Developer preview
Start in observe mode with no change to agent behavior. Understand what Admis would have done, then turn the same canonical outcomes into approval, rewrite, step-up authentication, deferral or block.
Developer preview interface. Package installation is shown as the launch target and remains subject to clean-machine release validation.
First run to second run
The first run must be safe enough to try and useful enough to keep. Admis starts by showing the determination. Local explanation, summaries and tuning make the next run more relevant without creating an invisible bypass.
The trace makes the decision legible across identity, authority, consequence, reversibility and applicable obligations.
Adjust thresholds, group repeated determinations and suppress low-value console output. Every action is still evaluated and every decision record remains available.
mode: observe tuning: thresholds: payments.transfer.approval_above: 10000 group_by: [action, outcome, reason_code] suppress_from_console: [docs.search]
allow and allow_with_logging_escalation may be hidden from the console. Every outcome that would change behavior always surfaces.$ admis summary --since 7d 47 consequential actions observed 12 would require approval 3 would have been blocked 5 new action fingerprints $ admis tune payments.transfer threshold updated · configuration hash recorded in every subsequent trace
Start where you build
Admis does not ask you to replace the model, framework or execution host. Add the decision boundary at the point where your system can still pause, constrain, rewrite or stop an action.
Install one plugin, begin in observe mode and translate Admis outcomes into the approval and execution controls available in the host.
# Launch target. Exact registry path is validated before release. openclaw plugins install @admis/openclaw plugins: admis: mode: observe intercept: - before_tool_call record_store: local # The agent runs unchanged while Admis records # the canonical outcome and translated host effect.
Normalize tool intent before execution and preserve one trace from the agent proposal through the final execution result.
from langgraph.prebuilt import create_react_agent from admis.integrations.langgraph import AdmisMiddleware agent = create_react_agent( model, tools=[transfer_funds], middleware=[AdmisMiddleware(mode="observe")], ) # Adapters normalize and translate. # The Admis core remains the decision authority.
Intercept MCP tool calls, evaluate a canonical Action Context and return the result through the protocol controls the client can faithfully enforce.
admis gateway start \ --upstream http://localhost:3001/mcp \ --mode observe # client -> Admis MCP Gateway -> tool server # | # +-> canonical determination # +-> admissibility trace outcome: allow_with_human_approval translated_effect: input_required
Your control point remains local. Admis owns the decision semantics and returns a deterministic contract your infrastructure enforces.
POST /v1/decisions/evaluate { "action": { "tool": "payments.transfer", "operation": "execute", "arguments": { "amount": 450000 }, "target": { } }, "execution_contract": { "decision_class": "deep", "deadline_ms": 750, "failure_policy_id": "finance.fail_closed.v2" } } # One response vocabulary across every host.
Mock Signal Server
Switch among ten deterministic scenarios with bundled evidence. The lab exercises the real runtime and produces a local trace, without connecting IAM, SIEM or production systems.
A treasury agent proposes transferring USD 450,000 to a new vendor with valid credentials but insufficient autonomous authority.
The action is legitimate, but its consequence exceeds the authority delegated for unattended execution.
One core, many adapters
Every host supplies different hooks and enforcement controls. Admis preserves one Action Context, one decision authority and one outcome vocabulary, then translates only what the host can faithfully enforce.
Transform the host action, identity, authority and evidence into the canonical Action Context.
Evaluate once through the independent Admis core. The adapter contains no decision logic.
Map the canonical outcome to the strongest control the host can faithfully enforce.
Host fidelity is explicit
A supported connection is not the same as semantic fidelity. The Integration Registry publishes the hook, maturity, control coverage and known limitations for every adapter.
Filter the current preview registry. Maturity uses one vocabulary across the platform: Planned, In development, Experimental, Supported and Admis Certified.
| Integration | Boundary | Maturity | Approval | Rewrite | Step-up | Block | Known limitation |
|---|---|---|---|---|---|---|---|
| OpenClaw | before_tool_call | In development | Host dependent | Limited | External | Targeted | Exact registry install and runtime behavior require release validation. |
| LangGraph | tool middleware | In development | Interrupt | Arguments | Custom | Short-circuit | Compiled in preview; full runtime matrix remains under validation. |
| MCP Gateway | tools/call | In development | Client dependent | Protocol dependent | Elicitation | Yes | Control fidelity varies by client and negotiated protocol capability. |
| Decision API | POST /v1/decisions/evaluate | In development | Caller enforced | Contract | Contract | Caller enforced | The calling gateway owns local enforcement and fail-mode handling. |
| Copilot Studio | POST /analyze-tool-execution | Experimental | Retry pattern | Flow dependent | Identity flow | Reason + stop | No native canonical approval primitive; requires an explicit workflow. |
| AWS AgentCore | pre-target interceptor | Experimental | Workflow | Target dependent | Custom | Short-circuit | Interceptor budgets, retry behavior and idempotency must be preserved. |
Preview registry shown for design and integration planning. No adapter is Admis Certified until its golden scenarios, compatibility tests and published limitations pass the conformance process.
Adoption path
Observe is a local mode inside Develop. Shadow is the production stage where the same determinations run against live traffic without intervention. Enforcement is always an explicit institutional choice.
Run locally, explain determinations and tune the developer experience without changing agent behavior.
Evaluate production traffic, establish baselines and see what Admis would have changed across real workflows.
Authorize host-native approval, rewrite, step-up, constraint, quarantine and block behavior by environment.
Operate one decision contract, evidence model and assurance posture across agents, platforms and teams.
Start in observe mode during development. Move to Shadow when you want the same determinations across production traffic, still without changing agent behavior.
Developer resources
Use the specifications, examples and conformance assets that match your insertion point. Every path resolves to the same Action Context, outcome vocabulary and trace model.
Run the canonical finance demo and inspect your first determination.
Open guide →CookbookAdd observe mode to the agent runtime where you already build.
View examples →GatewayProtect multiple tools through one shared action boundary.
Open guide →SpecificationInspect the Action Context, outcomes, trace and decision record.
View reference →RegistryReview current hooks, maturity, control support and limitations.
View registry ↑CommunityBuild an adapter, run golden scenarios and prove compatibility.
View on GitHub ↗An llms.txt index, Markdown references, OpenAPI and JSON Schemas make the developer surface available inside the tools doing the implementation.
Independent decision authority for consequential AI-agent action. One canonical core across the frameworks, gateways and platforms you already use.
Choose your insertion point