← Back to Get Started
Technical reference
Profiles, outcomes, and integration boundaries.
This page keeps the detailed research-preview material out of the quick-start path while preserving it for implementation and review.
Reference environment
Use the supported profile as a whole.
The candidate targets Linux/amd64 with CPython 3.12, Docker, and Compose. One-time setup may use the network; the measured demonstration is credential-free and offline after setup.
Operation outcomes
Read the operation result before calling anything again.
Effect retained
The governed effect occurred and its authoritative result is available.
Integration response: Return or inspect that result; do not invoke the original consequential effect again.
Effect not run
The action was not permitted and the governed effect did not occur.
Integration response: Return the denial as the final operation outcome without calling the original effect.
Effect not run
No final governed effect has occurred and the operation remains durable.
Integration response: Resolve it through an independently authorized MasuGate path, then read the authoritative operation state.
Reference integration profiles
Keep the contract fixed; vary the host edge.
These profiles describe exact reference bindings, not a broad compatibility promise. Open one only when you need the host-level detail.
reference · experimentalOpenClaw
OpenClaw keeps orchestration while declared consequential tool calls cross the MasuGate boundary.
- Host pins
- OpenClaw 2026.7.1
- Adapter boundary
- The MasuGate-backed tool replaces the original consequential purchase tool on the declared route. Unrelated host tools remain host-owned and are not mediated automatically.
- Deployment-owned configuration
- Service base URL · Agent-to-credential mapping · Finite governed-route catalog · Provider, execution position, and connector binding
reference · experimentalLangChain / LangGraph
A configured replacement tool binds graph invocation context to one MasuGate operation.
- Host pins
- LangChain 1.3.14 · LangGraph 1.2.9
- Adapter boundary
- The MasuGate-backed tool replaces the original consequential purchase tool on the declared route. Unrelated host tools remain host-owned and are not mediated automatically.
- Deployment-owned configuration
- Replacement tool registration · Trusted graph-context binding · Host-native result conversion
reference · experimentalMicrosoft Agent Framework
A typed replacement function returns the authoritative MasuGate result to the agent workflow.
- Host pins
- Microsoft Agent Framework Core 1.12.0
- Adapter boundary
- The MasuGate-backed tool replaces the original consequential purchase tool on the declared route. Unrelated host tools remain host-owned and are not mediated automatically.
- Deployment-owned configuration
- Typed function registration · Trusted invocation-context binding · Framework result conversion
reference · experimentalCrewAI
A governed task tool binds task and tool-call context to one replay-safe MasuGate operation.
- Host pins
- CrewAI 1.15.6 · CrewAI Core 1.15.6
- Adapter boundary
- The MasuGate-backed tool replaces the original consequential purchase tool on the declared route. Unrelated host tools remain host-owned and are not mediated automatically.
- Deployment-owned configuration
- Governed task-tool registration · Trusted task-context binding · Retry and resume handling
OpenClaw boundary
One declared route crosses the governance boundary.
OpenClaw 2026.7.1 retains orchestration. The configured masugate_governed_action route connects the selected action to the MasuGate execution boundary.
- MasuGate is the execute boundary for the declared purchase route, not a detached policy-check endpoint.
- The adapter governs only declared routes; it does not intercept unrelated OpenClaw tools.
- Trusted agent, session, replay, credential, provider, and route context do not come from model-controlled arguments.
- Stable action identity from trusted host context is required for safe retry and replay behavior.
- A committed result is authoritative; the caller must not invoke a second native purchase effect.
- Pending work requires a distinct, independently authorized resolution path.
- Providers remain responsible for their declared state views and governed effects.
Troubleshooting
Diagnose the observed boundary.
The host platform or tool version is rejected.
Check: Compare every local version with the exact candidate or selected-release profile.
Next: Use the declared profile; do not substitute an unverified package combination.
One-time setup cannot retrieve or verify a pinned input.
Check: Check that setup has network access and that the retrieved input matches its declared lock or digest.
Next: Preserve the integrity failure and retry only with the release-declared input.
Docker or Compose is unavailable.
Check: Confirm both required container capabilities before entering the reference workflow.
Next: Use the documented Linux/amd64 reference environment or request a guided demonstration.
masugated cannot be reached.
Check: Check the deployment-owned service location and the reference stack's service health.
Next: Restore the declared service boundary before retrying the same logical action.
A credential and expected principal do not match.
Check: Compare the deployment-owned credential mapping with the trusted principal recorded for the operation.
Next: Correct the trusted deployment mapping; never move principal selection into model-visible arguments.
A route, provider, execution position, or connector assertion is rejected.
Check: Inspect the finite route catalog and its provider, position, and connector binding as one contract.
Next: Correct the deployment-owned binding instead of bypassing the governed route.
An operation remains pending.
Check: Read the authoritative operation state and identify the independently authorized resolution path.
Next: Resolve or decline through that path, then read the same operation identity again.
Generated evidence does not pass verification.
Check: Compare the environment, immutable revision, expected result, and named gate with the selected release.
Next: Treat the mismatch as a failed verification; do not relabel the output or disable validation.