Repository navigation
Releases
v0.4.0 — OpenClaw Migration & Governance Audit
OpenClaw users: Anthropic is ending Claude subscription access for third-party tools on April 4th. Use AgentShift to migrate your agents — with governance preserved.
Quick migration
pip install agentshift
agentshift convert ~/.openclaw/skills/my-agent --from openclaw --to claude-code -o ./output/
agentshift audit ~/.openclaw/skills/my-agent --targets claude-code,copilotWhat's in v0.4.0
Governance Framework (L1/L2/L3)
- 3-layer governance model: prompt guardrails (L1), tool permissions (L2), platform-native controls (L3)
- Permission elevation engine — promotes L2/L3 rules to L1 instructions when target lacks native support
agentshift audit— Governance Preservation Rates (GPR) for any conversionagentshift audit-batch— batch audit across multiple agents
Governance Preservation Results (12 agents x 2 targets)
| Target | GPR-L1 | GPR-L2 | GPR-L3 | GPR-Overall |
|---|---|---|---|---|
| Claude Code | 1.00 | 0.93 | 0.00* | 0.83 |
| Copilot | 1.00 | 0.37 | 0.00* | 0.62 |
*L3 elevated to prompt instructions with 93.6% behavioral equivalence (tested across 330 LLM evaluations)
New Parsers & Emitters
- 7 parsers: OpenClaw, Claude Code, Copilot, Bedrock, Vertex, AGENTS.md, M365
- 7 emitters: Claude Code, Copilot, Bedrock, M365, Vertex, LangGraph, A2A
- EU AI Act compliance checker (
agentshift compliance) - Local agent registry for drift detection (
agentshift registry) - GitHub Action for CI/CD portability reports
- 1,037 tests passing
Links
- Website: agentshift.sh
- Migration guide: README
v0.3.0 — 5 targets: Claude Code, Copilot, Bedrock, M365, Vertex AI
AgentShift v0.3.0
Convert OpenClaw skills to 5 platforms with a single command.
pip install agentshift
agentshift convert ./my-skill --from openclaw --to allOutputs:
output/
├── claude-code/ ← CLAUDE.md + settings.json (run immediately)
├── copilot/ ← github.agent.md (run immediately)
├── bedrock/ ← instruction.txt + cloudformation (deploy + wire Lambdas)
├── m365/ ← declarative-agent.json + manifest (sideload to M365)
└── vertex/ ← agent.json (deploy to Vertex AI)
New in v0.3.0
--to vertex— GCP Vertex AI Agent Builder emitter--to m365— Microsoft 365 Declarative Agent emitter (schema v1.4)agentshift validate— per-platform schema validation- 5 targets in
agentshift diffportability scoring - Docs: agentshift.sh
All commands
agentshift convert ./my-skill --from openclaw --to all
agentshift diff ./my-skill --from openclaw
agentshift validate ./output/bedrock --target bedrock324 tests passing across Python 3.11, 3.12, 3.13
v0.2.1 — Microsoft 365 Declarative Agent emitter
What's new
Microsoft 365 Copilot emitter (--to m365)
Convert any OpenClaw skill to a Microsoft 365 Declarative Agent:
agentshift convert ./my-skill --from openclaw --to m365 --output ./my-skill-m365Generates:
declarative-agent.json— M365 Copilot agent manifest (schema v1.4, 8,000 char instructions)manifest.json— Teams app manifest with auto-generated UUIDREADME.md— how to sideload into M365 Copilot
Now 4 working targets
agentshift convert ./my-skill --from openclaw --to all
# → claude-code/ copilot/ bedrock/ m365/Install
pip install agentshift==0.2.1v0.2.0 — AWS Bedrock + agentshift diff + --to all
What's new in v0.2.0
AWS Bedrock emitter
Convert any OpenClaw skill to a complete Bedrock deployment package:
agentshift convert ./my-skill --from openclaw --to bedrock --output ./my-skill-bedrockGenerates: instruction.txt + openapi.json + cloudformation.yaml + README.md
agentshift diff — portability scoring
See exactly what converts cleanly before you commit:
agentshift diff ./my-skill --from openclawShows a rich table with per-component fidelity and a portability score (0–100%) per target.
--to all — convert to every target at once
agentshift convert ./my-skill --from openclaw --to all --output ./output
# → output/claude-code/ output/copilot/ output/bedrock/Supported targets
| Platform | Status |
|---|---|
| Claude Code | ✅ |
| GitHub Copilot | ✅ |
| AWS Bedrock | ✅ new |
| GCP Vertex AI | Coming soon |
Install
pip install agentshift==0.2.0Docs
Full guides at agentshift.sh
v0.1.1 — Precise Claude Code permissions
What's new
- 54/54 OpenClaw skills convert cleanly to Claude Code
- Precise
Bash(<binary>:*)permissions (e.g.Bash(gh:*)instead of genericBash(bash:*)) - MCP tool detection from prose and JSON action blocks
Write(path)permissions for skills that write data files
Install
pip install agentshiftv0.1.0 — OpenClaw → Claude Code
What's in v0.1.0
Convert OpenClaw skills to Claude Code format with a single command.
pip install agentshift
agentshift convert ~/.openclaw/skills/weather --from openclaw --to claude-code --output ./weather-claudeIncluded
- 🧠 Universal IR model — platform-agnostic agent representation
- 🔍 OpenClaw parser (SKILL.md → IR)
- 🔍 Claude Code parser (CLAUDE.md → IR)
- ✍️ Claude Code emitter (IR → CLAUDE.md + settings.json)
- ⌨️ CLI:
agentshift convert - 🧪 79 tests passing across Python 3.11, 3.12, 3.13
- 📦 4 worked examples (weather, github, slack, notion)
Known limitations
- Cron triggers not yet parsed from jobs.json (#2)
- Only Claude Code emitter — Copilot, Bedrock, Vertex AI coming in v0.2
What's next
Week 2: Copilot, Bedrock, and Vertex AI emitters