Roadmap
No marketing dates, no vaporware. This is the real state of the build — what's shipped, what's next, and what's vision. Updated as the code moves.
The foundation of the language. Lexer, recursive descent parser, AST tree-walking interpreter, and all the primitives that every phase above builds on.
.png){kind=logo}
string number bool null list dict** power · % moduloif/else if/else · for · while · matchtry/catch/raiseprint() · typeof() · line and block commentsaxon run · axon check · axon versionThe core of the Cognitive Flow Oriented paradigm. Agents, tasks, flows, the pipeline operator, and the full actor lifecycle managed by the C++17 runtime.
flow {} · agent {} · task {} · main {}Agent(input) · Agent->Task(input) · >> pipelinetask_graph {} with on_success and on_fail per edgevalidation {} — confidence threshold, max retries, on_fail behaviorparallel {} — real multi-core execution via C++17 thread poolcompute = auto/gpu/cpu · gpu_fallback = truenum_gpu · num_thread · f16_kv · low_vramdelay · timeout · sleep() per task, agent and flowThe explicit memory model. Every actor declares what it remembers and for how long. Four scopes, TTL support, disk persistence, and native methods for strings, numbers and lists.
runtime · session · shared · persistentmemory->store() · get() · has() · remove() · clear()memory->store(k, v, "runtime", 30)memory->persist(path) · memory->load(path)Search by meaning, not by exact text. A native HNSW index with Ollama embeddings that persists as a single portable file — no external vector database needed.
embedding {} — configure model, provider and dimensionsmemory->index(text, metadata) — index content as a vectormemory->search(query, top_k) — search by meaning, cosine similaritymemory->sem_persist(path) · sem_load(path) · sem_size()Native file operations with automatic type detection by extension. CSV and JSON parsers, date utilities, and regex support built directly into string methods.
file->read() auto-detects format by extensionfile->write() · append() · delete() · mkdir() · ls() · exists()csv->parse() · stringify() · write() · custom delimiterjson->parse() · stringify() · write() · configurable indentfind() · find_all() · replace() · split() · test()date-> — now() · format() · add() · diff() · before() · after()Native reading of Word and Excel documents, image manipulation, encoding utilities and cryptographic hashing — all without external runtime dependencies.
file->read(".docx") — plain text from Word documentsfile->read(".xlsx") — List of Dicts from Excel with auto type detectionimage->save() · resize() · info() · convert()codec->b64_encode/decode() · hex_encode/decode() · url_encode/decode()hash->sha256() · hash->md5()Two completely separate extraction paths for PDFs — text and images. Each optimized for its purpose, without one breaking the other.
file->read(".pdf") — plain text, simple modepdf->parse(path, "t") — structured object with text and pagespdf->parse(path, "i") — image extraction as base64Text extraction from images via Tesseract OCR, character-level string access with negative index support, and automatic detection of optional tools at configure time.
string[i] — character access by index including negativesocr->read(path, lang) · ocr->installed() · ocr->languages()The import system, multi-file architecture, and first-class functional programming. Functions as values, closures, recursion, higher-order functions, and the HTTP module for consuming external APIs.
import "./file.axon" — load agents and flows from external files_prefixflow {} in a projectFlowName.input · FlowName.versionfunc name(params) { body }func(params) { body }(params) => exprhttp->get(url, headers) · http->post(url, body, headers)http->json(url) — GET + automatic JSON parseenv("VAR") — read system environment variablesThe second layer of the functional programming system. Variadic functions, named arguments, tuple unpacking in loops, and selective imports to control what gets exposed between files.
*args — variadic functions: func sum(*args) {}**kwargs — named arguments: func log(**kwargs) {}for key, val in dict — tuple unpackingimport "./f.axon" { fn1, fn2 }(n: number) => n * 2any type for untyped parametersSame actor syntax, same validation contract, different provider. OpenAI, Anthropic and Cohere will work exactly like Ollama — no new syntax to learn.
gpt-4o · gpt-4-turbo · gpt-3.5-turboclaude-sonnet · claude-opus · claude-haikucommand-r · embed-english-v3http-> via libcurlllava · gpt-4o · claude)Train models directly from an Axon flow. Define a fine-tuning job as a block, point it at a dataset, and get a model file ready to load as an agent — without leaving the language.
finetune {} — training job definition block.gguf — load directly as an Axon agentWhat an operating system does for processes on one machine, the mesh does for agents across machines. Encrypted peer-to-peer communication, dynamic topology, and the Consensus module for collective decisions.
The stable, production-ready release. Full tooling ecosystem, package distribution via native package managers, and the official VS Code extension.
winget · brew · aptaxon install package-nameaxon new · axon init · axon publish