Skip to content

Capabilities and limits

This page states the current Python 0.3 MVP boundary. Labs experiments are listed separately and are not implied by the PyPI package.

Runtime

Capability Status Boundary
Agents and multi-turn model execution Included One provider per agent
Typed tools Included JSON-schema-compatible Python signatures
Permissions and approvals Included Runtime-enforced, deny when a declared permission is not granted
Cancellation Included Active runtime task and provider stream
Handoffs Included Targets declared in one Application
Workflows Included In-process deterministic runner and checkpoints
Artifacts Included In-memory or SQLite artifact store
Traces and usage Included Default tracer is in memory

Providers

Provider SDK path Text streaming Tool calls Usage Credential-backed qualification
FakeModel Core Yes Yes Deterministic Offline
OpenAI Official Responses API Yes Yes Yes Required per deployment account
Anthropic Official Messages API Yes Yes Yes Required per deployment account
Google Official Gen AI SDK Yes Yes Yes Required per deployment account

Adapter contract tests run without keys or network calls. They validate request translation, stream handling, complete tool arguments, usage, cancellation behavior, and safe errors. A real deployment must verify model access, region, quota, billing, safety policy, and data handling in its provider accounts.

Structured output, vision, audio, reasoning-specific controls, and embeddings are not declared as portable 0.3 capabilities.

Persistence

Store Status Boundary
In-memory sessions and artifacts Included One process lifetime
SQLite sessions, events, approvals, artifacts, and idempotency Included One active runtime process per file, modest writes
PostgreSQL Not included Implement behind public store protocols
Durable trace backend Not included Export or replace the default tracer

SQLite uses WAL, foreign keys, schema migration, bounded busy retries, atomic event sequences, and interrupted-run recovery. It is not encrypted by AgentMuru.

Server and Workspace

The package includes the FastAPI HTTP API, session WebSocket replay, bundled browser Workspace, trusted-host middleware, explicit HTTP and WebSocket origin controls, authentication protocols, session ownership, payload limits, and security headers.

The default anonymous configuration is for local use. AgentMuru does not ship a production user directory, identity provider, TLS terminator, tool sandbox, distributed scheduler, or managed deployment control plane.

Databricks

Databricks SDK, SQL, and Unity Catalog helpers are optional. Offline tests verify configuration, serialization, identity isolation, and input safety. Credential-backed workspace behavior must be qualified in the target account.

Labs

Go-native compilation, machine profiling, adaptive action routing, and local-model catalog work remain under Labs. They are not part of the PyPI 0.3.0 MVP or its support claim.