2pm.spaceOne AI workspace for your
customers, your data, your content
Build agents with skills, tools and guardrails — then put them to work answering Messenger and Zalo, turning your database into plain-language answers, and filling next week's content calendar. All on a shared Drive of live documents, mind maps, sheets and tables.
results = search_kb( "Q4 sales", top_k=6 )
chart = generate_chart( revenue_data, type="bar" )
Six products, one workspace
They share the same agents, the same memory and the same knowledge base — which is the whole reason they are in one place.
Agents that reason, act, and iterate
A visual builder for agents that plan, call tools, read your documents and adjust until the task is done — not a wrapper around one prompt.
- Simple single-call agents, or multi-node workflow graphs
- Skills, tools and guardrails shared across every agent
- Three-layer memory that persists between conversations
- Per-run logs with tokens, cost and every tool call
Every customer conversation, one queue
Messenger, Facebook comments, Zalo, Telegram and your website chat, answered by an agent that knows your products and hands over to your team on request.
- Seven channel types, plus a webhook channel for anything else
- Human handover mid-thread, and approval gates on risky actions
- Phone numbers detected and attached to the contact
- Care follow-ups with quiet hours and per-person caps
Your database, in plain language
Connect a warehouse, define your measures once, and let anyone ask a question and get an answer with the chart, the table and the SQL attached.
- 14 databases and warehouses, from Postgres to BigQuery
- A semantic model, so every answer agrees with the last one
- Multi-tab boards with filters and relative date windows
- SQL Lab, and verified question-to-SQL pairs
Plan it, write it, illustrate it, ship it
Brief a Brand once and let agents fill the calendar, draft the copy, generate the images and publish — with a human approval step in front of every post.
- Ideas planned against your brand voice and products
- Image generation on a canvas of rewirable steps
- An approval queue between drafting and publishing
- Publishing to Facebook Pages, WordPress or your own webhook
One record, however they reached you
Conversations from every channel, phone numbers caught mid-thread, and live data from your own systems resolve into a single customer your agents can answer from.
- Contacts matched across channels into one timeline
- Enrichment from an HTTP endpoint you control
- Spreadsheet import and cursor-based sync
- Permission-gated at the database, not just the API
The shared brain your agents read
Documents, mind maps, sheets and tables your team edits together in real time — and the same knowledge base your agents retrieve from when they answer.
- Real-time co-editing with live cursors
- Hybrid retrieval: vector search fused with keyword search
- Share by link with view, comment or edit access
- Unlimited mind maps, free, with no node caps
See it working
Drawn from the same design tokens as the product, so these are the screens as they look in your theme — not a screenshot of someone else's workspace.
Wire an agent together on a canvas
Drop in the skills it should know, the tools it may call and the guardrails it must stay inside, then watch it reason through a task one node at a time.
- Simple single-call agents, or a full workflow graph
- Skills and tools shared across every agent in the workspace
- Versioned canvas — restore any earlier shape of the graph
Every channel in one queue
Messenger, Facebook comments, Zalo, Telegram and website chat arrive together. The agent answers with your catalogue and your tools; your team takes the thread whenever it wants it.
- The customer’s history from other channels, on the panel beside the thread
- Phone numbers detected mid-conversation and attached to the contact
- Tool calls visible inline, so you can see what the reply was built from
A question in, a chart and the SQL out
Ask in plain language and get the answer with the query that produced it, compiled from measures you defined rather than guessed at from raw table names.
- 14 databases and warehouses, from Postgres to BigQuery
- The chart picked from the shape of the result
- Pin any answer onto a multi-tab board
A quarter of content, planned and shipped
Ideas planned against your brand fill the calendar, drafts get written and illustrated, and nothing reaches a Page until someone approves it.
- Drag a slot to reschedule it; re-roll the one field that missed
- An approval queue between drafting and publishing
- Publishing to Facebook Pages, WordPress or your own webhook
The agent builder underneath all of it
Everything above runs on the same builder — so a skill you write once is available everywhere you put an agent.
AI Agents
Build and customize intelligent agents tailored to your use cases. Define personas, instructions, and behaviors with a visual builder — then version the canvas so a change you regret is one restore away.
Custom Skills
Teach agents domain-specific skills and workflows. A skill can be injected into every prompt, or only when the question matches it — so a big library of behaviour does not cost tokens on every call.
Tool Integrations
Connect agents to external APIs, databases and services. Webhook tools, OpenAPI imports, sandboxed Python and JavaScript, MCP servers, or another agent exposed as a tool.
See everything it connects toGuardrails
Set safety constraints and content policies on what goes into an agent and what comes out — with an audit of which guard fired, what it did, and why.
Collaborative Drive
Documents, mind maps, sheets, and tables with real-time collaboration. Your team's shared knowledge base, and the one your agents retrieve from.
Try the free mindmap makerOne-Click Deploy
Put an agent on a channel, a schedule, or an API endpoint. Monitor performance, track cost per run, and iterate with the logs in front of you.
How it works
Go from idea to production in three steps. No complex setup, no infrastructure headaches.
Build your agent
Define your agent's persona, instructions and behaviour. Start from a marketplace template or from a blank canvas in the visual builder.
Give it skills, tools and knowledge
Add skills, connect tools, point it at the documents and databases it needs, and set the guardrails it must stay inside.
Put it to work
Send it to a channel, a schedule or an API. Watch every run with its tool calls and cost, and improve it from what actually happened.
Build your agent
Define your agent's persona, instructions and behaviour. Start from a marketplace template or from a blank canvas in the visual builder.
Give it skills, tools and knowledge
Add skills, connect tools, point it at the documents and databases it needs, and set the guardrails it must stay inside.
Put it to work
Send it to a channel, a schedule or an API. Watch every run with its tool calls and cost, and improve it from what actually happened.
Built on a powerful engine
Under the hood, 2pm.space is more than a chat interface. Every capability is a production-grade engine designed to make your agents smarter, faster, and safer.
Agents that think, act, and iterate
Every agent on 2pm.space is built on a ReAct loop — Reason, Act, Observe, repeat. Agents don't just respond; they plan, call tools, evaluate results, and adjust until the task is done.
- Simple mode: single LLM call for fast, direct answers
- Agentic mode: multi-node workflow graph with conditional routing and parallel execution
- Each node supports up to 10 tool-calling iterations by default
- Back-edges (cycles) allow agents to revisit steps without infinite loops
Memory that grows smarter with every conversation
Agents on 2pm.space remember. A three-layer memory system — global tenant memory, per-agent shared context, and per-user notes — ensures every interaction is informed by what came before.
- Automatic memory extraction from every conversation in the background
- Semantic retrieval with decay scoring — recent, relevant memories surface first
- Six memory types: fact, preference, instruction, feedback, episode, procedure
- Deduplication at three levels: LLM prompt, vector similarity, and periodic consolidation
Connect agents to everything
Agents can call any tool — built-in platform tools, external webhooks, custom Python/JS code, or any MCP server. The ReAct loop handles errors gracefully, retrying intelligently without manual intervention.
- Built-in tools: web search, image generation, code execution, and more
- Webhook tools: POST/GET/PUT/PATCH to any API with bearer or API key auth
- Code tools: run sandboxed Python or JavaScript directly from agent nodes
- MCP support: connect any MCP server for runtime tool discovery
Your knowledge base, always within reach
Upload documents, mind maps, spreadsheets, and tables to the Collaborative Drive. Agents retrieve the most relevant chunks on demand using hybrid search — combining dense vector and BM25 keyword retrieval.
- Hybrid search by default: 70% vector + 30% keyword via RRF fusion
- Semantic chunking with contextual embeddings for higher retrieval accuracy
- Two RAG modes: "tool" (agent-triggered) or "auto" (always injected)
- Tiered retrieval: chunk summaries first, full content on demand — saves tokens
Reusable instructions your agents can master
Skills are prompt templates that get injected into agent system prompts — either always, or automatically when the user's query matches. Define once, reuse across every agent in your workspace.
- "Always" mode: injected into every system prompt for guaranteed behavior
- "Auto" mode: injected only when query matches via full-text search — saves tokens
- Global skills apply to all agents; agent-specific skills stay scoped
- System skills are platform-wide, bypassing tenant restrictions
Full visibility into every agent run
Every execution is tracked end-to-end — per-node token usage, cost, latency, tool calls, and guard outcomes. Debug any failure, optimize costs, and understand exactly what your agents are doing.
- Per-node tracking: input tokens, output tokens, and cost per step
- Execution logs with full message history and tool call results
- Guardrails audit: see which guards triggered, what action was taken, and why
- Memory extraction logs: track what was learned from each conversation
Why teams choose 2pm.space
See how 2pm.space transforms your AI workflow from scattered tools to a unified platform.
One tool for the inbox, another for dashboards, a third for content — and none of them share what they know.
One workspace where the same agents, memory and knowledge base serve all three.
A chatbot on the Page that answers from a script and gives up when asked anything real.
An agent with your documents, your catalogue and your tools — and a teammate one click away.
Every question about the numbers waits in an analyst queue.
Anyone asks in plain language and gets the answer, with the SQL attached.
No memory between sessions. Context is lost every time a new chat starts.
Layered memory that persists across agents, channels and sessions.
Per-seat pricing that punishes you for inviting the rest of the team.
Every feature free, unlimited members, and usage billed at the model provider's own rate.
Access is all-or-nothing, so a contractor on one Page can see the whole workspace.
Per-resource permissions and channel-level access for members, roles and teams.
And much more
Built for teams that want full control over their AI — from the first prototype to production.
Fine-Tuning
Turn the conversations your team answered well into a training set, tune an open-weight model on it, and point an agent at the result.
How fine-tuning works hereRoles & Governance
Custom roles, teams, per-resource sharing and channel-level access — plus encrypted credentials and a log for every kind of run.
See the security controlsIntegrations & MCP
Channels, HTTP APIs, databases, sandboxed code and MCP servers — and an MCP endpoint of our own so an external client can drive the workspace.
Browse the catalogueAPI Access
Workspace API keys scoped to their creator's permissions. Drive agents, documents and boards from your own applications.
Scheduler
Run an agent on a schedule — a morning report, a nightly sweep, a weekly digest — with the run history kept beside every other log.
Marketplace & Apps
Install an agent template into your workspace and edit it, or ship an app built on the platform to a workspace of your own.
Playground
Test an agent before it meets a customer. Replay a conversation, inspect the thinking chain, and watch each tool call resolve.
Team Workspaces
Multiple workspaces, each with its own members, roles and data — with an interface available in eight languages.
Global Memory
A workspace-level memory every agent reads and writes — so what one agent learns about your business does not stay locked inside its own conversations.
Mobile App
The inbox in your pocket on iOS and Android, with deep links that open straight to the conversation.
Get the mobile appFrequently asked questions
Everything you need to know about building and deploying AI agents with 2pm.space.
A workspace where you build AI agents and then actually put them to work — answering customers across Messenger, Zalo, Telegram and website chat, answering questions about your database in plain language, and running your content calendar — all on top of a shared Drive of live documents, mind maps, sheets and tables.
Anything from a single-call assistant to a multi-node workflow with conditional routing, parallel branches and evaluation loops. Support agents, research assistants, content writers, data analysts and internal automations are the common shapes, and each one is assembled from the same skills, tools, knowledge and guardrails.
No. Agents are built visually — persona, instructions, skills, tools and guardrails are all configuration, and there is a marketplace of templates to start from. If you do write code, sandboxed Python and JavaScript tools, an API and an MCP endpoint are all there when you want them.
Three ways, together. Memory is extracted from conversations in the background and retrieved semantically, so an agent remembers what it was told. Documents in the Drive are retrieved on demand with hybrid vector-and-keyword search. And skills — reusable instructions you write once — are injected into the prompt either always or only when the question matches.
Yes. Messaging channels connect as first-class channels, HTTP APIs as webhook or OpenAPI tools, fourteen databases and warehouses as data connections, and any MCP server as a runtime tool source. Anything else reaches us through a plain webhook.
Every feature is free, with no paid tier and no per-seat charge. You pay only for AI usage — model tokens are billed at the provider's own rate with no markup, and the platform takes a single flat fee when you add credit. Storage above the included base is the one thing billed monthly.
Workspaces hold members, roles and teams, with permissions granted per resource and per action. On top of that, an individual agent, document, channel or board carries its own access list — so a contractor can be given one Page without being given the workspace.
Every run is logged end to end: the messages, each tool call and what it returned, the tokens and cost per step, and which guardrails fired and what they did. Scheduled runs, webhook deliveries and follow-up sends each keep their own log beside it.
Ready to build your first agent?
Every feature is free and there is no credit card to sign up. You only ever pay for the AI you actually use.