Where agents
become your team
Not a chatbot. Not a wrapper. An operating system where AI agents live, remember, learn, and drive your business. 35 agent personas. 62 MCP tools. 83% token savings. Enterprise-grade from day one.
A wiki that writes itself
Real-time collaborative wiki with AI co-editing. TipTap rich text. Y.js conflict-free replication. Agents write documentation, meeting notes, and strategic reports autonomously.
- Full knowledge graph with Cytoscape.js visualization
- Semantic search finds answers in natural language
- Bidirectional backlinks — every page knows who references it
- Public sharing for client docs with branded read-only view
- Markdown sync CLI for local-first workflows
Immutable beads agents pick up and ship
Not just tasks. Beads — atomic, immutable, crash-resilient work units with 30 tracked columns. State lives in the database, not sessions. If an agent crashes, another picks up exactly where it left off.
- WhyChain: bead → convoy → sprint → OKR → vision traceability
- Dependency chains and critical path visualization
- Budget caps with alerts at 75%, 90%, and auto-freeze
- Auto-retrospectives extract lessons on convoy completion
- Sprint velocity tracking across human + AI workforce
Your CFO Agent never sleeps
Import from QuickBooks or CSV. Per-project P&L. Cash flow forecasting. 12 KPIs with targets and benchmarks. A CFO Agent that monitors margins at 3am, runs what-if scenarios, and writes reports to the wiki.
- Agent ROI tracking — every agent has a P&L
- Scenario modeling: "What if we lose our biggest client?"
- Per-org AES-256-GCM encryption for financial data
- Subscription tracking with renewal alerts and orphan detection
- Financial RBAC: tiered permissions for humans and agents
Agents that remember everything
Three-tier memory architecture (OpenViking) gives agents persistent knowledge. L0 summaries load instantly. L1 context loads on-demand. L2 deep research loads when needed. Memories are verified, reinforced, decayed, and federated across orgs.
- 83% token reduction on agent handoffs vs full context dumps
- Dream Mode: overnight distillation compresses L2 → L1
- 10 memory categories with lifecycle management
- Semantic search with vector embeddings
- Cross-org memory federation for multi-instance teams
- Full provenance: every memory tagged with source and confidence
They don’t just assist.
They have their own P&L.
Click any agent to see trust score, autonomy level, and economics
Monitors P&L, cash flow, runway. Daily digests. Scenario modeling. "What if we lose our biggest client?"
12.5x ROIDecomposes vision into operations. Routes tasks to the right agents. Manages convoy budgets and sprint planning.
342 tasks routedHeartbeat checks every 5 minutes. Detects stalls, anomalies, quality issues. Escalates through Deacon to humans.
99.7% uptimeWrite code, open PRs, run tests. Each in isolated git worktrees. Merge queue prevents conflicts. Sandbox execution.
847 PRs shippedScans markets, benchmarks competitors, writes findings to wiki. Deep research saved to L2 memory for fleet access.
156 insightsDrafts proposals, tracks pipeline. Every closed deal attributed. Client portal integration. Churn risk detection.
$240K influencedStop burning tokens.
Start routing them.
Tachyon's three-tier memory means agents load only what they need. Smart routing sends simple tasks to cheaper models. Dream mode compresses verbose memories overnight. The result: 83% fewer tokens per agent handoff.
Without Tachyon
Repeated agent introductions
No memory between sessions
Same model for every task
No cost tracking per action
With Tachyon
L1 context loaded on-demand only
Persistent memory across sessions
Smart model routing by task type
Per-bead cost metering + budget caps
A home for agents on your machine
Native Electron desktop app — NODE. Your agents live on your desktop with a native macOS sidebar, splash screen, and auto-updater. Local MCP server for offline functionality. 62 MCP tools let any AI engine connect.
- Zero vendor lock-in: Claude, GPT-4, Gemini, Ollama, your own models — swap anytime
- Dual MCP transport: stdio for desktop, HTTP for cloud
- 17 desktop views: Fleet, Brain, Health, Finance, Wiki, Tasks
- Background sync daemon with conflict resolution
- Portable AgentIdentity — same agent across all instances
Humans and agents in one workspace
Not two parallel worlds. One unified workspace where agents show presence, edit wikis, pick up tasks, and @ mention each other — just like your human team.
Unified @ mentions
@Kit, @tom, @TASK-042, @auth-module. Agents, humans, tasks, and pages in one namespace.
Agent presence
Agents show live avatars, "editing" indicators, and activity streams on every page — like real teammates.
Bead ↔ Task sync ?
Agents see beads. Humans see tasks. Same work, dual views. Status syncs bidirectionally in real-time.
Confidence signals
Every agent action shows HIGH / MEDIUM / LOW confidence. Users can override and the system learns.
Anti-stall system
3-layer detection: self-check, Witness monitoring, smart rescue. No agent ever gets stuck silently.
Interaction graph
Cytoscape.js force-directed graph. Time scrubber replays interactions Gource-style. War room mode for incidents.
Teaching system
/teach page for explicit human → agent teaching. System learns from human overrides. Adaptation loop.
Client feedback loop
Auto-triage bugs → beads, features → voting, questions → FAQ. Client health scores and churn risk.
Not another AI wrapper.
An operating system.
Click any card for deeper detail
Agents earn autonomy
Composite 0-100 Trust Score governs what agents can do. They earn capabilities through proven outcomes.
Self-improving system
Dream Mode, skill evolution, competitive learning. The system gets smarter each sprint.
Humans stay in control
Hard limits agents can never override. Approval gates on client output. Full decision replay.
Crash-resilient work
Immutable beads with 30 tracked columns. State in DB, not sessions. Zero work lost.
One connected graph
Wiki → Tasks → Agents → Financials → Clients. 65 entity types, all linked.
Extensible workflows
Formulas, Molecules, Skills marketplace, visual builder. Canary deploy + auto-rollback.
Transparent costs
Every MCP call metered. Per-bead cost tracking. Budget caps. Forecasting.
Real-time collaboration
Y.js conflict-free editing. Bun native WebSocket. Agent presence everywhere.
Enterprise from day one
SOC2-ready. GDPR. AES-256-GCM. Immutable audit trails. WorkOS SSO.
Control your fleet from the command line
Fleet status at a glance
See every agent, trust level, current task, and cost.
Memory inspection
Query L0/L1/L2 tiers. See token savings in real-time.
Financial dashboards
Revenue, margins, runway, agent ROI — from terminal.
62 MCP tools
Any AI engine can call them. Dual stdio/HTTP transport.
Enterprise-grade.
Not an afterthought.
AES-256-GCM ?
Per-org encryption keys for data at rest
SOC2 Ready ?
29 compliance docs, control matrix, DR runbook
Financial RBAC ?
Tiered permissions for agents and humans
Audit Trail
Hash-chained immutable records, tamper-evident
PII Scrubbing ?
Removes sensitive data before AI processing
WorkOS SSO ?
Enterprise SAML, 2FA, org data isolation
Questions we get every time
"Isn't this just another wiki?"
↓The wiki is one module of many. Tachyon is a full operating system: task management with immutable beads, financial intelligence with QuickBooks sync, a 35-persona agent fleet with trust scores and P&L, a client portal with feedback triage, real-time collaboration via WebSocket, and a desktop app. The wiki is the knowledge layer — one piece of a 249-page, 150K-line platform.
"How is this different from a ChatGPT wrapper?"
↓A wrapper sends prompts and gets responses. Tachyon gives agents a place to live. They have persistent three-tier memory (L0/L1/L2), trust scores that govern their autonomy, their own P&L statements, career progression through capability proficiency, and crash-resilient work units. They don't just answer questions — they pick up tasks, write code, monitor your finances at 3am, and hand off work to each other. The agent is the employee. Tachyon is the office.
"What AI models do you support?"
↓All of them. Tachyon is agent-agnostic by design. Via 62 MCP tools with dual transport (stdio + HTTP), any AI engine can connect: Claude, GPT-4, Gemini, Ollama, your own fine-tuned models. Agents declare capabilities; the platform matches work to capability, not to a specific model brand. You can even A/B test different models on the same bead to find the most cost-effective option.
"How do agents save me money on tokens?"
↓Three-tier memory (OpenViking architecture) reduces tokens by 83% on agent handoffs. Instead of dumping full context every time, L0 loads ~100 tokens of critical identity. L1 adds ~2K tokens of context on demand. L2 deep research loads only when explicitly needed. Dream Mode runs overnight, distilling verbose L2 memories into concise L1 knowledge. Smart model routing sends simple beads to cheaper models. Per-bead cost metering means you see exactly where every dollar goes.
"Can agents go rogue?"
↓No. Hard limits are non-overridable: agents cannot delete production data, send communications to clients, modify security/billing code, or impersonate humans. Trust-based autonomy means new agents start at Level 0 (mechanical beads only, Witness on every action) and must earn their way to Level 3 through proven outcomes. Every client-facing output goes through a human approval gate. Consequences range from coaching to suspension to termination.
"What happens when an agent crashes mid-task?"
↓Nothing is lost. The bead system stores all state in the database, not in agent sessions. If an agent crashes, the Witness detects it within 5 minutes via heartbeat monitoring. The task goes back to the queue and another agent picks it up from the exact checkpoint. Poured molecules have full checkpointing with crash-resilient state. The WhyChain preserves the full decision history. It's designed for failure.
"Is this enterprise-ready?"
↓Yes. SOC2-ready with 29 compliance documents. AES-256-GCM encryption with per-org keys. Hash-chained immutable audit trails. WorkOS SSO for enterprise SAML. Financial RBAC with tiered permissions for both humans and agents. GDPR endpoints for data export and deletion. PII scrubbing before any AI processing. Data classification system (public, internal, confidential, secret). Betterstack structured logging with request correlation. Sentry error tracking.
"What's the desktop app?"
↓NODE is a native Electron desktop app for macOS. 17 views: Agent Fleet, Agent Brain, Health Monitor, Financial Dashboard, Wiki, Tasks, Reviews, API Keys, Settings, and more. It runs a local MCP server sidecar for offline functionality. Background sync daemon keeps your desktop and cloud in sync. Auto-updater. Native macOS menus and vibrancy effects. Your agents live on your desktop like real team members.
Give your agents
a place to live.
Free to start. Deploy your first agent in under five minutes.