@decionis/mcp 0.3.0 → 0.6.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/.claude-plugin/plugin.json +1 -1
- package/README.md +57 -5
- package/dist/AgentHookCli.js +2 -2
- package/dist/ManagedAgentHookCli.js +53391 -37117
- package/dist/agent/AgentIndex.js +2 -2
- package/dist/{chunk-ABHNQFWI.js → chunk-KV4WA4BV.js} +2 -2
- package/dist/{chunk-LIVLO4MV.js → chunk-XMQRTHQK.js} +32368 -19092
- package/dist/chunk-XMQRTHQK.js.map +1 -0
- package/dist/index.js +3112 -474
- package/dist/index.js.map +1 -1
- package/dist/policy_core.wasm +0 -0
- package/distribution/chatgpt/Install.md +40 -0
- package/distribution/chatgpt/Linked.md +25 -0
- package/distribution/chatgpt/Manifest.json +45 -0
- package/distribution/chatgpt/Mcp.json +9 -0
- package/distribution/chatgpt/Review.json +48 -0
- package/distribution/chatgpt/Setup.md +24 -0
- package/distribution/claude/Code.md +35 -0
- package/distribution/claude/Connected.json +13 -0
- package/distribution/claude/Install.md +47 -0
- package/distribution/claude/Manifest.json +11 -0
- package/distribution/claude/Mcp.json +8 -0
- package/distribution/common/Consent.md +22 -0
- package/distribution/common/Skill.md +111 -0
- package/distribution/copilot/Agent.json +22 -0
- package/distribution/copilot/App.json +34 -0
- package/distribution/copilot/AuthoringAgent.json +22 -0
- package/distribution/copilot/AuthoringApp.json +34 -0
- package/distribution/copilot/AuthoringInstructions.md +96 -0
- package/distribution/copilot/AuthoringM365.md +38 -0
- package/distribution/copilot/AuthoringPlugin.json +20 -0
- package/distribution/copilot/Connector.json +36 -0
- package/distribution/copilot/Instructions.md +21 -0
- package/distribution/copilot/M365.md +38 -0
- package/distribution/copilot/Plugin.json +20 -0
- package/distribution/copilot/Studio.md +16 -0
- package/distribution/schema/CopilotAgent.json +950 -0
- package/distribution/schema/CopilotApp.json +2927 -0
- package/distribution/schema/CopilotPlugin.json +723 -0
- package/distribution/schema/Mcp.json +120 -0
- package/distribution/schema/Plugin.json +65 -0
- package/distribution/schema/Source.md +36 -0
- package/package.json +15 -12
- package/dist/chunk-LIVLO4MV.js.map +0 -1
- /package/dist/{chunk-ABHNQFWI.js.map → chunk-KV4WA4BV.js.map} +0 -0
|
@@ -4,7 +4,7 @@
|
|
|
4
4
|
"version": "0.2.0",
|
|
5
5
|
"description": "Refuse an agent's tool call before it runs, using a policy you wrote. Evaluated locally through the real Decionis protocol evaluator \u2014 no account, no network, no credentials.",
|
|
6
6
|
"keywords": ["security", "policy", "guardrails", "execution-authority", "agent-safety", "hooks"],
|
|
7
|
-
"homepage": "https://decionis.
|
|
7
|
+
"homepage": "https://decionis.ai/mcp",
|
|
8
8
|
"repository": "https://github.com/decionis/mcp",
|
|
9
9
|
"license": "SEE LICENSE IN LICENSE",
|
|
10
10
|
"author": {
|
package/README.md
CHANGED
|
@@ -11,8 +11,9 @@ protocol's own schema-validated bundle ingestion, and evaluates through
|
|
|
11
11
|
locally is the verdict the platform would produce — with **zero network, zero database,
|
|
12
12
|
zero credentials, and nothing recorded**.
|
|
13
13
|
|
|
14
|
-
(The protocol service
|
|
15
|
-
|
|
14
|
+
(The protocol service also exposes a remote authenticated MCP surface. The local
|
|
15
|
+
execution path stays offline; optional connected PolicyAgent authoring is described
|
|
16
|
+
below.)
|
|
16
17
|
|
|
17
18
|
The package now exposes two complementary surfaces:
|
|
18
19
|
|
|
@@ -69,14 +70,20 @@ spends their money: `shield.request_purchase` returns ALLOW, ASK or BLOCK
|
|
|
69
70
|
against the limits the person set in the Shield app, with a signed Dispute
|
|
70
71
|
Passport behind every hold. Two ways to bind:
|
|
71
72
|
|
|
72
|
-
- **Pair from the app (ten minutes, no keys).** In the Shield app, open
|
|
73
|
-
spending allowance → Bind your own
|
|
74
|
-
Tell your assistant
|
|
73
|
+
- **Pair from the app (ten minutes, no keys).** In the Shield app, open _AI
|
|
74
|
+
spending allowance → Bind your own agent_; it shows a six-character code.
|
|
75
|
+
Tell your assistant _"pair with Shield, code ABC123"_ — it calls
|
|
75
76
|
`shield.pair`, which claims the code and keeps the production credential in
|
|
76
77
|
`~/.decionis/shield-credential.json` (`0600`; `DECIONIS_SHIELD_CREDENTIAL_FILE`
|
|
77
78
|
moves it). The app then lists the assistant with what it was allowed to
|
|
78
79
|
spend this week, and can revoke it. Pairing again from the same computer
|
|
79
80
|
rotates the credential rather than binding a second agent.
|
|
81
|
+
- **Sign in with the emailed code (no code from the app).** Tell your
|
|
82
|
+
assistant _"pair with Shield, my email is you@example.com"_. `shield.pair`
|
|
83
|
+
asks Shield to email the person a 6-digit code; they tell the assistant the
|
|
84
|
+
code, and Shield binds it the moment the code is confirmed. It keeps the same
|
|
85
|
+
credential file. The code only works for binding an agent, never for signing
|
|
86
|
+
in to the app.
|
|
80
87
|
- **Configure a key (integrators).** `DECIONIS_SHIELD_API_KEY` and
|
|
81
88
|
`DECIONIS_SHIELD_APP_ID` from `POST /v1/consumer/shield/integrations` win
|
|
82
89
|
over the paired file. `DECIONIS_SHIELD_API_URL` defaults to
|
|
@@ -361,3 +368,48 @@ printf '%s\n' \
|
|
|
361
368
|
the evaluate tool detects this and tells the agent how to fix the policy.
|
|
362
369
|
- Local evaluations mint no signed Decision Dossier; CI runs through
|
|
363
370
|
[`decionis/govern`](https://github.com/decionis/govern) produce the verifiable record.
|
|
371
|
+
|
|
372
|
+
## Governed PolicyAgent authoring
|
|
373
|
+
|
|
374
|
+
The local package can additionally connect to the governed PolicyAgent workspace.
|
|
375
|
+
This is opt-in: set `DECIONIS_POLICY_CONNECTED=1`, your
|
|
376
|
+
`DECIONIS_POLICY_WORKSPACE_ID`, and a private `DECIONIS_POLICY_OAUTH_TOKEN` issued
|
|
377
|
+
through an approved registered OAuth client. Without the opt-in flag the additional
|
|
378
|
+
authoring tools are absent. Existing execution hooks keep their existing behavior.
|
|
379
|
+
The token is read per call; a wrong workspace, expired grant or missing disclosure
|
|
380
|
+
approval fails closed. The package has no OAuth login or refresh-token store.
|
|
381
|
+
|
|
382
|
+
PolicyAgent reads authorized policy revisions, stores change proposals and requests
|
|
383
|
+
independent review. Human authors accept proposals into drafts and independent
|
|
384
|
+
reviewers attest in the workspace. Assistant confirmation never signs, approves or
|
|
385
|
+
deploys policy. Never put tokens in prompts or checked-in configuration.
|
|
386
|
+
|
|
387
|
+
### Policy authoring context for your assistant
|
|
388
|
+
|
|
389
|
+
On connect, the server sends short policy authoring instructions: write each rule as a
|
|
390
|
+
rule card, read it back, and propose it with `decionis.propose_policy_change`. Clients
|
|
391
|
+
that support MCP prompts and resources also get three guides:
|
|
392
|
+
|
|
393
|
+
| Prompt | Resource | Use it to |
|
|
394
|
+
| ----------------- | ----------------------------------- | ----------------------------------------------------------- |
|
|
395
|
+
| `policyauthoring` | `decionis://assist/policyauthoring` | Turn plain-English rules into PolicyAgent proposals. |
|
|
396
|
+
| `agentsafe` | `decionis://assist/agentsafe` | Put AgentSafe in front of an agent's writes. |
|
|
397
|
+
| `workflowsetup` | `decionis://assist/workflowsetup` | Set up governed CI and move a policy through its lifecycle. |
|
|
398
|
+
|
|
399
|
+
To keep the guides in a project as files, install them for one assistant:
|
|
400
|
+
|
|
401
|
+
```bash
|
|
402
|
+
npx @decionis/mcp context install --client claude # or codex, copilot, kiro, cursor
|
|
403
|
+
npx @decionis/mcp context install --client cursor --dir ./my-project --dry-run
|
|
404
|
+
```
|
|
405
|
+
|
|
406
|
+
The command writes the client's skills, agents, rules or steering files and adds the
|
|
407
|
+
Decionis servers to its MCP settings, keeping any other servers there. It updates files
|
|
408
|
+
it wrote before and skips any other existing file unless you pass `--force`. The files
|
|
409
|
+
name environment variables only; keep keys and tokens in your environment.
|
|
410
|
+
|
|
411
|
+
Provider package sources and installation instructions are under
|
|
412
|
+
`distribution/claude/`. Repository maintainers can run
|
|
413
|
+
`pnpm --filter @decionis/mcp package:policyagent claude` to create reviewable ZIPs.
|
|
414
|
+
See the repository's `docs/marketplace/policyagent/Claude.md` for the full review
|
|
415
|
+
exercise and publication requirements.
|
package/dist/AgentHookCli.js
CHANGED
|
@@ -1,8 +1,8 @@
|
|
|
1
1
|
#!/usr/bin/env node
|
|
2
2
|
import {
|
|
3
3
|
AgentHookCommand
|
|
4
|
-
} from "./chunk-
|
|
5
|
-
import "./chunk-
|
|
4
|
+
} from "./chunk-KV4WA4BV.js";
|
|
5
|
+
import "./chunk-XMQRTHQK.js";
|
|
6
6
|
|
|
7
7
|
// src/AgentHookCli.ts
|
|
8
8
|
void AgentHookCommand.run(process.argv.slice(2)).then((exitCode) => {
|