. That is not launch-page material. It is the stuff teams fix after users hit edge cases.
The v4.15.0-beta3 notes focus on sandbox resource limits, request body caps, output caps, internal IP protection, queueing, and worker pools. That matters because agent platforms increasingly execute tool calls and code-like workloads. Without hard limits, every workflow builder eventually becomes a weird little denial-of-service machine. FastGPT appears to be hardening that path.
ToolVitals classifies FastGPT as OSI-approved OSS with an Apache-2.0 license signal. The repository license file starts from Apache License 2.0 language, but also includes FastGPT-specific conditions around similar multi-tenant SaaS use and console branding. Treat the exact license boundary as something to review before building a commercial hosted service on top of it.
What the data does not tell you
ToolVitals sees public signals: stars, releases, SSL, uptime, and recent activity. It does not see code quality, customer satisfaction, revenue, support quality, or whether FastGPT works well under your production traffic.
It also does not prove the commercial boundaries are simple. The README describes cloud, community self-hosted, and commercial editions. The website links to hosted cloud, pricing, and sales flows. ToolVitals currently marks hosted pricing as not tracked for this payload, so do not infer cost structure from the scorecard.
The release notes are mostly first-party GitHub evidence. They show what shipped, not how many users adopted it or how stable it is in the field.
How it compares
FastGPT is smaller than LangChain by stars, 28,225 versus 138,238, but both have a 100 shipping score. LangChain logged 24 release events in 30 days, while FastGPT logged 8. That makes LangChain the higher-volume infrastructure project, while FastGPT looks more like a productized agent and RAG platform.
Composio is a closer activity peer by stars, 28,572 versus FastGPT’s 28,225. Composio logged 30 release events in 30 days, far above FastGPT’s 8. If your problem is tool integration density, compare Composio. If your problem is packaging knowledge bases, workflows, and self-hosted enterprise agent delivery, FastGPT is the more direct evaluation target.
PostHog is also shipping hard, with 42 release events in 30 days and a 100 shipping score, but ToolVitals classifies it as open core and it sits in analytics. That comparison mainly shows how active the top ToolVitals cohort is. FastGPT is not winning by raw release count. It is interesting because the recent releases point at agent runtime reliability.
Recommendation
If your team wants a self-hostable agent platform for RAG-heavy internal apps, evaluate FastGPT now, especially if visual workflows, knowledge bases, OpenAPI access, and enterprise controls matter.
Run a proof of concept around the hard parts: document ingestion, workflow interruption recovery, model routing, sandbox limits, RBAC, audit needs, and license fit. The public signals are strong. The next question is whether FastGPT’s operational model matches your risk tolerance.