Agent roster
Each agent gets its own persona, model and reasoning effort. Conversations belong to the agent, so contexts never bleed.
One person can still divide work like a team.
Self-hostedIn betaMulti-agent · Multi-machine · OpenAI compatible
Agents scattered across machines,collected into oneconsole.
MatrixBot brings agent runtimes from your laptop, your LAN and Tailscale — plus any OpenAI-compatible model — onto one workbench.
Chat, assign, watch, accept. All of it in a single window.
The product is in closed beta and there is no public server yet, so we do not offer an online trial. If you want a look, or a beta seat, just get in touch.
- Self-hosted, data stays local
- Multi-machine, multi-runtime
- Tasks and acceptance
- Files return to chat
TaskRelease notes
Running 03:24Turn this week's product updates into one page of release notes, export a PDF, attach a desktop screenshot.
16:42
Read 12 filesRun commandBuild PDF
Plan · 4 steps2 / 4
- Collect 12 changes and screenshots
- Merge into 3 release themes
- Lay out and export the PDF
- Return files to this conversation
Writing /workspace/release-notes.pdf
Say something to Local OpenCode, or @ another agent
Interface schematic · rebuilt from MatrixBot v0.4.0
Local instance · LiveMatrixBot v0.4.0
Collected 2026-11-09 17:11 CST15
Agent roster
Own persona and model
10
Connections
Local / LAN / Tailscale
73
Task records
History and in-flight
12
Service entries
Entry points and usage
Numbers come from the running MatrixBot instance API, not from marketing copy. Task count includes history.
In 30 secondsOne entry point,
One entry point,
every agent under control
No manual needed. Thirty seconds on how it pulls several machines, several agents and real services into one window.
- One entry point that wires scattered machines into a single net
- Laptop, server, phone — connect once, stay reachable
- One sentence to assign work, several agents start at once
- Progress, logs and results live in the same window
- It remembers what you told it
- Data, models and keys stay on your own machines
30s · 1080 × 1920 · Chinese subtitles · muted by default
30s
Tap to play · sound recommended
Why it existsThe more agents you run,
The more agents you run,
the more you need a console
MatrixBot does not replace your agents. It gathers connections, tasks and runtime state in one place, so assigning work and accepting it become one continuous line again.
Agents do not lack ability.
They lack one shared workbench.
01
Scattered entry points
A page on this machine, a terminal in the office, another window for the model endpoint. Every switch costs context, and you rebuild the scene from memory.
02
Blind process
Which step a long task reached, what tools it called, whether a dangerous command is waiting for approval — you can only guess at the screen. Step away for ten minutes and the story is gone.
03
Broken delivery
Reports, screenshots and audio sit on a remote machine. To deliver, you ssh over, haul the files back, and then explain which copy is final.
Core capabilitiesEverything is here,
Everything is here,
nothing to stitch together
From connection to delivery, eight things close the loop inside one system. Each one maps to a feature that really runs, not a concept drawing.
08 capabilities
Connections and runtimes
Hermes, OpenCode, dsh, ACP and any OpenAI-compatible endpoint, managed in the same place.
Local, LAN and Tailscale treated alike.
Task engine and acceptance
Goals and acceptance criteria come first. Every task gets a Markdown plan, with subtask dispatch and automatic roll-up.
Leave a long task running; it is still there when you return.
Service catalog
Registers entry points, summaries, auth methods and limits as a single source of truth agents can read directly.
Fewer guessed ports, fewer stale memories.
File and media delivery
Move any file in both directions, remote machines included. Images, audio and video render right in the conversation.
Deliverables land in the chat, not on some disk.
Voice in and out
Local faster-whisper for recognition and edge-tts for playback, or plug in an OpenAI-compatible speech API.
When your hands are busy, just talk.
Long-term memory and permissions
Agent memory is isolated per account. Admins, regular users, connection sharing and command approval policies are managed in layers.
Shared use, separate boundaries.
Scheduled tasks
Wake an agent on a schedule; results, traces and files return to the original conversation.
No babysitting. It runs when it is due.
Interface preview
Four screens, from conversation to delivery
Nothing is buried in a menu. Chat, tasks, agents and services each have a place — switch tabs to see the actual structure. The screens below are structural schematics; the product is in beta and no online trial is offered yet.
In development
SessionLocal OpenCode · Release notes
ConnectedCheck the health of these three services, put it in a table, and include the collection time.
curl /api/healthRead 3 services200 OK · 9/9 connections healthy
All three are online. The health check returned 9/9, collected at 16:55. The table and raw JSON are attached below and ready to download.
ServiceStatusLatency
MatrixTodoOnline18 ms
Matrix MailOnline42 ms
NotebookOnline23 ms
health.jsonservices.csv
Ask a follow-up, or @ another agent
Note: these are high-fidelity interface schematics reconstructed from the current MatrixBot build. No AI-generated screenshots were used.
How it worksFour steps to put agents
Four steps to put agents
back into the workflow
No changes to the agent code, and no need to move machines together. Connect, divide the work, then start assigning.
01
Connect a machine
Enter the address and key, pick a runtime: Hermes, OpenCode, dsh, ACP — or just one OpenAI-compatible endpoint. Connection health is visible in settings.
02
Create an agent
Write the persona, choose model and reasoning effort, say which machine it runs on. Every agent keeps its own conversation, memory and tool permissions.
03
Assign a task
One sentence for the goal, a few acceptance criteria. MatrixBot drafts a plan and keeps pushing; it stops to ask when approval is needed.
04
Accept delivery
Watch tool calls, interrupt or stop at any time. Files, screenshots and media return to the chat. Completion writes back to the plan.
connection.yamlconnected
# connect a machine that is already running
kind: opencode
name: MatrixUbuntu · OpenCode
baseUrl: /home/matrixko
approval: ask
# a single model is enough too
kind: openai-compatible
baseUrl: http://your-host:8788/v1
model: matrixtokenapi/MatrixFlash
# health
health: 9 / 9 connections ok
Connection health
9/9healthy
Runtimes connected
OpenCodeHermesdshACP
Architecture
Runtimes on both sides, console in the middle
MatrixBot does not move agents into the browser. A connection layer maps their runtime state and capabilities onto one interface.
runtime → console
01 / RuntimesAny machine · any network
OpenCode
dsh
ACP
Hermes
OpenAI compatible
02 / Connection layerAuth · reconnect · health checks
Local socket
LAN HTTP
Tailscale
HTTPS proxy
MatrixBot Core
one console · one source of truth
Session routing
Task engine
Command approval
File transfer
Audit log
Account isolation
Web console
Chat / tasks / settings
Voice
Input and playback
Channels
WeChat · Feishu
Big screen
Cards pushed by agents
Self-hosting and securityCapability can be shared,
Capability can be shared,
boundaries cannot blur
Opening a self-hosted agent tool to several people means the security model cannot hide in a footnote. These four are defaults, not options.
Security boundary
Sharing a Hermes connection means the other person can act on that machine. Approval policies cover commands the runtime flags as dangerous; they are not a sandbox. Only share connections with people you trust.
— from the project security model
Account and data isolation
Each user gets a separate SQLite database. Regular users cannot see other people's conversations, runtime state, screen or home directory files.
Sharing and approval policy
Admins can share a connection with specific users and set command approval policies; commands that need approval can be set to always deny.
Audit log
Sign-ins, account and sharing changes, messages (length only), approvals and stops are all recorded and visible to admins on the users page.
Data stays on your machine
Self-hosted with no third-party hosting layer. Databases, keys and files live on your machines; model requests go only to the endpoints you configure.
Deployment checklist, reverse proxy and the security model are all in the docs.
Read the security docsWhere it fits
Who uses it, and how
Not a general-purpose AI platform. It serves people already running agents who need to keep the process under control.
Solo developer01
Several machines, each with a job
Builds on your laptop, docs in the office, checks on the server. One window to switch between them, no need to remember every login.
Turn changes across three repos into one page of release notes.
Small team02
Share capability, not data
Share a connection with a colleague, keep conversations and files isolated per account. Dangerous commands need approval; actions are audited.
Give a new teammate a read-only connection with bulk delete blocked.
Ops and automation03
Scheduled runs, interrupt on trouble
Checks, backups and log archiving run on schedule. When something looks wrong you can add instructions in the same run instead of restarting.
Check service health at 09:00 daily and summarise anomalies into the chat.
Content and research04
Long tasks with a beginning and an end
Gathering, layout, screenshots and PDF export run as one chain; intermediates and the final file both stay in the conversation.
Research a topic and produce a report, images and a downloadable file.
What enterprises use it forWork done on a computer
Work done on a computer
can, in principle, be handed over
MatrixBot is not a chat window that only answers questions. It runs on real machines: it opens applications, runs commands, reads and writes files, operates databases and internal systems. What it does not know, it learns; missing tools, it installs; finished work, it verifies. You give the goal and the acceptance criteria; it runs the process.
Why it can take the work
01
Real execution
Not advice, but action. Windows, Linux or macOS, locally or over SSH; the command line, a browser, Office files, databases and ERP back ends are all within reach.
02
Self-iteration
Facing an unfamiliar system, it reads the docs, the source, probes the API. Missing dependencies get installed, failing scripts get fixed. Once a path works, it is captured as a skill and reused.
03
Automation you can manage
Every task has a goal, a plan and acceptance criteria. Dangerous commands can require approval. The whole run is logged, and interrupted work resumes instead of starting over.
Work it can take on
Each item below maps to a concrete task, not an adjective on a capability list.
01
Build and change business systems
ERP, CRM, inventory, OA, reporting: new fields, changed workflows, approvals, data migration. Change it the way you describe it, without waiting for a release window.
02
Contracts and quotes
Draft from templates, diff versions, flag risky clauses; quotes to PDF, Word or Excel, generated, renamed and sent in batches.
03
Finance and data
Excel cleanup, reconciliation and roll-ups; invoice filing; management reports; exceptions pulled out with the reason attached.
04
Customers and marketing
Websites and landing pages, SEO, product copy, campaign assets, email marketing, follow-up records.
05
IT and operations
Server checks, backups, log triage, batch deployment, account and permission reviews. It investigates before it reports.
06
Engineering
Requirement breakdown, code, bug fixes, review, tests, releases, documentation.
07
Daily office work
Meeting notes, action tracking, mail triage, research, moving data between systems, weekly summaries.
08
Captured know-how
The way your best person does it becomes a skill and a checklist the next person can call.
One workflow, end to end
It starts with an inquiry email. No hand-offs, no switching systems.
- 01
Inquiry
spec and quantity from the email
- 02
Check
price, stock and terms in the ERP
- 03
Quote
Excel and PDF generated
- 04
Contract
template plus customer clauses
- 05
Diff
changes marked clause by clause
- 06
Send
reply, then write back to the CRM
- 07
Follow up
reminder if no reply by the due date
Illustrative flow; actual execution depends on your systems' APIs and the access you grant.
Boundaries and prerequisites
Access
Accounts, APIs or a UI. It can only go as far as you let it.
Approval
Deletes, payments and outbound mail can be set to require a human.
Responsibility
Contract terms, accounting treatment and legal judgement still need a qualified reviewer. It puts the draft and the differences in front of them.
Data
Self-hosted. The model endpoint can be an internal gateway; sensitive material need not leave your network.
Evidence
Talk to usThis site, from design and code to deployment, was produced by an agent on this machine.
Bring one role or one process. We will tell you whether it can be handed over.
Made by STONBuilt by STON Technologies
Visit the company site
Built by STON Technologies
(Xi'an) Co., Ltd.
MatrixBot is developed and maintained by STON Technologies (Xi'an) Co., Ltd. Founded in Xi'an in 2022, the company works on applied technology for analytical instruments, small-batch laboratory equipment, and embedded hardware and software, including AI deployments.
Founded
2022
Based in
Xi'an, China
Focus
Analytical instruments · Engineering
ContactSTON Technologies
Product enquiries, deployment and custom development all start with the channels above.
ICP filing陕ICP备2023019059号
Public channels
WeChatScan to talk to us directly
WhatsAppOverseas customers and partnersFAQ
Questions, answered plainly
Anything not covered here is documented in more detail in the docs.
Read the docsHow is this different from opening a terminal or an agent's own web UI?
MatrixBot does not replace your agent runtime; it is the unified entry point. Several runtimes and models connect once, with shared conversations, tasks, approvals, files and acceptance. You still use OpenCode, Hermes or a model's own abilities — you just stop carrying context between windows.
Which runtimes and models are supported?
On the runtime side: Hermes, OpenCode, dsh and ACP. On the model side: any OpenAI-compatible endpoint, including OpenAI, DeepSeek, Qwen, Kimi, OpenRouter, Ollama and vLLM. Connections can coexist and are assigned per agent.
Does my data go to a third party?
MatrixBot is self-hosted on your machines. Conversations, tasks, files and audit data live in your local database and never pass through a MatrixBot relay. Model requests go straight to the endpoint you configure, and you decide what leaves the machine.
If several people use it, can they see each other's conversations?
No. Each user has a separate database. Regular users cannot see other people's conversations, runtime state, screen, memory or file roots. Admins can share connections, but what is shared is capability, not the other person's data.
How are dangerous commands controlled?
Runtimes such as Hermes flag dangerous commands, and MatrixBot requires approval before running them. Admins can set approval policy per user, including always deny. To be clear: approval covers commands the runtime flags, it is not a sandbox. Only share connections with people you trust.
How is it deployed, and under what licence?
Node.js 22+ is required; clone the repo, then npm install, npm run build, npm start. A Docker Compose setup is also provided. The project is MIT licensed, so you can self-host, modify and deploy it internally.
BetaIn closed beta.
In closed beta.
Want a look? Just ask.
There is no public server yet, so no online trial. Interface walkthroughs, beta seats and private deployments can all be arranged by phone or email.
In development · no public access yet15 agents · 9/9 connections healthy · 12 service entriesMIT License