git:20261027.84f5ce5 to git:20261028.5ba1357

# AbstractCore

> Unified Python interface for cloud and local LLMs: generation, streaming, host-controlled tools, typed output, sessions, media, embeddings and OpenAI-compatible HTTP serving.

Supported Python: 3.9+; CI tests Python 3.9–3.13.

Use an explicit provider, exact model and endpoint. Local model loads are offline/cache-first;

download weights deliberately. Tools are pass-through by default: the host authorizes and

executes calls. For durable agent orchestration, use [AbstractRuntime](https://github.com/lpalbou/abstractruntime).

## Agent Skill

- [AbstractCore skill](skills/abstractcore/SKILL.md): portable operating procedure for agents writing integrations and piloting an installation; keep its references folder.

- [Python recipes](skills/abstractcore/references/integrations.md): sessions, typed extraction, streaming, async, allowlisted tools, embeddings and media.

- [Operating recipes](skills/abstractcore/references/operations.md): endpoint discovery, models, engines, serving, caches and diagnostics.

## Start Here

- [README](README.md): project overview and quick start.

- [Documentation index](docs/README.md): complete user documentation and reading paths.

- [Installation](docs/installation.md): three settings: light, apple and gpu; platform requirements.

- [Getting started](docs/getting-started.md): first provider call, sessions, tools and media.

- [Integration and operations guide](docs/agent-usage.md): explicit endpoints, version checks, bounded recipes and lifecycle decisions.

- [Prerequisites](docs/prerequisites.md): provider keys, base URLs and hardware.

- [API map](docs/api.md): public entry points, response types and error boundaries.

- [API reference](docs/api-reference.md): signatures and retry configuration.

- [Architecture](docs/architecture.md): components, ownership, request lifecycle and native MLX runtime.

- [FAQ](docs/faq.md): conceptual questions and supported limits.

- [Troubleshooting](docs/troubleshooting.md): installation, authentication, generation and model-load failures.

## Python Integrations

- [Tool calling](docs/tool-calling.md): native/prompted syntax, tool definitions and host execution.

- [Tool rewriting](docs/tool-syntax-rewriting.md): tags and agent_format for clients.

- [Structured output](docs/structured-output.md): response_model, validation and retry budgets.

- [Generation parameters](docs/generation-parameters.md): token/sampling defaults and caller overrides.

- [Reasoning](docs/reasoning-control.md): thinking controls, effort transports and effective metadata.

- [Sessions](docs/session.md): history and persistence.

- [Prompt caching](docs/prompt-caching.md): backend capability matrix, stable prefixes, hit metadata and measured limits.

- [Async](docs/async-guide.md): asynchronous generation and streaming.

- [Concurrency](docs/concurrency.md): async versus backend batching.

- [Media input](docs/media-handling-system.md): images, documents, audio/video policies and fallbacks.

- [Request and output](docs/request-output.md): canonical multimodal request/output contract.

- [Capabilities](docs/capabilities.md): voice, vision, music, 3D and camera plugins; supported boundaries; voices by engine and language, voice design, cloned-voice management.

- [Vision](docs/vision-capabilities.md): native image/video input and captioning fallbacks.

- [Glyph compression](docs/glyphs.md): vision-based document compression and tradeoffs.

- [Embeddings](docs/embeddings.md): local/remote vectors and the embedding.text route.

- [MCP](docs/mcp.md): discover tool servers over HTTP/stdio; dispatch in the host.

- [Examples](examples/README.md): runnable examples by task and a guided learning path.

## Operating the Host

- [Local models](docs/models.md): catalog/fit, exact artifacts, downloads, jobs, deletion and verification.

- [Knowledge-graph extraction models](docs/knowledge-graph-extraction-models.md): measured local vs hosted model comparison for ontology-constrained extraction, prompt rules and second-pass strategies.

- [Recommended models](docs/recommended-models.md): capability choices by platform and memory.

- [Engines](docs/engines.md): inventory, install plans and platform support.

- [Configuration](docs/centralized-config.md): persistent routes, endpoint profiles, credentials and timeouts.

- [Server](docs/server.md): multi-provider /v1 endpoints, authentication and control planes.

- [Endpoint](docs/endpoint.md): one provider/model per HTTP worker.

- [Web console](docs/console.md): model/engine/config management through the browser.

- [Terminal console](docs/console-tui.md): terminal host management.

- [Chat CLI](docs/acore-cli.md): interactive conversations.

- [Memory and residency](docs/memory-management.md): loaded models, ownership, unload and memory accounting.

- [Memory blocs](docs/memory-blocs.md): text snapshots and native KV artifacts with live-binding checks.

- [Cancellation](docs/generation-cancel.md): cancel_event, per-lane support and safe eject.

- [HF compatibility](docs/huggingface-model-compatibility.md): Transformers/GGUF artifacts and loading limits.

- [Speculation](docs/speculative-decoding.md): native MTP controls and checkpoint support.

- [Native MLX](docs/native-mlx-runtime.md): opt-in scheduling, MTP, caching and cancellation.

- [MLX benchmarks](docs/native-mlx-benchmarks.md): measurements and comparison scope.

- [Fallbacks](docs/fallbacks.md): explicit compatibility behavior; no model-artifact substitution.

- [Events](docs/events.md): observability hooks.

- [Logging](docs/structured-logging.md): log configuration.

- [Tracing](docs/interaction-tracing.md): opt-in prompt/response/usage capture.

- [Email](docs/email.md): account configuration, Read only or Read and write access (`EmailAccount.access`, `set_access`), mailbox verbs, recipient policy and read/send tools.

## Contributing and Project Information

- [Contributing](CONTRIBUTING.md): source setup, focused checks and contribution conventions.

- [Changelog](CHANGELOG.md): version history and compatibility notes.

- [Security](SECURITY.md): vulnerability reporting.

- [Acknowledgements](ACKNOWLEDGEMENTS.md): upstream projects and communities.

- [License](LICENSE): MIT license.

- [Package metadata](pyproject.toml): install settings, dependencies and console scripts.

- [Expanded documentation](llms-full.txt): canonical user pages in one file; the sources remain authoritative.