Concepts and Features
A tour of everything Pulse can do, organised by how you actually use it.
On this page
Agents
The worker unit of Pulse. An agent subscribes to a topic of events, reasons about each one, and emits new events downstream. Four engines available:
Rule-based
Deterministic · Free · Predictable
Condition-action logic: if priority == 'high' and sender ends_with '@customer.com', emit on urgent-routing.
Use when: Logic is truly deterministic; you want zero LLM cost.
Routing by priority, schema filtering, dedup, count aggregation.
LLM
Reasoning · Flexible · Token-cost
A language model reasons about every event using a system prompt. Run a local model on your own machine, plug in any cloud LLM provider, or point to your own endpoint.
Use when: Input needs judgement (is this email angry? is this CV a good fit?).
Email triage, CV scoring, content classification.
MCP
Scripted · API-driven · Fast
Invokes Model Context Protocol tools in sequence, without LLM reasoning in between. A scripted orchestrator that knows how to use APIs.
Use when: Workflow is API-driven (fetch X, post Y, log Z), no judgement needed.
Scrape feed → post Slack summary; daily cleanup across services.
Streaming
Stateful · Exactly-once · Horizontally scalable
Production-grade stream processing built via a no-code wizard. Single-binary deploy, no native dependencies, event-time windows with watermarks, CEP / pattern matching, and stateful processing over millions of events. Scale horizontally when you need to — without rewriting your agent.
Use when: High-volume event streams, windowed aggregations, sequence detection, or stateful joins.
Real-time fraud patterns, KPI dashboards, sessionization, sliding-window analytics.
Chat with your agents
Every agent has its own scoped chat. Ask it 'what did you do today?', 'why did you reject event abc-123?', or replay a past event. The platform-wide Pulse Assistant has the full toolkit: create agents, deploy templates, run maintenance.
Pipelines
A named, versioned chain of agents that accomplish a business outcome. Build from scratch on the Flow page or deploy a template that provisions everything.
- •Versioning: every deploy creates a revision. Roll back if it misbehaves.
- •Observability: live throughput, error rate, last run, quick links to agents.
- •Composition: wire agents together via topics with a declared input source.
Shadow deploys (PRO+)
Build a candidate pipeline (e.g. a new LLM prompt). Deploy as a shadow, it sees real traffic but its decisions don't affect anything. After 100-1000 events, compare side-by-side. If it wins, promote in one click. How serious teams ship LLM changes without praying.
Templates
Pre-built pipelines for common automation scenarios. Pulse ships with 140+ templates, 10 free, 130+ premium (PRO or à-la-carte unlocks).
Each template includes:
- •Name, description, icon, category, browsable on the Templates page.
- •Prerequisites, API keys, plugins, uploads. Pulse checks BEFORE deploying.
- •Config fields, wizard form for template-specific choices.
- •Rendered pipeline, the agents, topics and rules that will be created.
A template deploy is idempotent, if it already exists, you see a 'reconfigure' option instead of duplicating.
MCP plugins
MCP (Model Context Protocol) is an open, widely-adopted standard for tool-use by language models. Pulse is an MCP host: any tool that speaks MCP works inside Pulse with zero integration code.
Community plugins (every plan)
Commercial plugins (PRO+)
Higher-effort integrations with paid licenses or maintenance contracts: premium Salesforce, advanced LinkedIn Sales Navigator, Exchange/Workday calendaring, premium market data. You can also write your own, the MCP SDK ships with Pulse, ~100 lines of Java or Python.
Chat
Pulse has two chat surfaces, worth understanding the difference.
Pulse Assistant
Global · Sidebar
Full platform toolkit, create agents, deploy templates, start/stop pipelines, adjust settings, run maintenance. The expert operator you hired, always available.
Agent-scoped chat
Per-agent · Detail page
Locked to one agent. Ten tools: get_my_status, get_my_config, get_my_metrics, query_my_state, list_my_recent_inputs/outputs, inspect_event, replay_event, pause_me, resume_me. For debugging and supervision.
Approvals for high-risk actions
Some tool calls are high-risk: delete_pipeline, create_agent, run_maintenance, stop_agent. The Pulse Assistant defers them, you see a card to approve, reject or edit. Mistakes don't become incidents.
Events & topics
The plumbing layer. Most operators never touch this directly but understanding it helps with debugging.
Guardians
Pulse's policy layer. Two complementary mechanisms decide whether an event is allowed to commit downstream.
Rule-based guardians
Deterministic rules that run BEFORE an event commits and can block it. Blocked events go to the DLQ with a structured reason, review, fix, reinject or discard.
Schema validation
Reject events missing required fields.
Rate limiting
Block when a source exceeds N events per minute.
Content policies
Block events matching a regex (e.g. PII patterns).
Business rules
Never send email to non-company addresses after 6 PM.
PVSC, consensus validation for critical AI decisions
For decisions where one LLM being wrong is unacceptable, auto-refunds, medical triage, trade execution, legal clause extraction, agents opt into PVSC mode: a committee of independent guardians reviews every output, agreement is required to commit, dissent blocks the event and escalates to a human with the full review record. Every decision leaves a tamper-evident audit trail.
Trade-off: PVSC adds latency vs the default fast path. You opt in per agent, use it where it's worth the wait.
Included on every plan, including FREE. This is one of Pulse's genuinely-unique architectural features, no other self-hosted AI platform ships a comparable consensus layer.
Read the full PVSC pageSources
External feeds that flow events into Pulse. Configure in the Sources page, each has a 2-minute wizard.
Schedules
Cron-style triggers for pipelines that run on a cadence. Enter a cron expression (0 7 * * 1-5) or pick a preset. Timezone aware. Pause/resume without deleting. Scheduler runs inside Pulse, no external cron needed.
Learning
Pulse watches what you approve, reject or correct and turns the feedback into long-term knowledge.
- •Corrections: when you edit an LLM's output before sending, Pulse remembers the pattern.
- •Preferences: ask the assistant to remember something ('I prefer short summaries').
- •Knowledge extraction: per-user memory that every LLM agent can optionally reference.
Approvals
The human-in-the-loop surface. Any LLM agent can mark a decision as 'needs human approval' before committing. Approvals collect in an inbox, you see input, LLM reasoning, proposed action, then approve / reject / edit. Rejected ones feed back into Learning.
Backups & restore
Automatic snapshots of everything Pulse persists. Available on every plan.
- •Scheduled backups: default daily, retaining 7 snapshots. Configurable cadence.
- •On-demand: 'Back up now' button takes an immediate snapshot.
- •Download: any snapshot exports as a single zip.
- •Restore: stage-then-apply-on-boot. A banner shows pending; cancel before restart.
Multi-user & audit
Plans scale with your team. Every admin action is written to an append-only audit log.
| Plan | Users | Audit log |
|---|---|---|
| FREE | 1 admin account | 30 days retention |
| PRO | Up to 5 users (admin / member / viewer) | 365 days retention |
| ENTERPRISE | Unlimited seats + SSO/LDAP + per-resource ACLs | Configurable, exportable |
Observability
Wire Pulse into your existing monitoring stack.
- •/health endpoint, JSON status of every subsystem.
- •/metrics endpoint, Prometheus-format text, scrape with any modern stack.
- •WebSocket dashboard, live agent status without polling.
- •Log files, rotating, gzipped, journalctl/Fluent Bit compatible.