Makion is a governed AI agent for SAP — it doesn't just chat about your systems, it works inside them: building ABAP, supporting incidents read-only, and producing consulting deliverables — safely, under a full audit trail, on your own infrastructure and your own AI.
Most AI tooling edits text and hopes for the best. Makion holds real tool access to your live SAP — with the guardrails engineered in from the first line.
Makion works through real protocols — it speaks SAP's official ADT interfaces to lock objects, write and activate code, create dictionary objects and carry them in transports. Work that autocomplete tools physically cannot do.
Guardrails live in code, not prompts: read-only roles are denied write tools at the server's gate, object deletion is disabled unconditionally, unattended runs pass a two-human gate, and every call lands in an audit trail.
Everything deploys on your servers. You connect the AI you already pay for — Claude, Gemini, Codex, Kimi, Grok or DeepSeek, by subscription or API key — so your code goes only to your own AI account, Makion never sees it, and there is no per-token metering from us. Makion is verified end-to-end on Claude — the experience we recommend today; other models are connectable on your own account, with results that vary by model.
The AI ABAP developer for your SAP.
Describe a task in plain language. Makion's agent works inside your own SAP system through the official ADT interfaces — reads the live system, writes and activates ABAP, creates DDIC objects and transports, then runs syntax checks, unit tests and ATC. Including on-prem and ECC, where cloud AI tooling doesn't reach.
Task chat. Describe the task, pick a role, send — the agent works in your SAP project's context with real ADT tools. Attach code, screenshots or files.
Chat is the interface; ADT is the engine. The agent inspects your live system — source, structures, where-used, SQL — then does the work and reports with proof.
Each role is a full ABAP methodology your architect edits in plain language — and read-only roles physically cannot write to SAP.
A rule-based style normalizer runs before activation: casing, indentation, spacing, modernization rules per ABAP release. Style only — never logic.
Unattended runs are structurally read-only: the agent only proposes changes into a ship queue. A human approves each one; a separate human-triggered Apply performs the write.
Configure the server in plain language — projects, access, user roles, base rules. Changes apply immediately; destructive actions are confirmed first. Config only, no shell.
A Telegram bot puts projects, roles, tasks and files in your pocket — kick off work and collect results from your phone while the server runs at home base.
SAP's Joule is cloud-gated — most of the SAP install base can't use it. Makion works through official ADT against your own system — S/4HANA and NetWeaver-based on-prem (incl. ECC on 7.40+). Verified by us on S/4HANA. We confirm reach on your system during the pilot.
Connect the AI account you already have — Claude, Gemini, Codex, Kimi, Grok or DeepSeek — with your API key or subscription, where the provider permits. Makion doesn't resell model access and doesn't charge for usage: what you pay us doesn't move with how much you run. Verified on Claude; other models are optional and results vary by model.
Unattended agents can propose but never apply. Object deletion is disabled for every role and every mode, unconditionally — a code gate, not a promise.
Curated ABAP knowledge packs — real-world patterns mined from years of production code, cleaned, anonymized and reviewed — plus your team's standards, written once in plain language.
Nothing irreversible happens without a human decision; unattended runs are hard-gated.
The full product story — features, safety architecture, demo and pilot options — lives on the Makion product site.
Prefer email? Write to contact@makion.dev — we read everything.