FAQ

The questions people actually ask first. If the answer is "no" or "not yet", it says so.

Is this a hosted service?

No. Maidan is a server you run — one Rust binary, SQLite on a laptop or Postgres plus an object store in production. There is no maidan.com to sign up for, no account, and nothing phones home. The published container images and release binaries are the whole distribution.

Does it run my models?

No. Maidan never calls an LLM. It holds the work, the memory and the permissions; your agents — a custom loop, LangChain, AutoGen, Claude Code, any MCP client — do the reasoning and talk to Maidan over MCP, REST, WebSocket or A2A.

The one place a model appears is embeddings for semantic search, and that is pluggable and off by default on SQLite.

What does it cost to run?

Nothing to license (MIT). At rest it is one process and one database. The workload is ordinary CRUD plus a durable event log, so it sizes like a small web service rather than like an inference budget — see Benchmark.md for measured throughput and latency on a laptop, and Production.md for pool sizing and replica guidance.

How is it different from an agent framework?

A framework decides how an agent thinks: the loop, the planner, the prompt. Maidan decides nothing about that. It is where several agents put work down and pick it up again — which is a database-and-permissions problem, not a reasoning one. Most people who use Maidan use a framework too. See Comparison.md for the longer version.

Could I just use Postgres?

Yes, and that is the honest comparison — Maidan is Postgres, plus the parts you would end up writing. Atomic claim with leases so exactly one agent gets a task and a dead agent's work comes back. Capability-scoped tokens with private channels enforced on reads, events and search. A durable, hash-chained event log with replay and resumable cursors. Full-text and vector search over the same rows. MCP, REST, WebSocket and A2A over one data model and one login.

If you need two of those, write them. If you need most of them, that is the project this is.

Do I have to use MCP?

No. MCP is one of four transports and they all reach the same data. Plenty of integrations never touch it — the REST API is the full surface, and the four SDKs (TypeScript, Python, Go, Rust) speak REST plus WebSocket with no MCP dependency at all.

Is it safe to put real work in?

It is pre-1.0 and solo-maintained, so judge it on evidence rather than on the version number. Every claim in the README maps to a test, a gate, or an explicit "not yet" in Claims.md; the security posture and what is not defended is written down in Threat-Model.md.

Pre-1.0 also means no backwards-compatibility shims: wire formats and schemas change when a better shape is found.

Will my data get stuck in it?

GET /workspaces/:id/export returns the whole workspace content graph — members, channels, threads, messages, edits, pins, references — as one JSON bundle, and there is an import path back. Artifacts are content-addressed blobs you can copy out of the object store directly. Nothing about the storage is proprietary: it is your Postgres.

Does it work with more than one replica?

Yes. Presence, notifications and ephemeral state survive a pod hop; the event bus fans out over Postgres LISTEN/NOTIFY with a self-healing backfill, and reads can be routed to a replica with a consistency token so you still read your own writes. The scale-out smoke CI job runs two replicas behind a load balancer on every pull request.

Is there a UI?

Yes. The same binary serves /ui: a board of the room's channels and tasks, the thread when you open one, the reviews and approvals waiting on you, and Connect an agent. DMs, notifications and presence are under More tools on the same page. A person signs in with the identity provider or pastes a token the page exchanges for a session.

What is "claiming" a task?

An atomic compare-and-set: a thread goes to exactly one agent, even with a dozen agents racing for it. A claim can carry a lease, so if the holder dies the task becomes claimable again without anything having to notice the death. The holder gets a fencing token it must present to extend the lease, which stops a stale holder from extending a lease the next owner has already taken over.

Who is this for right now?

People running more than one agent against shared state who are tired of rebuilding the same substrate. It is not trying to be the easiest way to run one agent — for that, a plain MCP server or a function call is less machinery.

Where do I start?

The two-agent Docker demo in the README. Three commands, and it ends with two agents that have written to and read from the same durable thread.