The product

Kodzonia from the inside.

A desktop application, hexagonal architecture, no mandatory server and no account. What follows is what exists in the code today.

A project is a file

Conversations, memories, encrypted credentials, subagents, workflows and the audit trail live in a single file you move, copy and version. No server, no account, no lock-in by default.

Twenty data sets in one file, migrated automatically on open
A document type registered with the operating system: opening it from the file manager works
Credentials encrypted with a key protected by the system keychain — the file from another machine does not decrypt

110 tools, 22 categories

Files, shell, Python, git, semantic code search, a driven browser, databases, scheduling and MCP clients. Every tool declares a stable name, a description and an impact level.

A core that is always advertised plus groups loaded on demand
Discovery by intent and eviction by LRU, solving the 128-tool-per-request ceiling
Deletion is reversible and every edit creates a restore point before the first mutation

Subagents that cannot escape their scope

Five factory specialists with written personas, ad-hoc specialists and a Specialist Builder that generates the specialist by analyzing the repository.

Maximum depth, a subset of tools and an inherited sandbox never wider than the parent's
A destructive brake that carries provenance: the child cannot escalate privilege
Background execution with exactly-once wake-up

Deterministic workflows

A script orchestrates subagents in phases. What introduces variation is forbidden inside the engine, with a named error — not a convention, a barrier.

A fresh engine per run, strict mode by default
A data boundary of plain JSON, with no reflection
A clarification channel with two-tier deadlines: the subagent asks instead of inventing

Persistent, reviewable memory

Seven memory kinds, a stable slug, a confidence level and declared provenance. Two operating modes: curated, where the agent proposes and you approve, or no limits.

Revertible versioning, a trash bin and a graph of typed relations
Hybrid search: full text plus local embeddings, fused by reciprocal ranking
Nothing leaves the machine to be indexed: the embeddings run locally

Multi-model by architecture

Six provider adapters and multiple accounts in the same project, with routing by model family. The user picks a model, never an endpoint.

Anthropic, OpenAI, Kimi and our own gateway, side by side
Switch provider without rewriting anything in the project
Every call audited with model, tokens, cache, duration and error

Security that is a program

A per-session canary with a deterministic guardrail in the engine, which aborts the stream before the leak. Tool results always arrive marked as untrusted.

A versioned corpus of 12 attacks, running against a real provider
A closed list of destinations for state-changing requests
Executable hooks authorized by hash, revalidated on every start

Semantic code intelligence

Real analysis, not regular expressions: a compiler for C# and language servers for TypeScript, Python, Rust, C and C++, Go, PHP, Swift, Lua, Zig and Pascal.

Scope-aware symbol renaming
An explicit refusal to produce a partial changeset while the index is not ready
Operational honesty by default: what could not be extracted is reported, not hidden

Want to see it running?

We show the product live, on your kind of code, and open the audit dashboard in front of you.

Talk to us