https://github.com/selfagency/agentsy
🤖 WIP agentic development framework
https://github.com/selfagency/agentsy
agentic agentic-ai agentic-framework agentic-workflow agents ai anthropic chatbot chatgpt json llm mistral models ollama openai parser xml
Last synced: about 2 months ago
JSON representation
🤖 WIP agentic development framework
- Host: GitHub
- URL: https://github.com/selfagency/agentsy
- Owner: selfagency
- License: mit
- Created: 2026-03-11T17:13:58.000Z (4 months ago)
- Default Branch: main
- Last Pushed: 2026-05-03T23:47:14.000Z (about 2 months ago)
- Last Synced: 2026-05-04T00:23:32.887Z (about 2 months ago)
- Topics: agentic, agentic-ai, agentic-framework, agentic-workflow, agents, ai, anthropic, chatbot, chatgpt, json, llm, mistral, models, ollama, openai, parser, xml
- Language: TypeScript
- Homepage: http://llmstreamparser.self.agency/
- Size: 1.13 MB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- Changelog: CHANGELOG.md
- License: LICENSE.md
Awesome Lists containing this project
README
# Agentsy
Composable infrastructure for LLM systems.
Build headless agentic workflows in Node.js-compatible runtimes, plug in only the layers you need, and stay aligned with open standards like MCP and AG-UI instead of a captive vendor stack.
[](https://github.com/selfagency/agentsy/actions/workflows/tests.yml) [](https://codecov.io/gh/selfagency/agentsy) [](https://app.codacy.com/gh/selfagency/agentsy/dashboard?utm_source=gh&utm_medium=referral&utm_content=&utm_campaign=Badge_grade) [](https://sonarcloud.io/summary/new_code?id=selfagency_llm-stream-parser) [](LICENSE.md)
## Why choose Agentsy
- **Streaming-first by default**: handle chunk boundaries, partial outputs, tool-call deltas, and finish states as first-class concerns.
- **Headless-first architecture**: ship robust automation, CLI tooling, and agent workflows before committing to a UI surface.
- **Composable package ecosystem**: adopt one package or a full pipeline — no all-or-nothing framework tax.
- **Type-safe and defensive**: strict TypeScript contracts and recoverable parsing behavior for untrusted model output.
- **Open ecosystem alignment**: practical interoperability with standards and protocols such as MCP and AG-UI.
## Already used in production
Agentsy already powers third-party model support in VS Code extensions for GitHub Copilot Chat:
- [Opilot](https://marketplace.visualstudio.com/items?itemName=selfagency.opilot) — Ollama models with local workflows, tool calling, vision, and streaming
- [Z.ai for Copilot](https://marketplace.visualstudio.com/items?itemName=selfagency.z-models-vscode) — Z.ai coding models with streaming, tool calling, and MCP-assisted capabilities
- [Mistral for Copilot](https://marketplace.visualstudio.com/items?itemName=selfagency.mistral-models-vscode) — Mistral AI models with streaming, tool calling, and vision support
## Example workflows
- [CLI log summarizer (easy)](./docs/examples/cli-log-summarizer.md) — stream model output into terminal-ready summaries with minimal wiring.
- [DNS blocklist workflow](./docs/examples/dns-blocklist.md) — ingest multiple logs, stream model output, validate schema, and update remote DNS when a source IP should be blocked.
- [Multi-provider policy gate](./docs/examples/multi-provider-policy-gate.md) — switch providers while keeping one schema-gated automation contract.
- [Agent tool loop with retries + continuation](./docs/examples/tool-loop-retries-continuation.md) — run a multi-step tool loop, retry transient failures, and resume interrupted streams.
- [Stateful ops copilot backend](./docs/examples/stateful-ops-copilot.md) — project stream events into state and AG-UI clients with continuation recovery.
- [All-tooling end-to-end workflow](./docs/examples/all-tooling-end-to-end.md) — combine stream ingestion, normalization, processing, tool loops, recovery, state projection, AG-UI conversion, rendering, and action gating in one flow.
## Package ecosystem
This repository is a **pnpm workspace monorepo** orchestrated with **Turborepo** and organized as focused `@agentsy/*` packages.
### Integration
| Package | Role | Status |
| ---------------------------------------------- | ------------------------------------------------------------------------- | --------- |
| [`@agentsy/vscode`](./docs/packages/vscode.md) | VS Code chat-provider, rendering, settings, and MCP integration utilities | Published |
### Runtime and orchestration
| Package | Role | Status |
| -------------------------------------------------------- | ---------------------------------------------------------------- | --------- |
| [`@agentsy/processor`](./docs/packages/processor.md) | Event-driven stream orchestrator and transform pipeline | Published |
| [`@agentsy/normalizers`](./docs/packages/normalizers.md) | Provider-specific normalization into a common stream shape | Published |
| [`@agentsy/agent`](./docs/packages/agent.md) | Multi-step agent loop and stop-condition helpers | Published |
| [`@agentsy/adapters`](./docs/packages/adapters.md) | Integration-oriented wrappers around stream-processing pipelines | Published |
| [`@agentsy/renderers`](./docs/packages/renderers.md) | Plain-text rendering plus shared renderer primitives | Published |
### Core parsing and shaping utilities
| Package | Role | Status |
| ------------------------------------------------------ | ---------------------------------------------------------- | --------- |
| [`@agentsy/thinking`](./docs/packages/thinking.md) | Incremental `` content extraction | Published |
| [`@agentsy/tool-calls`](./docs/packages/tool-calls.md) | XML and native tool-call helpers | Published |
| [`@agentsy/structured`](./docs/packages/structured.md) | JSON parsing, repair, validation, and streaming helpers | Published |
| [`@agentsy/context`](./docs/packages/context.md) | Context extraction, dedupe, and normalization helpers | Published |
| [`@agentsy/formatting`](./docs/packages/formatting.md) | Display-focused text sanitization and formatting utilities | Published |
| [`@agentsy/recovery`](./docs/packages/recovery.md) | Recovery snapshots and continuation prompt helpers | Published |
| [`@agentsy/xml-filter`](./docs/packages/xml-filter.md) | XML tag filtering and privacy-oriented scrubbing utilities | Published |
| [`@agentsy/sse`](./docs/packages/sse.md) | Server-sent-event parsing utilities | Published |
| [`@agentsy/types`](./docs/packages/types.md) | Shared type contracts across the ecosystem | Published |
### State and protocol
| Package | Role | Status |
| -------------------------------------------- | ------------------------------------------------------ | --------- |
| [`@agentsy/ui`](./docs/packages/ui.md) | Conversation-state store and processor binding helpers | Published |
| [`@agentsy/ag-ui`](./docs/packages/ag-ui.md) | AG-UI protocol bridge utilities | Published |
### Private package
| Package | Role | Status |
| -------------------------------------------------------- | -------------------------------------- | ------- |
| [`@agentsy/integration`](./docs/packages/integration.md) | Cross-package integration test harness | Private |
## Pick your adoption path
- **I need robust Node streaming pipelines** → start with [`@agentsy/normalizers`](./docs/packages/normalizers.md) + [`@agentsy/processor`](./docs/packages/processor.md)
- **I need JSON schema gates before automation** → add [`@agentsy/structured`](./docs/packages/structured.md)
- **I need multi-step agent behavior** → add [`@agentsy/agent`](./docs/packages/agent.md)
- **I need VS Code integration utilities** → add [`@agentsy/vscode`](./docs/packages/vscode.md)
- **I need protocol/state helpers** → add [`@agentsy/ag-ui`](./docs/packages/ag-ui.md), [`@agentsy/ui`](./docs/packages/ui.md), and focused utility packages
If you prefer a thinner integration layer, start with `@agentsy/normalizers` + `@agentsy/processor`, then add `@agentsy/structured` once you want strict schema gates before automated actions.
## Documentation map
- [Documentation home](./docs/index.md)
- [Getting started](./docs/getting-started.md)
- [Why Agentsy](./docs/why-agentsy.md)
- [Examples](./docs/examples/index.md)
- [API index](./docs/api.md)
- [Package catalog](./docs/packages.md)
- [Roadmap](./docs/roadmap.md)
- [Developer guide](./docs/developers/index.md)
Deep architecture material lives under [`docs/architecture/`](./docs/architecture/index.md), while contributor and automation specifics remain in the developer docs and repository instructions.
### Monorepo commands
```bash
pnpm install
pnpm build
pnpm check-types
pnpm test
```
## License
[MIT](LICENSE.md)