For smaller models
Supervisor
Split the work into focused roles with scoped context, so the model has less to handle in each step.
Coordinates the request. Delegates as needed.
- PlannerOrganize the work
- WorkerUse tools
- ReviewerReview results
A structured agent loop
for smaller local models.
An open-source TypeScript runtime for local AI agents. Give smaller models focused roles, or let more capable models run the tool loop directly.
New in 2.2.0 · Continue with ChatGPTBuilt around your model
One runtime. Two execution paths.
Not every model needs to handle everything at once.
For more capable models
Execution Agent
Let one agent own the tool loop and final response, with planning and auditing available when needed. Resources adds embedding-backed search and focused reads of your indexed text files, text-bearing PDFs and DOCX documents across conversations.
Runs tools directly. Calls for advice as needed.
- PlannerSuggest a plan
- AuditorCheck the evidence
NEW IN 2.2 · CONTINUE WITH CHATGPT
Your ChatGPT plan.
Your ABot workspace.
Use an eligible ChatGPT plan with ABot’s agents and tools. Connect your account in the local Web UI, without setting up an API key for this connection.
How to connectOpenAI determines eligibility, models and usage limits. Sign-in happens in your local ABot installation.
A connection, in three steps
Choose Continue with ChatGPT
During setup, or when adding a model in Models.
Sign in and grant permission
Complete OpenAI’s sign-in and consent in your browser.
Choose your model
Return to ABot. Review context capacity and reasoning support, then save.
Runtime and browser on the same computer. Embeddings are configured separately.
Introduced in 1.7.0
A simpler setup. More ways to extend ABot.
MICROSOFT STORE · WINDOWS
Windows Companion
Install ABot AI Companion from the Store. Review the Runtime address and approve the connection.
- InstallFrom Microsoft Store
- OpenLaunch from ABot
- ConfirmRuntime address & access
Runtime stays in Docker or WSL. Local forwarding is required; computer permissions are configured separately.
BETA · PLUGINS · RUNTIME 1.7+
Community Marketplace
Find free, open-source plugins. Review and install them in your own Runtime.
- GitHub sourceChoose a plugin
- Review in RuntimeCommit, tools & settings
- Install + ApplyYour explicit approval
With Computer access paired, the website opens Discover for review, not an automatic installation. Or paste a GitHub URL directly in Runtime.
ABot Studio Since 1.6
Tools and agents.
One reusable workflow.
Connect the steps in a visual editor—or ask your agent to draft them. Run on demand, on a schedule, or when matching files arrive.
Studio
Connect tool and agent steps, including parallel branches. This is a draft: the model and permissions must be configured before running. Tools shown reflect this installation.Swipe the image sideways to explore.
View full size 
Start from a template, then choose your inputs and model. Available tools depend on the plugins configured in your Runtime.Swipe the image sideways to explore.
View full size - Build it your way
- Start from a template. Reuse agents, connect inputs, and set result checks.
- Choose when it runs
- Manual runs, schedules, or folder polling. Publish and activate explicitly.
- See what actually happened
- Inspect saved outputs, tool receipts, and run reports—not just the final answer.
Drafting does not grant permissions or start a run. Runs use saved Full or FULL+ permissions while Runtime is active. Reruns require confirmation.
Explore Studio’s releaseABot SPARK HOME · New in 2.3
A home for your day.
A place to keep working.
Keep your day, your system and your ongoing work in view. Shape your HOME when you choose, with SPARK Chat always close by.
HOME
Collection pausedLearning pausedProactive paused
Local time
09:41
Monday, 9 November 2026
A little context for your day.
November 2026
System monitorComputer
Native readings, with scope in view.
CPU12%
Memory38%
Temperature—
Live widgetExample values
Pick up where you left offConversations & recent activity
Plan the weekConversation ↗
Review project notesConversation ↗
No active tasksActivity
Talk to SPARKOpen chat ↗
Product illustration · example data
One chat per environment, with its own tool permissions. Collection and proactive messages are off by default. Chat starts in Ask mode and uses your selected Proactive model; activity text goes to its configured model provider.
- Keep the conversation going
- Your persistent SPARK chat stays close, with attachments and optional suggestions from the activity you allow.
- Shape your HOME
- Generate and calibrate a dashboard explicitly. Bring reusable views together with live system metrics.
- You decide when it changes
- Review approval requests and cancel when needed. A failed or cancelled generation keeps your previous HOME.
HOME generation is manual and uses your configured HOME and SPARK models. Metrics identify the computer or container being measured; available readings depend on supported sensors. Collection and proactive suggestions remain separate, optional settings.
Explore SPARK HOMEComputer use
Let your agent see what’s on your screen and take action.
See the screen
Capture the screen or a selected area.
Work with windows
Find windows and bring the right app into focus.
Take action
Click, type, drag, and scroll.
MEMORY
Keep what matters.
Carry selected facts and preferences into future conversations, with saved memories you can review and edit.
Stored memories
EnabledPrefers concise summaries with sources linked inline.
Added in Web UIAvailable for relevant future conversations
SCHEDULED WORK
Put it on the calendar.
Schedule a request in its conversation, with a saved model, tool permissions, and a run history.
Scheduled task
Weekly project review
active- Schedule
- Monday at 09:00
- Time zone
- Etc/UTC
- Model
- Your saved model
- Tool access
- Full
Project planning
Run historyNo recorded runs yet.
Jobs run while Runtime is active. Memory is optional.
Quick start
From an empty directory to a running agent.
Install ABot, start the local Web UI, and let guided setup connect your model and tools.
SETUP
Create a project
$ mkdir my-abot
$ cd my-abot
$ npm init -y
$ npm install @abot-ai/runtimeCONFIGURE
Choose a provider in guided setup
Run npx abot start and open http://127.0.0.1:5177. On a fresh installation, choose llama.cpp, Ollama, OpenAI with an API key, or Continue with ChatGPT. Configure your model, optional embeddings, and plugins and tools, then finish setup to start chatting.
For ChatGPT, use a browser on the same computer as Runtime. Complete OpenAI’s sign-in and consent, select your account and model, then review its context capacity and reasoning support. Existing installations can add the connection in Models. Embeddings use a separate provider.
Prefer the terminal? Initialization is also available with these commands:
Ollama
$ npx abot init \
--provider ollama \
--model <model-id> OpenAI
$ npx abot init \
--provider openai \
--model <model-id> llama.cpp · New in 2.0
$ npx abot init \
--provider llamacpp \
--model <server-model-id> \
--base-url <server-origin> \
--context-window <context-tokens>Using llama.cpp? Follow the llama.cpp setup guide for server compatibility, authentication, thinking budgets, and embedding connections.
RUN
Start ABot
$ npx abot start
# Local Web UI
http://127.0.0.1:5177Open source
Inspect the runtime.
Extend what it can do.
ABot Runtime is available under the MIT license. Read the source, inspect the public contracts, or build a capability around its plugin system.