Harness Hub
All startups

Prefect

Workflow orchestration, now stewarding FastMCP.

Prefect builds dataflow and workflow orchestration infrastructure, and in 2025 took stewardship of FastMCP — the most popular way to author MCP servers. A data-platform company moving squarely into the agent harness layer.

Pros & cons

Each point cites its sources and says what kind of evidence backs it. A second reviewer tried to refute every claim before it was kept.

Pros

  • DocsThe Prefect server and SDK are Apache-2.0 and can be self-hosted on SQLite for a single box or on PostgreSQL 14.9+ for production and HA.[1][2]
  • DocsThe Prefect team builds FastMCP (about 28k stars, Apache-2.0), a widely used Python framework for MCP servers and clients, plus Horizon, a commercial MCP gateway with SSO, tool-level RBAC and audit logs.[1]
  • DocsUnder its agreed acquisition of Dagster Labs, Dagster's FAQ commits to keeping the Dagster open-source project, leaving existing pricing and contracts unchanged, and not merging Dagster into Prefect.[1]
  • DocsThe free Hobby tier on Prefect Cloud includes 2 users, 5 deployments and 500 serverless minutes a month.[1]

Cons

  • DocsSSO, SCIM, RBAC, IP allowlisting and PrivateLink are Enterprise-only, and the self-hosted server lacks RBAC, audit logs and SSO entirely.[1][2]
  • DocsTeam costs $100 per user per month and still caps run retention at 14 days and audit logs at 24 hours; Hobby keeps runs for only 7 days and has no audit log.[1]
  • DocsThe post-acquisition roadmap is undefined: the Dagster FAQ commits only to "stability and continuity" and future notice about product direction.[1]
  • DocsPrefect itself is a workflow orchestrator; its agent features (FastMCP, Horizon) are separate products, not a full agent harness.[1][2]
Best for
Python teams that already orchestrate data or ML pipelines and want durable, observable execution for agent jobs plus MCP server tooling.
Watch out for
Governance features (SSO, RBAC, audit) need Prefect Cloud Enterprise, and the Dagster deal may reshape the roadmap.

Assessed Oct 29, 2026 · adversarially reviewedSuggest a correction

People

Open roles (5)

Their open-source work

Featured

PrefectHQ/fastmcp

The fast, Pythonic way to build MCP servers and clients.

Maintained by Prefect

v4.2.3released

Activity

Last push
· yesterday
Latest release
v4.2.3
· yesterday
Last 90 days
20+ releases
  1. v4.2.3v4.2.3: Trace Elements

    · yesterday

    FastMCP 4.2.3 adds an opt-in parent_task_spans setting for background tasks: turn it on and each task's worker span nests inside the tools/call request that submitted it, so the whole job reads as one trace (off by default, where the worker still starts its own trace linked…

  2. v4.2.2v4.2.2: Tea for Two (Too)

    · 3d ago

    FastMCP 4.2.2 fixes skills discovery for tool-only clients. Adding a server transform such as ResourcesAsTools made skills/list come back empty even though the skill:// files stayed readable; skills are now listed whenever their files are actually served. Thanks to our 2…

  3. v4.2.1v4.2.1: 4, 2, 1... Liftoff

    · 4d ago

    FastMCP 4.2.1 is a security and correctness patch on top of 4.2.0. It tightens OAuth and token handling: JWTs aren't accepted before their nbf, provider required_scopes survive a custom verifier, pre-registered OAuth clients can be pinned to their issuer, and consent, metadata…

  4. v4.2.0v4.2.0: Mad Skills

    · 8d ago

    FastMCP 4.2 serves agent skills from the server. A skill used to live in a folder on each machine, so every agent kept its own copy and an update meant asking everyone to reinstall. Now the server holds one copy next to the tools the skill describes, and a deploy updates every…

  5. v4.2.0a5v4.2.0a5: Skillium FancysonPre-release

    · 8d ago

    FastMCP 4.2.0a5 adds SkillManifestProvider, which serves skills from their manifests (frontmatter plus each file's digest and size) and reads a file only when a client reads it, so skills kept in a bucket or database don't have to be loaded to answer skills/list.…

Checked against GitHub on Nov 21, 2026. Release notes are shortened to plain text.

On X

No posts from them in our X feed yet.

Follow @prefectio on X · The X feed

Serverless cloud built for AI and agentic compute.

Modal lets teams run inference, fine-tuning and arbitrary agent workloads as serverless Python without managing infrastructure — a common substrate for harness backends at scale.

4 pros · 4 cons

Durable execution for long-running agent workflows.

Temporal's durable execution engine is widely adopted to make agent pipelines reliable: state survives crashes, retries are automatic, and workflows run for minutes or months. A key piece of serious harness infrastructure.

$430M total · $300M Series D (Feb 2026) at $5B valuation

3 pros · 4 cons

Durable functions and workflows for AI agents.

Inngest runs agent logic as durable steps — queueing, retries, concurrency control and human-in-the-loop waits — without operating your own workflow infrastructure.

4 pros · 5 cons

Connections

Relations from the Harness Hub knowledge graph, as charted in the State of Harness reports.

Maintains · 2

Built on · 1

The Harness Hub newsletter

New harness companies, projects, releases and the monthly State of Harness report, in your inbox.