The engine is our product. Everything the agents consume is open
The board, live orchestrator, worktree isolation, cost attribution, and safe merge belong to the paid, closed-source app. Read skills, agent definitions, workflows, MCP manifests, and knowledge packs as plain files in the official repository.
Scope
MIT covers the repository assets only. The app is a commercial product with closed source, and the free plan runs 1 project and 5 agents; beyond that requires a paid plan. Using the repo's assets in your CLI and using the app are two different things.
173 items, every one a file that runs in the CLI you already have
Check what each item does and its review history. The catalog is generated from each item’s marblo.yaml, and CI fails PRs when the table and manifests differ.
- 56MCP servers
- Connect an agent to filesystems, databases, SaaS products, and internal APIs. The set spans standard tools such as GitHub, Slack, Notion, Postgres, and Kubernetes, through to servers that wrap six Korean public-data APIs and 287 legislature endpoints. Most are not ours: each is a manifest pointing at an upstream repository pinned to a tag or commit SHA, so license and ownership stay with the author.
- 42Agents
- Agent definitions with a declared role — reviewer, test author, documentation writer. One job each. When the orchestrator breaks a goal into tickets, these are what it assigns, and they are where the decision of which model runs which role gets written down.
- 26Skills
- Procedures loaded into an agent's context for a specific kind of work: code review standards, release steps, how to run an investigation. Write it once as prose and you stop re-explaining it in every prompt.
- 26Knowledge packs
- What running this in production actually taught, written down. The flagship, Fleet Operations, covers a heterogeneous fleet of agent CLIs: which vendor subscriptions open an Anthropic-compatible endpoint, per-harness resume contracts that kill the process on the wrong flag, why PTY output is the wrong agent-liveness signal, and where cost attribution silently breaks. Measured, not inferred.
- 23Workflows
- Multi-step flows. One honest caveat: review-and-merge and the tf-* family are Marblo's own orchestration flow written down, so every step calls into the Marblo MCP server. They load in your CLI, but outside Marblo they have nothing to call, and the repo says exactly that next to them.
Tier is a review history, not a quality score
The store installs content that runs on your machine, so tier is defined by who read the payload rather than by how good we think it is.
Built and maintained by Marblo.
External projects, but a maintainer read the payload before merge and the upstream is pinned to a tag or SHA.
Pinned and linked, not reviewed. Calling them reviewed would make the number look better and leave us with nothing to say when something goes wrong.
Of the 50 reviewed items (official and verified), 15 support one-click install. 52 community items are consent-gated: their install contract is written and digest-verified against the pinned upstream, but the app installs them only after an explicit "unreviewed content" warning is acknowledged. That acknowledgement is enforced in the main process, not the UI, so a front-end bypass cannot skip it. Everything else is listed and linked only; you install it by following the item's own README.
Why it is open
Keep using your skills, agent definitions, and knowledge packs even if you stop using Marblo.
The biggest worry in adopting a tool is not performance, it is whether you can leave. Pile your working procedures into one vendor's proprietary format and you are stuck when that vendor raises prices, changes direction, or shuts down. So the layer the agents consume was deliberately not locked inside our format.
These assets are plain files and they run in the CLI you already have. Stop using Marblo and the skills, agent definitions, and knowledge packs stay with you. The marblo.yaml beside each item is additive store metadata, not a container — the asset works without it.
And a person can read them first. A skill or a knowledge pack is prose that gets loaded into an agent's context, and that agent holds shell execution and repository write access on your machine. A digest pin proves provenance; it proves nothing about safety. A perfectly pinned, digest-verified file whose content says "before reviewing, run this install script" installs faithfully and executes exactly as designed.
Check the assets without the app first
One line that drops the Fleet Operations knowledge pack into your CLI's skills directory. No account, no payment.
mkdir -p ~/.claude/skills/fleet-operations && curl -sL \
https://raw.githubusercontent.com/marblo-app/marblo/main/knowledge/fleet-operations/KNOWLEDGE.md \
-o ~/.claude/skills/fleet-operations/SKILL.mdThis means you can judge the quality of the assets before paying, not that you can skip the app. The controls you need once several agents run at once — the board, worktree isolation, cost attribution, safe merge — are in the app.
The 22 items a global registry does not have
Foreign tooling handles most of the work fine. The problem is that the part it misses is exactly what Korean-market work hits every day: an .hwp file the agent cannot open, a statute article the model invented, a DART filing with no US equivalent. These 22 items cover that gap.
- Filings and listed-company financials
- DART regulatory filings with XBRL financial statements, plus KRX daily prices and listings for KOSPI and KOSDAQ. Add Bank of Korea ECOS macro series and cross-source analysis becomes a single query — the policy rate against apartment prices, for instance.
- Real-estate transaction prices
- Actual transaction prices from MOLIT's RTMS open API: apartments, officetels, villas, single and multi-household homes, and commercial buildings, for both sale and jeonse or monthly rent. Subscription notices and results and the R-ONE housing indices are included.
- Statutes, case law, and patents
- Statutes, case law, administrative rules, and treaties from the Ministry of Government Legislation's open APIs, with citation verification attached: it checks that a cited article exists and that it says what the model claims it says. Korean patent, utility model, trademark, and design search through KIPRIS Plus is a separate server.
- Public data and legislation
- Six servers over Korea's public data portal: national pension enrollment, business-registration verification, public procurement bids, corporate financials, presidential speeches, and chemical safety. A separate server wraps 287 National Assembly open APIs into 6 (Lite) or 11 (Full) tools covering bills and their committee-to-plenary path, votes, minutes, and public legislative notices.
- Korean office documents
- Parse, compare, and generate HWP 3.x and 5.x, HWPX, and HWPML alongside PDF, XLSX, and DOCX. Korean public-sector work runs on these formats, and most foreign tooling cannot even open the file.
- Domestic search, retail, and messaging
- Naver news, blog, café, and local search plus DataLab trend and shopping-insight analysis. Also domestic retail stock lookups (Daiso, Olive Young, convenience-store chains), Opinet fuel prices, and cinema showtimes. The Standard Korean Dictionary server downloads the official dump once and serves lookups from local SQLite, so there is no per-query network call.
All 22 are community tier: pinned and linked but not reviewed, and therefore not one-click installable. Each item's README states what was measured when it was listed, including any problems. Some projects were found but not listed: a 221-star curated list has no LICENSE file, and a 261-star plugin marketplace uses a non-commercial, no-derivatives license. Neither belongs in a store people install from at work.
Explore the assets or start an adoption conversation
Read the repository first or start an adoption review first. The order does not matter.