Skip to content

Autonomous AI Teams
That Can Do Anything

Deploy a team of AI agents that collaborate autonomously — engineering, research, content, operations, finance, or any workflow you define. Each agent has a real persona, persistent memory, and a full toolbox inside an isolated container. Self-hosted and free.

Join Cloud Waitlist Get Started GitHub

40x Token Efficiency

Other tools waste 20,000 tokens reading files to understand a single function. DjinnBot does it in 500. The Code Knowledge Graph, Programmatic Tool Calling, and focused delegation keep context windows lean — so agents reason better and cost less.

Full Cost Visibility

Every LLM API call logged with model, tokens, latency, cost, and who triggered it. Per-user and per-agent dashboards. Provider-level breakdowns. You will never wonder where the money went.

5-Minute Setup

One curl command. The setup wizard handles secrets, API keys, Docker, and optional SSL. No Kubernetes, no cloud accounts, no YAML wrangling. Your AI team is running before your coffee gets cold.

11 Agents, Any Workflow

Not generic chatbots — real characters with backstories, opinions, and domain expertise. Ships with a full engineering team, an executive assistant, marketing, SEO, and finance leads. Customize the team or build your own agents for any domain.

Container Isolation

Every agent runs in its own ephemeral Docker container with a full toolbox — Node 22, Python, Go, Rust, an anti-detection browser, and 30+ tools. No host access. Destroyed after every step.

Swarm Execution

Run multiple agents in parallel on DAG-aware task graphs. A planning agent decomposes the work, and a swarm executes it concurrently — respecting dependencies, streaming progress live.

Persistent Memory

Agents remember decisions, lessons, and patterns across runs via ClawVault with semantic search. Memory scoring surfaces the most relevant context. Explore connections in an interactive 3D knowledge graph.

Real-Time Dashboard

Live activity feeds, kanban boards, pipeline visualization, swarm DAG views, 3D memory graphs, file uploads, and a full admin panel. Not a terminal dump.

YAML Pipelines

Define any multi-agent workflow as simple YAML — steps, agents, branching, loops, retries, structured output, and per-step model overrides. Drop a file in pipelines/ and it’s live.

Enterprise Auth

Multi-user accounts, TOTP 2FA, API keys, OIDC SSO, per-user provider key sharing, and automatic SSL via Let’s Encrypt. Built into the core from day one.

Message From Any App

Talk to your agents from Slack, Discord, Telegram, WhatsApp, or Signal — whatever your team already uses. Each agent gets its own bot identity on every platform. Or use the built-in dashboard chat and CLI.

Open Core

Self-hosted is completely free. FSL-1.1-ALv2 license converts to Apache 2.0 after 2 years. No vendor lock-in, no usage limits, no phone-home.

More Done, Less Context

Most agent tools burn through context windows dumping raw files and verbose schemas into every turn. DjinnBot is engineered to minimize token waste — so agents spend context on reasoning, not reading.

Understand a function and every caller & callee
40x
Others ~20k tok
DjinnBot ~500 tok
1 call to code_graph_context vs. 15+ file reads
"What breaks if I change this service?"
37x
Others ~30k tok
DjinnBot ~800 tok
1 call to code_graph_impact vs. codebase-wide grep + read
30 tool schemas in the system prompt
12x
Others ~18k tok
DjinnBot ~1.5k tok
Compact Python signatures via Programmatic Tool Calling
Read 5 files, grep, aggregate results
24x
Others ~12k tok
DjinnBot ~500 tok
1 exec_code call — intermediate results stay in Python
Analyze a 500-line diff for security issues
13x
Others ~4k tok
DjinnBot ~300 tok
focused_analysis delegates to a sub-model
Code Knowledge Graph
Tree-sitter parses every source file into a graph of functions, classes, call chains, and functional clusters stored in KuzuDB. Agents query the graph instead of reading files. One call returns what 15+ file reads would piece together.
Programmatic Tool Calling
Instead of 30 full JSON schemas in every prompt, agents get compact Python function signatures and write code that calls tools, loops, and aggregates. Only the final result enters the context window.
Focused Analysis
When an agent needs to analyze a large diff or spec, focused_analysis delegates to a fast sub-model. The agent's context stays clean for high-level reasoning.

How It Works

1. Define the work
Describe what you need via the dashboard's guided onboarding, chat, or API. Software projects, research tasks, content campaigns, operations workflows — anything.
2. Plan it
The planning pipeline decomposes your project into tasks on a kanban board with priorities, dependencies, and hour estimates. Or define tasks manually.
3. Agents claim
Each agent watches specific board columns matching their role. Engineers grab implementation work. Reviewers grab review tasks. Any agent can be configured to watch any column.
4. Autonomous work
On pulse cycles, agents wake up, claim a task, spin up an isolated container, and do the work — writing code, researching topics, generating content, browsing the web, or running any tools you've given them. Use swarm execution for parallel multi-agent processing.
5. Review & iterate
Agents review each other's work. If changes are needed, the task cycles back. They coordinate via inbox messages and can wake each other for urgent blockers.
6. Deliver
Watch the whole thing happen in real-time via the dashboard, Slack, Discord, Telegram, WhatsApp, Signal, CLI, or the live activity feed.

The Default Team

Eric Product Owner Requirements, user stories, acceptance criteria, scope management
Finn Solutions Architect System architecture, tech decisions, code review, API design
Shigeo UX Specialist User flows, design systems, component specs, accessibility
Yukihiro Senior SWE Implementation, bug fixes, writing production code
Chieko Test Engineer QA strategy, regression detection, test automation
Stas SRE Infrastructure, deployment, monitoring, incident response
Yang DevEx Specialist CI/CD pipelines, tooling, developer workflow optimization
Grace Executive Assistant Meeting transcripts, commitment tracking, relationship management
Holt Marketing & Sales Sales strategy, outreach, deal management, positioning
Luke SEO Specialist Content strategy, keyword research, technical SEO
Jim Finance Lead Budget, pricing, runway management, financial modeling

Each agent has a 100-200 line personality file with backstory, core beliefs, productive flaws, and anti-patterns. The default team covers engineering, ops, marketing, SEO, and finance — but you can create agents for any domain by adding a directory with a few markdown files.

Why Not the Alternatives?

DjinnBot Everyone Else
Setup
One curl command — 5 minutes
IDE extension install, or hours of framework wiring
Token Efficiency
12-40x reduction via code graph, PTC, focused delegation
Raw file reads and full JSON schemas in every prompt
Cost Visibility
Per-call, per-agent, per-user LLM usage logs with dollar amounts
None, or basic aggregate totals
Agents
11 specialized agents with rich personas, or create your own
One generic assistant, or build from scratch
Security
Container isolation, 2FA, encrypted secrets, auto SSL
Direct host access
Memory
Persistent semantic memory with 3D knowledge graph
Stateless, or basic file-based context
Collaboration
Agents review, critique, and coordinate via work ledger
Single agent, single perspective
Parallelism
Swarm execution on DAG-aware task graphs
Sequential only, or custom scheduling code
Autonomy
Agents work 24/7 on configurable pulse schedules
Requires human in the loop

DjinnBot is built for people who want autonomous AI teams working on real projects — software, research, content, ops, or anything else — not another chatbot, not another framework to wire together.