A thread is the unit of work, whatever carries it — Slack today, with Discord, Telegram and WhatsApp on the same shape. Slack rides a long-lived Socket Mode connection, so there is no public endpoint and no tunnel. DJINN reads the thread, and the answer stays attached to the question that caused it.
Company brain integrations
DJINN
Adjinntify your mind.
Think of this as an extension of your (or your company's) mind. DJINN gives you tools to literally make your digital wishes come true, swiftly, reliably and easily.
The descent
Watch one request fall through the company brain.
Architecture descent
A request moves through six production gates.
The path stays compact: ingress, memory, runtime, tools, verification, and handoff remain readable as DJINN moves through the company brain.
DJINN coreCompany brainmemory + tools + review
Every message is journalled to Postgres. A nightly pass promotes what is durable into an Obsidian vault — plain linked markdown you can open, read and edit yourself. Nothing is pinned into the prompt: DJINN retrieves only the notes a question actually needs, so context stays small and the answer stays cited. Talking to DJINN is what grows its memory.
Classification and scoring run on Haiku in batches. Reasoning and drafting run on Opus, and only for what survived. The split is why a nightly digest costs cents rather than dollars.
DJINN can only call what the registry lists, and most of what exists on the box is never wired to the model at all. Each exposed tool carries its own rule: run it, or stop and ask a person first. Writes go to a sandbox clone rather than the live tree, and anything that moves money or ships to production has no tool behind it — unreachable by construction, not by instruction.
Code written by one model is reviewed by another before it can merge, and a bare pass is not enough — the findings have to be clean. Work loops back for as many rounds as that takes.
Work comes back as something with a URL that outlives the conversation. Sending email, posting to a new channel, or shipping to production stays with a person — the tool layer enforces that rather than trusting the agent to remember.
Company memory
The vault grows itself.
One note per idea, linked to the notes it belongs with. A nightly pass files what is durable, connects what should have been connected, and heals what rotted. Nothing here was written by hand.Real topology from a live vault. Titles withheld — the shape is the point.
Capabilities
A platform, and the packs that run on it.
Each pack combines connectors, memory, approvals, monitoring, and a human-readable handoff in the channel you choose.Company Memory
A journalled message log, a linked note vault that grows from your own conversations, and cited search across both.
Connector Workers
Each connector has its own page and a Sigil UI/DSL to manage credentials, scopes, actions, and human boundaries.
Agent Runtime
Model routing by task, retries, tool boundaries, budget visibility, and an approval gate on anything that reaches the outside world.
Engineering Handoff
A tracked issue can become a branch, implementation, cross-model review, and pull request, or you can let DJINN go wild and free inside the boundaries you set.
Observability
Process health, recurring failures, and error signatures surface in your channel of choice, with optional automatic recovery, resilience, and cleanup procedures.
Spend Analytics
Where the model budget actually went, per project and per day, with the change worth making named and priced.
What holds it together
Every pack above reads and writes the same memory.
Connectors
Integrate the channels your company already works through.
Each connector has its own page and a Sigil UI/DSL for managing credentials, scopes, actions, and the human boundary.Browse every integration →Request channels
Route each request through the channel that already has the customer, operator, or team context.
CoreSlackCommand room
FinishingGmailEmail threads
FinishingWhatsAppCustomer chat
FinishingInstagram / MessengerSocial inbox
Work systems
Turn requests into repo work, tickets, deploy checks, documents, calendar actions, and database updates.
CoreLinearIssues
CoreGitHubCode + PRs
CoreVercelDeploys
FinishingGoogle WorkspaceDocs, Drive, Sheets
Business data
Pull analytics, ads, CRM, support, commerce, market, and vertical data into the same reviewed agent loop.
CoreSearch ConsoleSEO signals
FinishingMetaAds + leads
FinishingHubSpotCRM memory
FinishingStripe / ShopifyRevenue context
Connector catalog
Slack
Command room
Linear
Issues
Gmail
Google Calendar
Schedule
GitHub
Code
Vercel
Deploys
Supabase
Data
Community
Customer chat
Meta
Ads + APIs
Social DMs
Messenger
Page inbox
Discord
Community chat
Microsoft Teams
Enterprise chat
Telegram
Light bot
Search Console
SEO
Google Analytics
Analytics
Google Docs
Docs
Google Drive
Files
Google Sheets
Ops data
WordPress
CMS
Webflow
CMS
HubSpot
CRM + marketing
Salesforce
Enterprise CRM
Notion
Docs + wiki
Jira
Engineering
Zendesk
Support
Intercom
Support + sales
Stripe
Billing
Shopify
Commerce
Airtable
Ops database
Zapier
Fallback automation
Make
Fallback automation
Google Ads
Paid search
Mailchimp
Lifecycle email
Calendly
Scheduling
Asana
Project work
Trello
Project boards
ClickUp
Project work
Confluence
Knowledge base
X
Social monitoring
Named agents
DJINN Implementations
Beast
Trading agent
Signal monitoring and risk summaries on top of mechanical entry gates that re-quote the book at decision time and refuse a drifted or collapsed fill.Guardy
Agent-file integrity
Hashes every file that instructs an agent — skills, hooks, MCP configs, CLAUDE.md — then reports what appeared, what changed, and what reads as an instruction rather than a description.No public demo yet.
Aiwario
News intelligence
Reads a widening set of sources, tracks how one story moves across them over weeks, and checks each against what the company already knows — then says what changed and what to do about it.No public demo yet.
Toky
Spend analytics
Reads every model call the company has made and reports where the budget went — by project, by day, by model — then names the change worth making and prices it.No public demo yet.
Prospector
Lead engine
Sweeps public sources for founders blocked on engineering, scores each against the actual offer, drafts an opening message, and posts a queue to work through.Operator
Pipeline agent
Picks up where Prospector stops. Tracks every conversation that got a reply, chases the ones going cold, and keeps the deal state in the tracker instead of in somebody's head.Pushy
Observability agent
Service health, jobs, logs, deploy status, error signatures, and Slack-visible next steps.Dr I
Argentina legal services
Legal intake, case memory, document drafts, source-grounded answers, and lawyer review.Luxy
Conversational support
Customer context, support replies, escalation, account memory, and handoff trail.Cairn
JTreks implementation
A vertical company-brain implementation for JTreks operating memory and workflows.Wavey
Keonda implementation
A Keonda-specific DJINN implementation tuned to its workflow memory and team operations.Integration path
How the DJINN pipeline becomes your business's active brain.
Map the company brain
Choose the operating surfaces, source systems, Slack channels, owners, approval rules, and cost limits.
Wire the core
Wire integrations, memory, search, queues, model routing, audit trail, and deployment health checks.
Integrate service packs
Add legal, trading, support, deploy, observability, marketing, sales, hospitality, or custom loops.
Run and improve
Watch cost, failures, user requests, and repeated workflows. Package what works into reusable agents.
Open-source track
DJINN is becoming an open-source engine.
A clean-room Rust rewrite: every behaviour re-specified from the running service, none of it copied across. Five crates, some 14,500 lines, 360-plus tests. Every commit passes a gate that fails the build on eighteen patterns — a public IP, an email address, fifteen shapes of credential, or the name of the company the engine was extracted from.
Landed
- Journal and facts on SQLite, full-text search
- The vault: lint, heal, link graph, nightly rollups
- Three model backends, two families, routed per task
- The capability registry — run, ask, or refuse
- The dispatcher: durable jobs, one turn per thread
- The reply judge — it advises, it never blocks
Not yet
- The Slack transport
- A runnable example pack
- Published binaries
Private packs, customer data, credentials, transcripts, vault content and deployment details stay out. Nothing runs end to end yet, and the repo opens only once the sanitization audit and a human read are clean.
Clean-room RustMIT licensedSQLite defaultNot yet published
Request access
Integrate the first service pack into your company brain.
Start with one high-value loop: support, deploys, observability, legal, trading intel, sales follow-up, or a vertical implementation.