First-party app · Claude Code
claude-code-agent
Run Anthropic's Claude Code as an agent (acp:claude-code) via the Zed ACP adapter. It runs with a Claude config of its own that starts empty, and loads nothing from the folder it works in (no .claude settings, .mcp.json servers, CLAUDE.md or skills), so no auto-approve rule of yours or of a repository's comes along: apart from the file reads and read-only commands Claude Code allows itself, every tool asks through PersonalClaw's approval gate. Sign in to Claude once for it (Settings → Providers → Claude Code → Sign in); PersonalClaw stores no key. Turn off Isolated Claude settings to use your own ~/.claude, rules included, and the Claude settings of the folder it works in. Enabling it installs its ACP adapter (@agentclientprotocol/claude-agent-acp) from npm into your PersonalClaw home when no copy is installed; a failed install says why on its card, with Retry.
Before you install
A first-party app of PersonalClaw v0.3.9
No scan on record before install
PersonalClaw publishes no scan of its own apps before you install one.
PersonalClaw runs its own scan of the actual code of every app it installs: this one is rescanned when you install, and that scan is what can stop an install.
Its row on Compare: install rescan
- Compatibility
- Works with PersonalClaw v0.3.9
What it gets
- Permissions
- network: true
- Its own network permission
- On. Its own network permission is a declaration, not a fence.
What it starts
The programs on this machine it starts, as install consent names them.
claude: Claude Code does the work of each chat. Its ACP adapter starts it and hands it each turn.While Isolated Claude settings is off, it runs with your own claude sign-in, settings and auto-approve rules, and with the claude settings of the folder it works in, which can add rules of their own and commands for it to run. What those rules allow, it does without asking you here first. Isolated Claude settings is on until you turn it off.
What it reaches
- Hosts it names
- Not declared in this release's manifest
Its own code has no host list at this release: any hosts it names are for the programs, servers and sources it starts or fetches, not everything it can reach. Its own network permission is a declaration, not a fence.
Secrets you would enter
- Settings marked secret
- Not declared in this release's manifest
What installing it gives and runs
The manifest's own declarations, under the names install consent shows them by and in its order. The Store adds, at install, which of them run as you.
- Permissions it is granted
{ "network": true }- Programs on this machine it starts
[ { "program": "claude", "why": "Claude Code does the work of each chat. Its ACP adapter starts it and hands it each turn.", "inherits": [ "sign-in", "settings", "auto-approve-rules", "folder-settings" ], "inheritsWhile": { "setting": "isolated_config", "value": false } } ]- npm packages installed for it
[ "@agentclientprotocol/claude-agent-acp" ]
- Places outside its folder it writes
[ { "path": "cc-config", "why": "The Claude config Claude Code runs with while Isolated Claude settings is on. It starts empty, and signing in to Claude for this app keeps that sign-in there." } ]- Provider modules the gateway loads
[ { "type": "agent", "implementation": "provider:create_provider" } ]
Not declared in this release's manifest:
- Scheduled jobs
- Python packages installed for it
- Engine packages it can install
- What it needs that PersonalClaw does not install
- Pages it adds to the dashboard
- Components it adds to the dashboard
- A server process of its own
- The sandbox tier its server runs in
- Command run after install
- Command run after an update
- Command run when it is switched on
- Command run when it is switched off
- Command run before it is removed
- What it runs for personalclaw setup
- What it runs for personalclaw doctor
- Connector sources it parses
- MCP servers it adds
- Skills it installs
Where it comes from
- Publisher
- PersonalClaw
- License
- MIT
- Pinned commit
- 169cbc4f8fc59cd42a65a6a3546235bc49ab3b22
- Signer
- Not declared in this release's manifest
From review to enable
- Held and scanned. An install stages the app in quarantine, checks its signature, and runs the shared supply-chain scanner over the staged bytes before anything is shown for consent. A registry's recorded verdict is not read at install: the scan that can stop it is this one.Its row on Compare: install rescan
- A dangerous verdict ends it. A dangerous verdict returns a refusal before consent is asked, so there is no consent to give and no switch that skips it. A warning verdict asks for your consent instead.Its row on Compare: dangerous verdict
- You review and consent. The review records a digest of the staged bytes, and an install committed against different bytes is refused: it needs consent again, because the app changed after it was reviewed. An update that changes what the app gets asks for consent again too.Its row on Compare: consent
- Its setup runs, then you switch it on. Nothing it declares runs before your consent; an app already installed is offered as an update instead.
To install it and switch it on
- In PersonalClaw, open the Store.
- Find claude-code-agent.
- Review what it gets and runs, and consent.
- Install it, then switch it on.
Evidence
claude-code-agent/app.json at the pinned commit, as written, with the blob hash git lists for it: c81920ca9a68232bf417d431ab1d0be9c6c365eb
{
"name": "claude-code-agent",
"version": "0.2.7",
"displayName": "Claude Code",
"description": "Run Anthropic's Claude Code as an agent (acp:claude-code) via the Zed ACP adapter. It runs with a Claude config of its own that starts empty, and loads nothing from the folder it works in (no .claude settings, .mcp.json servers, CLAUDE.md or skills), so no auto-approve rule of yours or of a repository's comes along: apart from the file reads and read-only commands Claude Code allows itself, every tool asks through PersonalClaw's approval gate. Sign in to Claude once for it (Settings \u2192 Providers \u2192 Claude Code \u2192 Sign in); PersonalClaw stores no key. Turn off Isolated Claude settings to use your own ~/.claude, rules included, and the Claude settings of the folder it works in. Enabling it installs its ACP adapter (@agentclientprotocol/claude-agent-acp) from npm into your PersonalClaw home when no copy is installed; a failed install says why on its card, with Retry.",
"icon": "SquareTerminal",
"author": "PersonalClaw",
"license": "MIT",
"tags": [
"agent",
"acp",
"claude-code"
],
"permissions": {
"network": true
},
"launches": [
{
"program": "claude",
"why": "Claude Code does the work of each chat. Its ACP adapter starts it and hands it each turn.",
"inherits": [
"sign-in",
"settings",
"auto-approve-rules",
"folder-settings"
],
"inheritsWhile": {
"setting": "isolated_config",
"value": false
}
}
],
"writes": [
{
"path": "cc-config",
"why": "The Claude config Claude Code runs with while Isolated Claude settings is on. It starts empty, and signing in to Claude for this app keeps that sign-in there."
}
],
"dependencies": {
"npmPackages": [
"@agentclientprotocol/claude-agent-acp"
]
},
"provider": {
"id": "agent",
"label": "Claude Code as an agent",
"type": "agent",
"implementation": "provider:create_provider",
"capabilities": [
"acp"
],
"settingsSchema": {
"type": "object",
"properties": {
"isolated_config": {
"type": "boolean",
"default": true,
"x-meta": {
"label": "Isolated Claude settings",
"help": "On: Claude runs with a config of its own under PersonalClaw's folder that starts empty, and loads nothing from the folder it works in (its .claude settings, .mcp.json servers, CLAUDE.md and skills), so no auto-approve rule of yours or of a repository's comes along: apart from the file reads and read-only commands Claude Code allows itself, every tool asks here first. Sign in to Claude once for it. Off: Claude uses your own ~/.claude, auto-approve rules included, and the folder's own Claude settings, whose rules, hooks and MCP servers apply without asking here first."
}
},
"model": {
"type": "string",
"default": "",
"x-meta": {
"label": "Default Model",
"help": "Optional. Leave empty to use the Claude CLI's own current default (recommended). The Claude adapter advertises the live model set for selection; set this only to pin a specific model."
}
},
"acp_bin": {
"type": "string",
"default": "",
"x-meta": {
"label": "ACP Adapter Path",
"help": "Optional absolute path to the claude-agent-acp adapter. Empty finds it: PATH \u2192 node-manager dirs \u2192 the copy PersonalClaw installed in its home when you enabled the app \u2192 npx @agentclientprotocol/claude-agent-acp. Equivalent to the CLAUDE_CODE_ACP_BIN env var.",
"tags": [
"advanced"
]
}
}
}
}
}
}
- Scan verdict
- No scan on record before install
- Scan time
- No scan is published before install
- Blocking findings
- No scan is published before install
- Scanner version
- No scan is published before install
PersonalClaw publishes no scan of its own apps before you install one.
Every app is scanned again when you install it, whatever its listing says.
An install stages the app in quarantine, checks its signature, and runs the shared supply-chain scanner over the staged bytes before anything is shown for consent. A registry's recorded verdict is not read at install: the scan that can stop it is this one.
apps/app_manager.py at v0.3.9A dangerous scan verdict stops the install, and nothing overrides it.
A dangerous verdict returns a refusal before consent is asked, so there is no consent to give and no switch that skips it. A warning verdict asks for your consent instead.
apps/app_manager.py at v0.3.9Consent covers exactly what you reviewed, and an app that changed asks again.
The review records a digest of the staged bytes, and an install committed against different bytes is refused: it needs consent again, because the app changed after it was reviewed. An update that changes what the app gets asks for consent again too.
apps/app_manager.py at v0.3.9A setting an app marks secret is kept in the credential store and never shown back.
The app's settings file holds only a reference, the value lives in the credential store, and every route that shows settings masks it. The app's own code receives the value when it runs.
apps/secret_fields.py at v0.3.9
PersonalClaw/PersonalClawApps @ 169cbc4f8fc5 · claude-code-agent/app.json