Stop wiring agents to ten systems. Curate your business context once. Kartha ACP serves it to any AI agent via MCP — pre-joined, dimensionally structured, ready to reason on. One call instead of three. 85% fewer tokens. Same model, an order of magnitude smarter.
Every business function now has a purpose-built AI agent being sold to it. Each one is smart in isolation and blind to everything else. The sales agent doesn't know about the support tickets. The support agent doesn't know about the deal pipeline. This is the pre-Snowflake analytics world, replayed for agents.
ACP solves the integration problem once, ahead of time. Pipelines pull and normalize your data before any agent asks a question. The result is pre-built, structured context stored in Postgres — reusable by any agent, any framework, any number of times.
CRM, billing, support, ERP, spreadsheets, Postgres, REST APIs, GraphQL. Drop a CSV, point at an API, write a YAML mapping. The CLI auto-suggests mappings to the 7-dimension schema.
Every entity becomes a structured profile organized by what / how much / who / when / where / why / how. Pre-joined across systems, deep-merged on upsert, change-tracked. Numbers are numbers. Dates are dates. Identities are resolved.
MCP server for Claude Desktop, Cursor, Claude Code. REST API for everything else. CLI for humans. Changefeed for proactive agents. Skills for repeatable workflows. Same data, three doors.
Most data platforms dump raw fields into a bag and make the agent figure out what's a date, what's a number, who to contact. ACP organizes every entity into 7 semantic dimensions — so the agent navigates structured context, not a search problem.
Core identity — name, status, type, segment.
Numbers to reason about — KPIs, scores, counts.
People to contact — owners, champions, signers.
Dates that drive urgency — renewals, SLAs, activity.
Geography, territory, timezone, jurisdiction.
Strategy and risk — churn signals, expansion potential.
Workflow state — stage, onboarding, support tier.
An agent assessing risk goes straight to measures
and temporals. An agent finding who to contact reads
actors. No prompt engineering about data structure. No system-specific field
names. The schema does the translation once — every agent, every framework, benefits.
MCP server ships over Streamable HTTP transport. Five tools are all an agent ever needs: read entities, search them, follow the changefeed, list activity, and write decisions back.
Full context profile for one entity, with the last 10 transactions. One call instead of three.
Find entities by JSONB filter (eq, gt, gte, lt, lte, contains) or text. Pre-joined results.
Activity history for an entity — assessments, status changes, agent decisions. Indexed by time + type.
The changefeed. Polling agents call this to react to data changes since their last cursor.
Agents write back: risk assessments, escalations, recommendations. Visible to all other agents on next poll.
Same data, three doors. REST for non-MCP agents. CLI for humans and CI. Changefeed for autonomous loops.
Your team shouldn't have to change frameworks to adopt ACP. MCP-native clients
auto-discover the five tools. Non-MCP frameworks use the acp
CLI as a subprocess or call the REST API directly. Managed agent platforms like
Bedrock and Vertex point at the REST endpoint.
/mcp endpoint. Tools auto-discovered. Zero integration code — just edit the MCP config.
acp CLI as a subprocess. Clean JSON on stdout. CLI handles the MCP handshake — the agent never sees JSON-RPC or SSE.
/v1/objects endpoint. Works with any platform that can make API calls. Ideal for managed cloud agent services.
A skill is a prompt template + metadata mapped to an APQC business process. Your agent runtime reads it and follows the workflow. No vendor lock-in. No new login. Edit the prompt to customize. Five ship today. Hundreds are possible.
General-purpose business analyst prompt. Paste into a Claude Project and start exploring data immediately.
Polls the changefeed, classifies risk against thresholds, records risk_assessed transactions.
Flags stale deals, missing context, overvalued pipeline. Records deal_risk_assessed.
Groups overdue invoices by aging bracket, cross-references customer health for prioritization.
Context-aware escalation using customer ARR, health score, and renewal proximity — not just SLA timers.
Scores vendors on delivery, quality, commercial, and relationship dimensions. Flags renewals.
Pre-curated context isn't just convenient. It's measurably cheaper, faster, and more accurate. An LLM reasoning over 800 tokens of clean data makes better decisions than the same LLM reasoning over 5,300 tokens of raw API noise.
Agent calls 3 raw MCP connectors. Gets HubSpot objects, Stripe blobs, Zendesk audit trails. Spends 80% of effort parsing and correlating.
Agent calls get_entity. Gets a pre-joined, typed, dimensioned profile. Spends 100% of effort on the actual question.
The platform is schema-agnostic — it stores any JSONB context, deep-merges it, tracks every change. The CLI is a pure HTTP client; the API and MCP server share the same core library and database pool. No HTTP hop between them. No surprises.
ACP is open-source at the core and self-hosted by default — the same deployment model as Postgres, Kubernetes, and Terraform. Run it on your laptop in Docker, ship it to production on Kubernetes. Commercial support, compliance tier, and managed cloud available when your governance team is ready.
Docker Compose for development. Kubernetes (Helm), AWS CDK, GCP, Azure, or Terraform for production. Bring your own managed Postgres. Deploy inside existing VPCs with standard service-mesh and ingress patterns.
Architected for SOC 2 and HIPAA alignment from day one. Audit log of every agent read and write. Field-level PII tagging and agent-scoped ACLs on the near roadmap. BAA available through Kartha Cloud for healthcare deployments.
Designed for change control, not migrations. Schema evolution is a YAML
config change — no ALTER TABLE, no downtime.
OpenTelemetry tracing and Prometheus metrics out of the box. Continuous backup
to your S3 or GCS bucket.
git clone to "Claude is reasoning over my data" in 10 minutes.
Three steps. No accounts, no API keys, no waitlist. Postgres + REST + MCP runs in Docker
on your laptop. The acp CLI links itself to your shell.
Installs deps, builds packages, brings up Postgres + API + MCP in Docker, links the CLI globally.
20 customers, 31 invoices, 14 cases, 8 vendors. Five accounts have rich risk stories worth reasoning about.
Auto-finds your Claude config and adds the MCP server entry. Restart Claude. Ask anything.
Then load your own data. Drop a CSV in data/,
run acp connect add csv for auto-mapping,
acp connect sync to validate and load.
Pipelines are YAML — version-control them with your project.
CrewAI, LangGraph, AutoGen, Mastra, Claude, GPT — all reason well. The wall every one of them hits is the same: where does the agent get good business context?
get_entity call. Pre-joined by the pipeline. 7-dimension structured profile, 800 tokens.The 7-dimension schema is domain-agnostic. A patient is a customer with different fields. An encounter is a case with different fields. A claim is an invoice. The platform doesn't change — only the templates and skills do.
Customers, invoices, deals, vendors, contracts, employees.
Patient, encounter, claim, provider, medication. FHIR-ready.
Matter, client, filing, billing entry, engagement.
Work order, inspection, part, supplier, production run.
Start free with the full-featured open source platform. Self-host on your infrastructure. Move to Kartha Cloud when you want managed Postgres, RBAC, and shared context across teams.
Self-hosted, fully featured. Run on your laptop or your cluster.
Production deployment with architecture review, compliance support, and named technical contact.
Managed ACP for teams. Shared context across people and agents.