Grinta Agent Engine
August 12, 2026 ยท View on GitHub
Grinta ships one production agent engine: the Orchestrator. Browser, MCP, debugger, LSP, memory, and editing features are capabilities exposed to that engine; they are not separate agents.
Runtime model
user input -> SessionOrchestrator -> Orchestrator -> Planner -> tools
<- observations / state / validation / completion evidence
The engine proposes tool calls and consumes observations. The orchestration layer owns lifecycle, confirmation, middleware, retries, pending actions, and completion policy. The execution layer performs approved actions.
Main implementation
backend/engine/orchestrator.pycoordinates a model turn.backend/engine/orchestrator_helpers/contains step, recovery, prompt, and condensation helpers.backend/engine/planner.pybuilds and mode-filters the model toolset.backend/engine/executor.pyandexecutor_mixins/process streaming and non-streaming model responses.backend/engine/function_calling/validates and dispatches tool calls.backend/engine/prompts/builds the system prompt from Python renderers and markdown partials.backend/engine/tools/defines native tools.backend/engine/tool_registry.pyvalidates the LLM-facing tool registry.
The exact toolset depends on configuration and interaction mode:
- Chat exposes read/discovery tools plus user questions.
- Plan adds task state, task tracking, and acceptance criteria.
- Agent exposes the configured execution surface, including edits, terminals, browser, debugger, memory, and MCP tools.
The canonical mode policy is backend/core/interaction_modes.py; toolset
construction is in backend/engine/planner.py.
Browser and MCP
Native browser execution lives under backend/execution/browser/, with the
model-facing tool in backend/engine/tools/browser_native.py. External MCP
servers are integrated under backend/integrations/mcp/ and reached through
the compact MCP gateway tool.
Extensibility boundary
The supported extension points are tools, MCP servers, provider/model configuration, and orchestration services. Grinta does not currently ship an Echo, Locator, Auditor, or other peer engine, and the repository does not document a stable third-party agent-class plugin API.
See ARCHITECTURE.md,
INFERENCE_AND_INTEGRATIONS.md, and
backend/engine/README.md.