AI Mode configuration¶
Experimental
Provider-backed AI, Planner behavior, investigation sessions, code patches, memory, and Studio AI internals are experimental in v0.7.0.
Aksara configuration uses the global aksara.conf.settings object. Do not add
an AKSARA = {...} dictionary; the runtime does not read that pattern.
Enable provider-backed AI explicitly:
Then configure a provider with the current AI Hub path:
See AI providers for current, compatibility, and deprecated provider layers.
Runtime limits¶
The real prompt-pack runtime accepts in-process limits per call:
from aksara.ai.limits import AgentRuntimeLimits
from aksara.ai.runtime import run_prompt_pack
result = await run_prompt_pack(
prompt_pack,
limits=AgentRuntimeLimits(
run_timeout_seconds=30,
provider_timeout_seconds=20,
token_budget=2_000,
),
)
These limits reset with the process. There is no AI_AGENT_RUNTIME dictionary,
stable AgentRuntime class, or durable provider budget in v0.7.0.
MCP is separate¶
MCP tool execution is a stable boundary retained in v0.7 and does not require a model provider. Enable its Streamable HTTP server separately:
MCP clients connect to /mcp/. The /ai/tools/mcp route is an HTTP JSON
inspection catalog. Follow the MCP quickstart to
add server-side Principal resolution before enabling tool execution.