Shakudo

Kaji reads, writes and acts across a company's tools through MCP integrations that an administrator adds once: ClickUp tasks, Notion pages and databases, Dremio and Supabase queries, Neo4j graphs, Mattermost channels, Gamma presentations and more. It picks the right integration from the request or uses the one a user names. KajiCore also runs bash commands and edits files inside its workspace, and the AI gateway turns an uploaded OpenAPI specification into MCP tools, so internal systems with an API become agent actions without custom code.

Sourcedocs.shakudo.io/kaji/kaji-integrationsread 2026-10-18

Subkajis split complex work into parallel child agents, each with its own context and workspace, that run independently and report back while the parent conversation tracks their progress. A successful multistep conversation can be saved as a workflow that captures its phases, steps and validation rules, and Kaji follows those phases and checks exit criteria when the workflow runs again. Underneath, an event bus on NATS JetStream spawns and manages a KajiCore process per conversation, with circuit breakers that stop loops and runs of consecutive errors.

Sourcedocs.shakudo.io/kaji/kaji-capabilitiesread 2026-10-18

Kaji grounds its answers in the company's own systems by querying them at run time: SQL against a Dremio data lake or Supabase, graph queries in Neo4j, Notion pages and wikis, Context7 for reference material and MarkItDown to convert documents, plus online research for current information. The platform can host vector databases and retrieval pipelines next to the data, and a team can build its own search over documents that way. Kaji does not come with a document index of its own, though. What it keeps between conversations is the memory graph of facts it has learned.

Sourcedocs.shakudo.io/kaji/kaji-integrationsread 2026-10-18

Before restarting a service, modifying production data or taking any other high risk action, Kaji stops and asks the right person on the team to approve. In the browser it hands control back when a login or approval step needs a person, and users can tell it to stop before a final form submission. KajiCore has a tool for asking the user questions mid task, and circuit breakers end runs that loop or keep failing. At the gateway, include and exclude lists strip PII and sensitive fields before data reaches a model.

Sourceshakudo.io/kajiread 2026-10-18

KajiChat users sign in through Keycloak single sign on connected to the company's existing identity provider, and role based access control is built in and linked into each component of the stack, so permissions follow users into the tools Kaji touches. MCP integrations are configured centrally by administrators. The AI gateway enforces organization standards on every model call, keeps an identity linked record of each configuration change, and strips PII from data before the model sees it. Shakudo is SOC 2 Type II certified and mitigates OWASP Top 10 LLM risks.

Sourceshakudo.ioread 2026-10-18

The AI gateway observes every model call and keeps a permanent, identity linked record of configuration changes, and the platform keeps audit trails and data lineage across the whole stack. Kaji shows what each job costs, and conversations, messages, tool runs and session state persist in PostgreSQL and object storage inside the customer's cluster, where KajiCore pods expose health and metrics endpoints. Subkaji progress is visible from the parent conversation as work runs.

Sourceshakudo.io/ai-gatewayread 2026-10-18

Kaji keeps a knowledge graph of entities, facts, relationships and episodes in FalkorDB, inside the customer's cluster. It saves key facts automatically after meaningful work and when it compacts a long conversation, and users can tell it to remember or search for something explicitly, such as when the fiscal year starts or the steps of a past database migration. Memory is shared and searchable across all of a user's conversations, so decisions and past work carry forward. Learned skills and session state persist separately in object storage.

Sourcedocs.shakudo.io/kaji/kaji-capabilitiesread 2026-10-18

Shakudo and Kaji deploy as Kubernetes workloads in the customer's own environment on AWS, Azure, Google Cloud, Oracle or on premises hardware, with no Shakudo infrastructure involved and no data leaving the customer's boundary. Air gapped deployments run the full AI and data stack with no outbound path, and workloads move between providers without rework. Prompts, files, memory and institutional knowledge all stay inside the perimeter, and local or private models can serve inference there.

Sourceshakudo.io/product/multi-cloud-deploymentread 2026-10-18

Built in skills ship preloaded with instructions for common tools and workflows, such as a Google skill for calendar events, a Neo4j skill for graph queries and a workspace cleanup skill, and a skills browser lists what is available. Teams save their own learned skills and workflows for reuse across conversations. The platform side adds a catalog of more than 230 integrations deployed in a few clicks and more than 30 documented use cases, from clinical documentation to DOT compliance records.

Sourcedocs.shakudo.io/kaji/kaji-capabilitiesread 2026-10-18

AutoKaji turns a plain language request into a recurring task on an hourly, daily or weekly schedule or a custom cron expression, reusing the same chat or starting fresh on each run, and automations can be paused, resumed or deleted at any time. The event bus routes @kaji mentions and prompt triggers to the right agent process, so work can start from a mention in a connected chat tool as well as from KajiChat. Kaji Cloud's hosted access is on a waitlist.

Sourcedocs.shakudo.io/kaji/kaji-capabilitiesread 2026-10-18

Model access runs through Shakudo's Envoy based AI gateway, deployed and configured by the customer, which routes to OpenAI, Anthropic, Gemini and other providers with priority based fallback between backends. An LLM router classifies each prompt as low, medium or high complexity and sends it to a model tier through virtual model names, so simple steps use cheaper models. Because the gateway runs in the customer's environment, its administrators decide which providers and models sit behind each tier.

Sourcedocs.shakudo.io/kaji/kaji-architectureread 2026-10-18

Kaji is reached through KajiChat and through mentions routed by its event bus. There is no public API, SDK or MCP server for driving Kaji, starting its tasks or managing its agents from another application, and the hosted Kaji Cloud that might offer one is still behind a waitlist.

Extension runs the other way: administrators add MCP integrations for Kaji to use, and the AI gateway converts OpenAPI specifications into tools. The wider platform deploys open source components that each bring their own interfaces, such as Airflow, Spark and Jupyter, but none of them is an interface to the agent itself.

Sourcedocs.shakudo.io/kaji/kaji-introread 2026-10-18

Saved workflows carry validation rules, and Kaji checks each phase's exit criteria as it runs them again, while multi agent workflows include built in verification of the work before it is handed back. These checks judge a single run as it happens. There is no suite of test tasks, expected results or scores for comparing one version of a skill, workflow or model against another before rolling it out.

Sourcedocs.shakudo.io/kaji/kaji-capabilitiesread 2026-10-18

Kaji launches a real browser and works websites itself: it goes to URLs, clicks links, waits for pages to load, fills forms, selects options, presses keys, reads content, takes screenshots and downloads files. A user can ask it to fill a form and stop before the final submit, or open a dashboard and capture the errors on screen, and when a login or approval step needs a person it hands the browser back temporarily before carrying on.

Sourcedocs.shakudo.io/kaji/kaji-capabilitiesread 2026-10-18