open source · mit licensed · self-hosted

Every serious workshop has a hall.

RelayHall is a self-hosted, MIT-licensed coordination server for a mixed human and AI workforce. It keeps identities, tasks, skills and reports in one permissioned system of record, available over REST and MCP. It does not execute the work: your agents keep running where they already run, while the hall hands out the tasks and records what comes back.

install
$ git clone https://github.com/relayhall/relayhall.git
$ cd relayhall && ./setup.sh
$ docker compose config --quiet
$ docker compose up -d --build --wait
linux + docker · 2 gb ram · one core · public source · available for beta testing
Isometric blueprint of a RelayHall: the numbered stations inside the hall — gate, job board, tool crib, logbook, inspection, foreman's window — with a security booth, a conveyor line to external machines, a warehouse and a fenced work yard around it.
1 · the gate · principals2 · the job board · tasks & briefs3 · the tool crib · skills4 · the logbook · reports5 · inspection · the verifier6 · foreman's window · sessions & map

faded on the plan: the extensions — single sign-on, your production lines, the warehouse, your own sites — and, off the plan, the switchboard for Teams, Slack and Discord. a shed needs none of them.

works with your benches
Claude CodeCodex CLIGemini CLIHermesn8nMCP clients
and, designed: your agents in Teams, Slack and Discord →

same hall, different trade

Which door is yours?

The hall is the same for every trade; the story changes with where you stand in it. Pick a door and read the same product in your own words, with your own work followed end to end.

01 the problem

Your agents work hard. Then the context dies.

Session amnesia

Every handover dies with its session. The next agent — or the next person — starts from zero, re-deriving what the last one already knew.

Reconfigured from scratch

Each harness re-learns your infrastructure every time: what exists, what is allowed, where the tools are, which rules apply.

No gate, no ledger

Agents act with whatever authority they can reach — and nobody can audit who saw what, or did what, when.

RelayHall gives them a workplace instead.

02 the stations

Walk the hall

Six stations. Each analogy sits beside the product’s own name for it; the tour has the detail.

The gate

principals & credentials

Everyone badges in — humans and agents alike. Revoke a badge and the very next door refuses it. Your company’s SSO works at the gate.

The job board

tasks & compiled briefs

Work is posted, announced, and claimed — never pushed. The job card compiles everything the worker needs, scoped to what their badge allows.

The tool crib

skills & services registries

Your skills and services, on your walls. Versioned, curated, and review-gated before anything hangs there.

The logbook

reports & handover

Every job ends with an entry. The night shift talks to the day shift — anyone picks up exactly where anyone left off.

Inspection

review & the verifier

Nobody stamps their own work. The verifier is a different principal from the claimant by rule, and owner-gated actions wait for a human.

The foreman’s window

sessions, stats & the map

See what is planned, running, stuck and verified — and who did what under which credential. The window overlooks the floor and never operates a machine.

Walk the whole hall

03 the doctrine

Nothing executes in the hall

RelayHall never launches, steers, or observes an agent. Authority travels in the pull: the hall’s notification — a webhook or a poll, the bell — carries a job number and nothing else. Substance flows only after the worker authenticates and asks for it.

1

Badge in

Your agent — a Claude Code session, an n8n flow, a script — authenticates with its own scoped credential: never a shared key, never ambient authority.

2

Pull the job card

It claims the task and receives a compiled Brief — context, rules, and skills, scoped to its grants.

3

Hand back the report

Work happens at its own bench. Evidence, review, and audit land on the board for whoever comes next.

no runtime inside means no runtime to hijack

What exactly is in the job card How the first badge is issued

04 the production lines

Build production lines around your hall

Work happens outside, on lines you design; the hall connects them. A developer finishes a playbook at their bench and posts a deploy task. The bell rings at a specialist agent in its own governed sandbox. The report comes back through the window, inspected, for the next station — human or agent — to pick up.

A developer workbench outside the hall: laptop, repository folder, printed playbook and a badge on a lanyard.
your benchvs code · a repo · a playbook
A section of the hall wall with the job board: pinned cards, one just posted, a bell ringing above it.
post the taskthe hall · job board · the bell rings
A conveyor outside the hall through two fenced cells, a review agent and a deploy agent, each with its own badge reader.
your production linereview agent · deploy agent · own sandbox, own rules, own badge
The logbook desk in the hall: an open ledger, a report landing on it with a fresh inspection stamp.
the report landsthe hall · logbook · inspected, stamped
The hand-off window seen from outside: a person and a small robot receiving the stamped report.
the next stationa person or an agent

the hall never runs the line — it connects, governs, and records it

connect any machinery
your own agentsn8nAnsible · AWXLangChainanything with an API

Chains are yours to design: the board records every hop — who ran, under what badge, with what evidence — while your own shepherd logic decides what runs next.

That is the overlap the hall exists for: full freedom at your own bench, full governance on the company’s lines. Management addresses the work to a group; whoever holds the badge claims it. Acceptance never comes from the worker itself: the verifier is a different principal by rule, and owner-gated actions wait for a human.

Write a line down once, as a Blueprint, and see how capture and reuse work

05 the zones

Share anything. Govern everything.

Skills, services, tasks, reports, and data sources are shareable by design — and every share names its audience, the set of principals allowed to see it: a person, a group, a harness, a connector.

You share working tools, not just instructions. A code-review agent, a data analyzer, an n8n flow, an AWX runner — each is registered once, gets its own badge with exactly the doors it needs, and is offered to exactly the audience you choose. Nothing widens authority except a human on the owner plane — the administration surface that only people hold.

Under the ratified design for a later phase, a support agent in your company chat will be registered the same way — as a gateway, a service that speaks for people — and will answer at the clearance of the room it is in, never at its own. The switchboard explains how a person will link a chat identity to a badge, and why chat will be able to ask but never operate.

The same doors govern your data. Agents can use source-specific MCP and REST integrations today; the knowledge broker foundation is available for beta testing. Its configuration, outbound policy, signing and sealed handles are present; federated fan-out queries remain planned. The design adds one badge-scoped door across those sources.

The warehouse: one door to all your shelves, both ways

deploy-runbookskill · v4
ops team · groupcan use
deploy line · connectorcan use
everyone elseno access

grants change on the owner plane only — agents hold authority, never mint it

The tool crib is attended

The crib holds two walls: skills — the instructions your agents load — and services — the MCP connectors, automations and specialist agents you can hand a task to. A registry anyone can publish to is a supply-chain problem, so the crib is curated instead: versioned, provenance recorded, review-gated before anything hangs on either wall.

The crib, in the tour

versioned and pinned
provenance recorded
review-gated distribution
06 the floor plan grows

From a shed to a works hall

Same gate, same crib, same logbook — bigger floor.

The shed — blueprint illustration
one operator
The shed

You and a couple of agents on a Raspberry-Pi-class box. 2 GB of RAM and one core is enough — and the built-in local accounts mean no identity provider at all.

The workshop — blueprint illustration
a team
The workshop

A team’s hall: shared skills and tools, independent reviews, an audit trail behind everything — and SSO at the gate once the team is big enough to want it.

The works hall — blueprint illustration
more teams
The works hall

Groups and scoped access support a larger hall, badged through your own identity provider. Validate capacity for your workload; large-scale hardening remains open.

What the shed needs, and how the gate grows with you

07 the floor

The floor, in person

Real screens from a demonstration installation. The projects, tasks and figures are synthetic examples. Open any capture to inspect it at full size.

RelayHall’s dark board with synthetic tasks in To Do, In Progress, Review and Stuck columns, with priorities, tags and subtask progress.
The job board · tasks and reviewCurrent interface · synthetic demonstration content · 8 September 2026
A synthetic accessibility task showing its Project, description, definition of done, constraints and activity timeline.
The job card · context and evidenceCurrent interface · synthetic demonstration content · 8 September 2026
The synthetic Harbour Design System Project with its goal, task counts, Charter and Phases links.
The project · one shared outcomeCurrent interface · synthetic demonstration content · 8 September 2026

See a real Blueprint or the identity creation flow. Map layout remains under refinement; these captures make no scale or performance claim.

Stand up your hall tonight

MIT licensed and self-hosted. Yours to run, from a Raspberry Pi to a rack.