Flapjack is an AI agent platform with memory, web tools, sandboxed compute, knowledge base RAG, MCP server integration, multi-channel deployment (Slack, Telegram, WhatsApp), multiplayer chat, scheduled tasks, custom webhook tools, database integrations, and automated runner pipelines with budget controls.
Every conversation starts cold. Your agent doesn't know the user's name, their last request, or what they care about. You end up building a custom context layer — and it breaks across sessions. Flapjack memory works at three scopes: agent-wide, per-thread, and per-resource. Memories are stored, recalled semantically, and injected automatically.
Your agent is stuck in its training data. It can't check today's docs, look up a live API response, or research a topic in depth. Four tools out of the box: search, deep research, URL reading, and recursive crawling. Toggle each independently.
The agent can talk about code but can't run it. It suggests a fix but can't verify it compiles. It writes a script but can't test the output. Sandboxed environments where agents execute code, run shell commands, and read/write files. Ephemeral per-thread or persistent across sessions.
Your agent doesn't know your product. It hallucinates features and references stale docs. Upload documents. Flapjack chunks, embeds, and retrieves the relevant pieces at query time. Agent-scoped or org-wide.
Complex tasks turn into a wall of text. The agent tries everything at once, loses track, skips steps. Structured plans with nested todos, status tracking, and turn-end reminders so nothing falls through.
Your agent lives in a chat widget. But your users are in Slack, Telegram, or WhatsApp. Each platform is a different backend. Connect an agent to any channel in minutes. Session strategy, contact tracking, and delivery — handled per-platform.
Every tool integration is bespoke. Custom auth, custom schema, custom error handling. Attach any MCP-compliant server. Stdio or HTTP, OAuth 2.1, encrypted credentials, tool namespacing. Browse the registry or bring your own.
Your agent can't see your data. It needs to answer questions about users or orders but you'd have to build a query layer. Connect Postgres or Supabase. Schema introspection, scoped table access, encrypted credentials. Read-only, no injection risk.
The agent needs to call your system. Create a ticket, charge a card, deploy a build. Define webhook tools with JSON schemas. Every call is HMAC-signed so your endpoint can verify it's real.
One user, one agent, one thread. But real work happens in teams. Multi-user threads with sender tracking, @mention routing, and configurable invocation. The agent knows who said what.
Your agent only responds when spoken to. But you need daily reports, check-ins, reminders. Cron-based scheduling with timezone support. One-time or recurring. Can target specific channels.
You built the agent, now someone needs to use it. Generate a public URL with a custom slug, optionally password-protected. No login required for end users.
Your agent is good enough to share. Publish to the Flapjack marketplace with a handle, avatar, description, category, tags, and pricing tier. Other teams can discover and use it.
You need the agent in your product, not a separate tab. @maats/flapjack gives you React hooks, streaming, and embeddable chat components. Drop it in and ship.
Your pipeline is a chain of API calls held together with glue code. Each step type has different failure modes and you handle them all manually. Four step types: agent (LLM reasoning), computer (deterministic code), webhook (external calls), and condition (branching logic). Each with configurable retries, timeouts, and input mapping.
Something needs to kick off the pipeline. You build a cron job, a webhook listener, and a manual trigger — three different systems. Six trigger types in one place: manual/API, cron, webhook, GitHub poll, bulk import, and an embeddable HTML button. All create runs with idempotency protection.
A runaway loop burns through your API credits before anyone notices. Set a USD budget per runner or per run. Hit the limit and the run pauses or fails — your choice. Resume with additional budget when you're ready. Warnings at 80% and 95%.
You need to know what happened, when, and why it stopped. Full lifecycle tracking: pending → running → completed, failed, cancelled, or paused. Each step records duration, token usage, cost, and a trace of every tool call. Idempotency keys prevent duplicate runs.
Your runner steps need the same tools, databases, and MCP servers as your agents. Runner-level defaults for web tools, MCP servers, database integrations, custom tools, knowledge, and skills. Agent steps inherit and can override. No duplicate configuration.
You're flying blind on cost and performance. Per-run and per-step metrics: token usage, estimated cost, duration, and tool call traces. See which steps are expensive, which are slow, and where runs fail.
| Feature | Agents | Runners |
|---|---|---|
| Interaction model | Conversational — responds to messages | Step-based — executes a pipeline |
| Triggered by | User message or channel event | API call, cron, webhook, poll, or button |
| State | Thread history — grows with conversation | Run context — accumulates across steps |
| Channel support | Slack, Telegram, WhatsApp | — |
| Budget controls | — | Per-runner and per-run USD caps |
| Multiplayer | Multi-user threads with @mention routing | — |
| Tools & MCP | Yes | Yes (runner defaults + agent overrides) |
| Knowledge / RAG | Yes | Yes (runner-scoped + org-wide) |
| Web access | Yes | Yes |
| Compute sandbox | Yes | Yes (as computer steps) |
Everything you need to go from prototype to production — without reinventing the plumbing.
Every conversation starts cold. Your agent doesn't know the user's name, their last request, or what they care about. You end up building a custom context layer — and it breaks across sessions. Flapjack memory works at three scopes: agent-wide, per-thread, and per-resource. Memories are stored, recalled semantically, and injected automatically.
Your agent is stuck in its training data. It can't check today's docs, look up a live API response, or research a topic in depth. Four tools out of the box: search, deep research, URL reading, and recursive crawling. Toggle each independently.
The agent can talk about code but can't run it. It suggests a fix but can't verify it compiles. It writes a script but can't test the output. Sandboxed environments where agents execute code, run shell commands, and read/write files. Ephemeral per-thread or persistent across sessions.
Your agent doesn't know your product. It hallucinates features and references stale docs. Upload documents. Flapjack chunks, embeds, and retrieves the relevant pieces at query time. Agent-scoped or org-wide.
Complex tasks turn into a wall of text. The agent tries everything at once, loses track, skips steps. Structured plans with nested todos, status tracking, and turn-end reminders so nothing falls through.
Your agent lives in a chat widget. But your users are in Slack, Telegram, or WhatsApp. Each platform is a different backend. Connect an agent to any channel in minutes. Session strategy, contact tracking, and delivery — handled per-platform.
Every tool integration is bespoke. Custom auth, custom schema, custom error handling. Attach any MCP-compliant server. Stdio or HTTP, OAuth 2.1, encrypted credentials, tool namespacing. Browse the registry or bring your own.
Your agent can't see your data. It needs to answer questions about users or orders but you'd have to build a query layer. Connect Postgres or Supabase. Schema introspection, scoped table access, encrypted credentials. Read-only, no injection risk.
The agent needs to call your system. Create a ticket, charge a card, deploy a build. Define webhook tools with JSON schemas. Every call is HMAC-signed so your endpoint can verify it's real.
One user, one agent, one thread. But real work happens in teams. Multi-user threads with sender tracking, @mention routing, and configurable invocation. The agent knows who said what.
Your agent only responds when spoken to. But you need daily reports, check-ins, reminders. Cron-based scheduling with timezone support. One-time or recurring. Can target specific channels.
You built the agent, now someone needs to use it. Generate a public URL with a custom slug, optionally password-protected. No login required for end users.
Your agent is good enough to share. Publish to the Flapjack marketplace with a handle, avatar, description, category, tags, and pricing tier. Other teams can discover and use it.
You need the agent in your product, not a separate tab. @maats/flapjack gives you React hooks, streaming, and embeddable chat components. Drop it in and ship.
Ready to build?
AI agents, without the infrastructure headache.
Make your first agentAlready have an account? Sign in