New the coordinator owns the card · you hear back with a result or one question →
Hermes.Crew · the smart kanban replacement for Hermes

You ask once.

/crew <ask> asks what is missing, up to five questions in one batch, writes the contract and hands the card to the coordinator. From there the coordinator owns it: every stop gets one decision. A fix, a smaller scope, a split, or one concrete question back to you. Done is a proof command that exited 0, run by a tool.

›/crew add Stripe checkout for the Pro plan, in test mode

Hermes · any profile · your chat turn ends when the card opens — what you hear back, and when →

NewSeven stations, you at both ends
6cards in parallel from one ask
100%of finished cards proven before done
24/7a coordinator owns every card to the end
Self-hostedruns inside your own Hermes

Built in, not promised: one ask fans out into up to six cards that run side by side, a card cannot be marked done until its proof passes, every run stops at its token budget, and the whole crew runs inside your own Hermes, on your own machine.

the launch film53 s · one ask, start to verified
crew · launch filma Pro plan launch, start to finish

One ask. Four deliverables. All verified. Built in parallel, each one proven before it is called done.

how a card travels

Seven stations. You stand at both ends.

Each station is a role with its own profile, model and tools. The two that decide whether work is done run no model at all. Why a kanban for AI agents needs a proof →

Youthe ask

Type it in the chat you already use. Say how you would know it is done.

your words · one message
Chatthe contract

Reads the ask as a skeptic, asks once for what is missing, opens the card. Then the turn ends.

your chat model · no polling
Coordinatorthe decisions

Wakes on every card event with the whole record in hand. One bounded turn, one decision.

strong model · 1 call per event · read-only
Workerthe work

One writer per card, inside the contract's scope. Stops itself after five failed tool calls in a row.

its role’s model · 60 steps
Proofthe exit code

The proof command runs in a tool, exactly as it was written when the card opened. 0 is the only pass.

no model · one verdict line
Auditthe second run

The coordinator re-runs the proof before the card can close. A card that asks for it gets one cheap independent verifier run instead.

no model, or one at 30 steps
Youthe ending

One line in the thread you asked from. Or one question, after the coordinator has tried twice.

at most 2 retries before you hear
one card, openedthe step panel: transcript · calls · details
crew · card · pricing pageworker · run 2 · glm-4.6
the card page: the pricing card's graph with its running worker selected, and the panel open on the worker's transcript

The panel reads like a transcript, not a log dump. What the worker is doing right now, what it said, which calls it made and the model it runs on. When a run stops, the same panel says why, and the coordinator’s fix sits on the card.

new in 0.10

Crew in Hermes Desktop.

Crew is now a native page in Hermes Desktop’s sidebar, built from the same app as the browser dashboard and drawn in the app’s own colours. It needs no separate server and no systemd. What the board guarantees, section by section →

crew · hermes desktopthe card page and the Crew flow pane
Crew in Hermes Desktop, dark: the Crew sidebar entry, a card page with its flow graph and button bar, and the Crew flow pane with its live-card selector

Setup is four short steps: Crew in Hermes Desktop, in the README.

01enforced bytool guard

One entry, and it ends when the card opens.

/crew runs in your chat and opens the card with the board's own tool, so the ending comes back to the thread that asked - no polling, no watcher. A card needs a full contract, and a script that writes board state or consent is refused before it exists.

ordinary messageno card, no questions
/crew, vagueone batch of questions, no research
/crew, completecontract shown, card opened, turn ends
answer turnopens the card within 30 minutes
02enforced bycoordinator

Every stop gets one decision.

The mechanical cases never reach a model: a held workspace, a dead model pin, a proof that now passes. Everything else gets one bounded turn: two retries, then one question to you. The decision turn runs -t file: it decides, does not do, and treats worker text as untrusted data.

  1. workerrun 1 · stopped · 5 failed tool calls in a row
  2. coordinatorretry · fix: verify the webhook against the raw body
  3. workerrun 2 · proof exit 0 · PASS by writer
  4. coordinatoraudit · proof re-run · PASS · done
  5. you
03enforced byverdict log

Nobody can talk a card into done.

A crew card closes only on a PASS line written by the tool that ran the proof command as it was snapshotted when the card opened. That holds for the writer, for the verifier, and for your own chat. Two FAILs go back to the coordinator with both outputs.

writer says donenot evidence
proof exit 0, by writerevidence
proof exit 0, by coordinatorcloses the card
edited proof linecannot pass
Verify: proof is the tool plus the audit. Verify: independent adds one cheap verifier run with one check the command cannot make. A proof command Hermes would flag waits for your own /crew-proof <card> yes — a run cannot approve its own proof.
04enforced byrole profile

Tokens go to work.

Every API call re-reads the whole context, so cost is calls times context, and Crew keeps both short. A run stops itself after five failed tool calls in a row. Every card carries a token budget, and a run that reaches it stops instead of burning on.

thrash stop5 failed calls → transient block
iterations60 per worker run · 30 per verify
token budgetper card · the run stops at it
first callprinted by install --check, under budget
Role profiles carry only the crew skills and the tools the role needs. A role cannot flip approvals.mode or write a crew profile’s config - that switch is yours, set once with /crew-safety.
05left toyou

Ask, look, stop.

Four commands cover the daily loop, and the cheatsheet below lists all eight. Unblocking, retrying and verifying are the coordinator's job, and the board shows what it decided: each card carries its verdict, its budget and its last decision.

/crew <ask>start a card /crew-statuswhat is running /crew-stop <id>stop a card /crew-diagnoseread-only look at a card
Served from your own machine on 127.0.0.1:8799. No external service, no telemetry leaving the box.
Hermes Agent logobuilt for Hermes Agent

Three asks to start with.

Paste one into the chat you already use. Each comes back verified, or with one question you answer in a line.

ask 1
›/crew add a pricing page with a monthly / annual toggle

✓ verifiedthe page and an e2e test

ask 2
›/crew write the launch post for v2, 600 to 900 words

✓ verifiedthe post, length and every link checked

ask 3
›/crew fix the failing CI on main

✓ verifieda green run, proven by the CI command

Install.
Give it an ask.

Install once per profile; running it twice changes nothing. Then type the ask in the chat you already use. The card, and its ending, come back to that thread.

›Set up Hermes.Crew for my Hermes profile NAME: run hermes plugins install macd2/Crew, then python3 ~/.hermes/plugins/crew/install.py --profile NAME and hermes -p NAME plugins doctor crew, and fix anything they report. Finish by sending /crew-status in my chat and tell me what it answered.

Star on GitHubFree and open source · Apache-2.0

cheatsheeteight commands, one line each
you will use
the edges

/crew and /crew-diagnose are skills that run in your own session; the other six are plugin commands. Stop, unstick, safety and proof are there so the coordinator never has to ask you to type something you did not want to type.

support

Where it runs.

Crew needs Hermes Agent 0.21.5 or newer with the kanban board, and Python 3.11+. The plugin declares requires_hermes: ">=0.21.5", so an older Hermes refuses to load it.

platform
terminal · browser dashboard on 127.0.0.1:8799
Hermes Desktop · Crew page in the sidebar
Linux
Full: the dashboard service and the nightly proofs.
Native Crew page, no daemon: Desktop’s own backend serves the data.
Windows, native
With Git for Windows (proofs run in Git Bash). The graph-http service and --nightly-proofs are skipped.
Works too: the Crew page needs no dashboard server, no systemd and no cron.
Windows, WSL
Both unchanged, as on Linux.
—

Terminal and Desktop read the same kanban board and session store, so a card looks the same in either, and you can use both at once. Windows details: WINDOWS.md.

commercial services

If the crew runs in your company.

The plugin is free and open source, Apache-2.0. If you use Hermes.Crew in an enterprise and would benefit from commercial support, extra tooling, or managed model spending, write to crew@forgecoreai.com and tell us what a card has to prove in your setup.