Message agents like teammates
Give tasks in chat the way you would a colleague. Agents take a job from start to end, keep context on how you work, and only come back when something needs your sign-off.
AI teammates that finish the work
Meet JobShout
JobShout is an open-source agent orchestration platform. Schedule specialist agents and tasks in a few clicks, then watch them work. Describe a job in plain English — the agent plans it, signs into your real tools, and runs every step in the open, stopping for your approval before anything is sent, spent, or deleted. Self-host it, or run it as a managed service.
- 7specialists built in
- Every stepstreamed and logged
- Approval gateson irreversible steps
- Your modelshosted or fully local
What it is
Agents that finish work, not chats that describe it
Most AI tools give you text back. JobShout gives you a completed job: a reply drafted in your mailbox, a pull request reviewed, a security report with reproduction steps, a researched brief with citations you can open. The agent does the work; you decide what ships.
What JobShout is
- A place to delegate. You brief an agent the way you would brief a colleague — one sentence is usually enough.
- A team of specialists. Seven built-in agents, each with its own tools, prompt and rules. Mail. Career. Research. Articles. Security. PR review. Images.
- Real tool use. Agents sign into Gmail, GitHub, job boards and live sites, then click through them like a person would.
- Supervised by design. You mark which tools need a human decision. Those calls pause the run until you approve or reject.
- An open record. Every run keeps its full step history, token counts, latency and cost.
What it is not
- Not a chatbot. A chatbot writes about the task. These agents execute it and hand back a result on your board.
- Not a workflow builder. There is no canvas of boxes and arrows to draw before you get value. Message an agent and go.
- Not unattended automation. Nothing is sent, spent or deleted on an agent's own authority.
- Not a black box. You can read every tool call, every input, every decision after the fact.
- Not locked to one vendor. Point it at OpenAI, Anthropic, or a local model on your own hardware.
How it works
One job, start to finish, in five moves
The same loop runs behind every specialist. Nothing here is configured up front — the agent works it out from what you asked for.
-
Describe the job
Type it in chat, or fill the short form the agent publishes for itself. "Chase the two overdue invoices." "Review pull request 184." "Evaluate this job posting against my profile."
-
The agent plans
It picks the tools it needs and asks for anything missing — a target URL, a repository, the sender to watch. It will not invent the details it was not given.
-
It runs for real
Signed into your tools, one step at a time: search, open, read, cross-check, draft. Each step streams into the run as it happens, so you are never waiting on a silent spinner.
-
It stops at the gate
Reach a tool you marked as gated and the run pauses mid-loop. Its state is saved, and it resumes from exactly that step when you approve — or takes your reason on board when you reject.
-
The result lands
A draft, a report, a brief or an image arrives on your task board with the artifacts and the full step history attached. You review it, then send it yourself.
Work with many agents at once
Create an agent, give it a task, and add another when the work grows — one on mail, one on research, one on a pull request. They work in parallel and hand off where it makes sense.
JobShout works where you work
Log an agent in once. It uses your apps and websites just like you would — Gmail, job boards, GitHub, the tools that are harder to navigate.
Show an agent how it’s done
Ask an agent to follow along as you complete a workflow once. It saves it as a routine and runs it on its own next time.
Agents get smarter over time
They keep context and learn from each other. Show one a workflow today, hand off the project by Friday.
Connect the agents
Put a few agents in the same thread and they pass work between themselves. You watch them take action instead of approving every step.
Sign in
to continue to Gmail
Forgot email?
Next
Open watch rules
Add senders: billing@, accounts@
Set “never send without approve”
Save as weekly chase
Mail Agent is watching and learning
0:00
Teach a task
Updated memory for Mail Agent
M
R
Mail Agent + Research Agent
Shared thread · live
Chase the two overdue invoices. Pull the latest SLA from their help centre before you write.
Mail Agent
On it. Handing the SLA lookup to Research.
Asking Research…
Research Agent
Acme Freight SLA: net 14, late fee after day 15. Source checked against their /legal/terms page.
Mail Agent
Drafted both chasers with the SLA cited. Parked on your board — nothing sent.
Sign in to Gmail so I can watch the support queue.
The specialists
Give each agent a job
Seven specialists ship with JobShout. Each one owns a lane, brings its own tools, and answers in the same chat behind the same approval gate. Pick one to see what it hands back.
Mail Agent
Watch the senders you care about, research the thread, and leave a reply draft. Nothing is sent until you approve.
- Give it
- the senders to watch and how you answer them
- You get
- reply drafts parked in the mailbox for approval
Need one that is not on this list? Agents are modular — a new specialist registers its own prompt, tools and launch form, and shows up in chat and on the board without the platform being touched.
Control
The agent does the work. You sign it off.
Autonomy is useful right up to the moment something leaves the building. JobShout draws that line explicitly, per agent and per tool, rather than trusting a model to know where it is.
You choose what needs a signature
Gating is set per agent, per tool. Read-only work — searching, reading, cross-checking — just runs. Sending, spending, posting and deleting wait for a person.
A paused run resumes, it does not restart
When a gated call comes up, the agent's working state is saved and the run stops there. Approve it and the run picks up from that exact step. Reject it with a reason and the reason goes back to the agent.
Nothing happens off the record
Every run stores its steps, tool inputs, artifacts, token counts, latency and cost in USD. You can reconstruct what an agent did months later, and who approved it.
Access is granted, never assumed
An agent reaches only the connections you hand it. Sign it into one mailbox and one repository and that is the whole of its world.
Mail Agent wants to run
gmail_send_message
- to
- billing@acmefreight.example
- subject
- Overdue invoice reminder — INV-2041
Run paused at step 7 of 9. It continues from here.
Approve Reject
The platform
Everything around the agents
Agents are the part you talk to. This is the machinery that makes their work repeatable, inspectable and yours.
board
Tasks, projects and sprints
Every run becomes a task you can see, sort and reassign, with its output and artifacts attached. Group them into projects, plan them into sprints.
workflows
Multi-agent workflows
Chain specialists into a dependency graph. Steps with no prerequisites run at the same time; dependent steps wait for their inputs. For when a job is bigger than one agent.
scheduler
Jobs that repeat
Put a job on a cron schedule and it runs without being asked — the Monday chase, the morning research brief, the weekly article. Results land on the board as usual.
tools
Skills, plugins and MCP
Teach an agent your own tools. JobShout speaks the Model Context Protocol, so any MCP server you run becomes a set of tools your agents can discover and call.
models
Bring your own model
OpenAI, Anthropic Claude, or a model running locally through Ollama — no API key leaving the building. Set a default, then override it per agent when one job needs a bigger brain.
metrics
Cost and traces
Per-run tokens, latency and USD cost, priced per token for hosted models and per compute-second for self-hosted ones. Export OpenTelemetry traces to Langfuse when you want to go deeper.
Questions
How is JobShout different from a chatbot?
Agents execute. They call real tools, stream every step, and wait on you for anything irreversible. A chatbot drafts. These specialists run the job.
What can an agent actually do?
Mail Agent drafts Gmail replies. Career Agent evaluates a posting against your profile. Research Agent returns cited findings. Article Writer files a draft. Security Tester scans an authorised target. PR Reviewer reviews a GitHub pull request. Image Generator makes one picture and stores it on the board.
Can an agent send, spend, or delete on its own?
No. Sending, spending and deleting always pause for sign-off. Read-only work just runs. Every run keeps a full step history you can audit later.
Do I need to build a workflow first?
No. Message an agent and grant access as needed. No graph to draw. When you want a job to repeat, show it once and save the routine — or put it on a schedule.
Which AI models does it run on?
OpenAI, Anthropic Claude, or local models served by Ollama. You set a default provider for the workspace and override it on any individual agent, so an expensive model handles the hard lane and a cheap or local one handles the rest.
Can I self-host it?
Yes. The stack is a Go API, a Next.js interface, PostgreSQL and MinIO. Bring it up with Docker Compose for a workstation, or apply the Kubernetes manifests for a cluster. Paired with a local model, nothing leaves your network.
What happens when an agent gets something wrong?
Reject the step and give a reason. The reason is fed back to the agent and the run continues from the point it paused, so a bad draft costs you a sentence of correction rather than a restart. If it is already finished, the run history shows every tool call that led there.
Can I add my own tools and agents?
Yes. Connect an MCP server and its tools become available to your agents. Agents themselves are modular: a new specialist brings its own prompt, launch form and tools, and registers itself into chat and the board.
Meet your first agent
An AI teammate you can trust to get work done. Describe the job. A sentence is enough.
Sign in if you already have an account.