@getmarrow/mcp 3.9.13 → 3.9.14
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/README.md +4 -73
- package/package.json +1 -1
package/README.md
CHANGED
|
@@ -63,7 +63,9 @@ Accounts with <7 days of activity AND <20 decisions get an onboarding payload sh
|
|
|
63
63
|
|
|
64
64
|
---
|
|
65
65
|
|
|
66
|
-
## What's New in v3.9.
|
|
66
|
+
## What's New in v3.9.14
|
|
67
|
+
|
|
68
|
+
v3.9.14 is a README-only patch release for npmjs. It keeps the package page focused on the current install surface and sends full feature history, examples, and API reference to [getmarrow.ai/docs](https://getmarrow.ai/docs/).
|
|
67
69
|
|
|
68
70
|
### Decision Brief Tool
|
|
69
71
|
|
|
@@ -84,78 +86,7 @@ Marrow returns aggregated prior failure categories only. It does not expose raw
|
|
|
84
86
|
|
|
85
87
|
`marrow_decision_brief` is additive. It gives the pre-action operating brief, then the agent still logs intent with `marrow_think`/passive auto-logging and commits the verified outcome with `marrow_commit`.
|
|
86
88
|
|
|
87
|
-
|
|
88
|
-
|
|
89
|
-
Agent Status Tool
|
|
90
|
-
|
|
91
|
-
New MCP tool: `marrow_agent_status`.
|
|
92
|
-
|
|
93
|
-
Use it at session start or before owner reporting to prove Marrow is passively active. It returns connected state, signal quality, non-sensitive proof, and next actions for the configured agent. Pass `agentId` explicitly when you want a different agent or fleet-scoped status.
|
|
94
|
-
|
|
95
|
-
```json
|
|
96
|
-
{
|
|
97
|
-
"active": true,
|
|
98
|
-
"state": "proving_value",
|
|
99
|
-
"signals": {
|
|
100
|
-
"decisions_logged": 42,
|
|
101
|
-
"outcome_coverage": 0.88,
|
|
102
|
-
"success_rate": 0.81
|
|
103
|
-
},
|
|
104
|
-
"quality": { "enough_signal": true, "measurement_risk": "low" }
|
|
105
|
-
}
|
|
106
|
-
```
|
|
107
|
-
|
|
108
|
-
No raw action, context, or outcome text is returned.
|
|
109
|
-
|
|
110
|
-
### Previous: v3.9.9 — Hook Noise Filtering
|
|
111
|
-
|
|
112
|
-
MCP hooks now skip auto-logging for read-only operations. Tools like `read`, `grep`, `ls`, `sessions_list`, `marrow_list_memories`, and other context-read operations no longer generate Marrow decisions. Only meaningful tool calls (edits, writes, deploys) are logged.
|
|
113
|
-
|
|
114
|
-
This reduces noise in your decision history and keeps your Marrow training data clean — automatically, no agent discipline required.
|
|
115
|
-
|
|
116
|
-
### Previous: v3.9.4 — Standalone CLI + Multi-API-Key Management
|
|
117
|
-
|
|
118
|
-
**Standalone CLI commands (new in v3.9.0):**
|
|
119
|
-
```bash
|
|
120
|
-
npx @getmarrow/mcp keys create --name "Prod" --type live
|
|
121
|
-
npx @getmarrow/mcp keys list
|
|
122
|
-
npx @getmarrow/mcp keys rotate --id key_abc123
|
|
123
|
-
npx @getmarrow/mcp keys revoke --id key_abc123
|
|
124
|
-
npx @getmarrow/mcp keys audit --limit 20
|
|
125
|
-
```
|
|
126
|
-
|
|
127
|
-
**MCP tools (for Claude/Cursor agents):**
|
|
128
|
-
|
|
129
|
-
- `marrow_create_key` — Create named, scoped API keys (`live`/`test` type, optional scopes and agent bindings)
|
|
130
|
-
- `marrow_list_keys` — List all API keys with masked display and tier limits
|
|
131
|
-
- `marrow_get_key` — Get key details, usage stats, and expiry info
|
|
132
|
-
- `marrow_revoke_key` — Permanently revoke a key
|
|
133
|
-
- `marrow_rotate_key` — Atomically rotate (revoke old + create new, returns new key once)
|
|
134
|
-
|
|
135
|
-
Create/rotate return the full key exactly once with a warning. All other tools return masked keys only.
|
|
136
|
-
|
|
137
|
-
```bash
|
|
138
|
-
npx @getmarrow/mcp --key mrw_...
|
|
139
|
-
# Now available: marrow_create_key, marrow_list_keys, marrow_get_key,
|
|
140
|
-
# marrow_revoke_key, marrow_rotate_key
|
|
141
|
-
```
|
|
142
|
-
|
|
143
|
-
---
|
|
144
|
-
|
|
145
|
-
### Previous: v3.7.0 — Passive Context Injection
|
|
146
|
-
|
|
147
|
-
Marrow's passive loop is now complete. v3.2.0 shipped the **write side** — every tool call your agent makes is auto-logged via a PostToolUse hook. v3.7.0 ships the **read side** — Marrow's intelligence is auto-injected into your agent's prompt window via a UserPromptSubmit hook. No agent discipline required for either side.
|
|
148
|
-
|
|
149
|
-
When you run `npx @getmarrow/mcp setup`, both hooks install automatically. From the next user prompt forward:
|
|
150
|
-
|
|
151
|
-
1. You type a message to your agent
|
|
152
|
-
2. Hook fires, takes your prompt, queries Marrow for relevant warnings + hive patterns + similar past decisions
|
|
153
|
-
3. Marrow context (if any) is injected into your agent's view as `additionalContext`
|
|
154
|
-
4. Agent responds knowing what failed last time, what worked, and what patterns to leverage
|
|
155
|
-
|
|
156
|
-
The hook is fail-silent: 2-second timeout on the Marrow API call, errors swallowed, never blocks your prompt. No signal → no injection. `MARROW_AUTO_HOOK=false` disables both hooks. `MARROW_CONTEXT_HOOK_DEBUG=true` enables stderr diagnostics.
|
|
157
|
-
|
|
158
|
-
This closes the loop Buu identified: Marrow now *just works* — your agent reads from the hive without you ever telling it to.
|
|
89
|
+
Full feature history, examples, and API reference live at [getmarrow.ai/docs](https://getmarrow.ai/docs/).
|
|
159
90
|
|
|
160
91
|
---
|
|
161
92
|
|