AI Mode¶

Experimental in v0.7.0

Provider-backed prompt execution, AI Console and flows, planner quality, investigations, code/patch generation, autonomous behavior, and Studio AI internals are experimental. Session and budget state may be process-local.

Aksara includes useful experimental primitives for provider calls, bounded prompt-pack execution, plan data structures, project context, declarative patches, and query-plan validation. These APIs are documented so developers can experiment with what the package actually exports; their presence does not make them part of the stable v0.7 contract.

Stable MCP is separate¶

Generated MCP tool execution is retained as a stable v0.7 boundary and does not require provider-backed AI:

Path	Purpose
`/mcp/`	official-client Streamable HTTP protocol endpoint
`/ai/tools/mcp`	HTTP JSON inspection catalog for generated tool metadata

Start with the MCP quickstart for the model, Principal, official client, and persisted database journey.

Current experimental entry points¶

Goal	Current documentation
Configure provider-backed prompts	AI Hub, Providers, Bring Your Own LLM
Run bounded prompt packs	Execution runtime
Describe and execute deterministic plans	Planner primitives
Build project context	Context engine
Validate query plans	Query engine
Describe code or patch operations	Code generation, Patch engine, Safety
Explore Studio AI	Studio

There is no public AgentRuntime or Planner class in v0.7.0. Pages with those historical names now describe the narrower real exports or label conceptual architecture as pseudocode.

Provider configuration¶

Use AI Hub for the current experimental provider path:

aksara ai-hub configure openai
aksara ai-hub status
aksara ai-hub doctor

Provider credentials stay in environment variables or application secret storage. Compatibility profile fields remain available for existing v0.6 code, but new setup should not use an AKSARA = {...} dictionary.

Safety boundary¶

Experimental helpers do not grant authority. Applications still resolve the Principal, authorize operations, constrain writable fields and tenants, and own human-review workflow and durable external side-effect policy. Review AI/MCP security boundaries before connecting agents to data-changing operations.