StateWeave

# StateWeave > `git` for agent brains — debug, time-travel, and migrate AI agent state across 10 frameworks. StateWeave lets you see exactly where your AI agent went wrong, rewind to any checkpoint, and optionally migrate between frameworks (LangGraph, MCP, CrewAI, AutoGen, DSPy, LlamaIndex, OpenAI Agents, Haystack, Letta, Semantic Kernel) — all while keeping accumulated knowledge intact. ## Install ``` pip install stateweave ``` ## Quick Start ```python from stateweave import LangGraphAdapter, MCPAdapter from stateweave.core.timetravel import CheckpointStore # Create a sample agent state (no frameworks needed) payload = LangGraphAdapter().create_sample_payload("my-agent") # Checkpoint it (like git commit for agent brains) store = CheckpointStore() store.checkpoint(payload, label="initial state") # Migrate to a different framework MCPAdapter().import_state(payload) # Modify state, checkpoint again, then diff payload.cognitive_state.working_memory["new_insight"] = "Q4 trending up" store.checkpoint(payload, label="after analysis") print(store.diff_versions("my-agent", 1, 2).to_report()) ``` ## One-Line Auto-Instrumentation ```python import stateweave stateweave.auto() # Auto-checkpoint + confidence alerts. That's it. ``` ## MCP Server StateWeave ships as an MCP server. Add to your MCP config: ```json { "mcpServers": { "stateweave": { "command": "python3", "args": ["-m", "stateweave.mcp_server"] } } } ``` ### Tools - `export_agent_state` — Export cognitive state from any supported framework - `import_agent_state` — Import state into a target framework with validation - `diff_agent_states` — Compare two states and return a change report ### Resources - `stateweave://schemas/v1` — Universal Schema specification - `stateweave://migrations/history` — Migration history log - `stateweave://agents/{id}/snapshot` — Live agent state snapshot ## Key Features - One-line auto-instrumentation (`stateweave.auto()` — like Sentry for agents) - One-command migration (`stateweave migrate --from langgraph --to crewai --agent my-agent`) - Round-trip fidelity benchmark (`stateweave benchmark` — tests all 10 frameworks) - Agent time travel (checkpoint, rollback, branch, diff) - git-style CLI: `log`, `blame`, `stash`, `pop`, `replay`, `status`, `compare` - Payload inspection (`stateweave inspect` — pretty-print any payload file) - Interactive migration picker (`stateweave try` — choose frameworks interactively) - Agent dashboard (`stateweave stats` — aggregate checkpoint statistics) - Shell completions (`stateweave completions bash/zsh/fish`) - Shareable markdown reports (`stateweave report` — paste into PRs/Slack) - Git pre-commit hook (`stateweave hook install` — CI on every commit) - Checkpoint pruning (`stateweave clean --keep 5` — disk management) - CLI config management (`stateweave config get/set/list` — no TOML editing) - Self-update check (`stateweave upgrade` — checks PyPI for new versions) - Environment snapshot (`stateweave env` — Python, frameworks, store, config) - Checkpoint search (`stateweave search "query"` — find state across history) - Visual checkpoint comparison (`stateweave compare agent v1 v2`) - Live agent dashboard (`stateweave watch`) - CI regression detection (`stateweave ci` — exits non-zero on behavior regression) - AES-256-GCM encryption with Ed25519 payload signing - 10 framework adapters (star topology — N adapters, not N² translations) - Smart auto-checkpoint middleware (every-N, delta-threshold, manual-only) - Delta state transport (send only changes) - State merge with CRDT conflict resolution - A2A protocol bridge - 12 UCE compliance scanners (serialization safety, import discipline, etc.) - GitHub Action template for CI integration - 730+ tests (unit, integration, adversarial red team) ## Links - Source: https://github.com/GDWN-BLDR/stateweave - PyPI: https://pypi.org/project/stateweave/ - License: Apache 2.0