接入你的智能体
发布者自己的命令。不是 PlugMCP 端点。
用我们的密钥与日志托管一份副本。
npx -y @ctliz/[email protected]Claude Code 没有针对本地命令的安装链接。复制下方命令并在终端中运行。
claude mcp add ctliz-agent-intercom-claude -- npx -y @ctliz/agent-intercom-claude@0.16.1在终端中运行。
{ "mcpServers": { "ctliz-agent-intercom-claude": { "command": "npx", "args": [ "-y", "@ctliz/[email protected]" ] } }}claude_desktop_config.json
远程 URL 还是本机命令？ 两种安装方式的区别
来自发布者
Agent Intercom is a cross-harness, same-machine messaging system for coding agents. Its Pi, Codex, Claude Code, OpenCode, Grok Build, and AGY adapters share one local broker and protocol, so sessions can discover and message each other regardless of which harness they run in.
Claude Code MCP plugin and wakeable worker for local messaging with Pi, Codex, OpenCode, Grok Build, and AGY agents.
| Harness | Repository |
|---|
| Core / Protocol | agent-intercom-core |
| Pi | agent-intercom-pi |
| Codex | agent-intercom-codex |
| Claude Code | agent-intercom-claude |
| OpenCode | agent-intercom-opencode |
| Grok Build | agent-intercom-grok |
| AGY | agent-intercom-agy |
| Fleet lifecycle | agent-intercom-orchestrator |
Grok Build and AGY are supported as first-class protocol peers through two dedicated npm packages:
The host packages depend on @ctliz/agent-intercom-claude and load its MCP runtime internally. Users do not need to install or place claude-intercom-mcp on PATH separately. Once connected, Grok and AGY sessions share the same local broker and protocol as Pi, Codex, Claude Code, and OpenCode, and expose the MCP operations: intercom_whoami, intercom_list, intercom_send, intercom_ask, intercom_reply, intercom_pending, intercom_status, intercom_team, intercom_join, and intercom_set_summary.
Install the host adapter before installing its plugin:
npm install -g @ctliz/agent-intercom-grok
npm install -g @ctliz/agent-intercom-agy
For multi-pane or Auto-Team-style use, the supervisor must give every MCP child a unique literal identity and the same scope as its intended peers:
AGENT_INTERCOM_SESSION_ID=<unique-pane-or-worker-id>
AGENT_INTERCOM_SESSION_NAME=<human-readable-name>
AGENT_INTERCOM_SCOPE_ID=<shared-team-or-workspace-scope>
The plugin manifests intentionally do not contain static session IDs, because sharing one ID across concurrent panes would create identity collisions. CLAUDE_INTERCOM_SESSION_ID and CLAUDE_INTERCOM_NAME remain higher-priority compatibility aliases.
Grok Build and AGY currently provide polling-only MCP integrations. Messages are durably retained, but there is no host-specific wake bridge to inject a new turn. Each active agent should call intercom_pending at startup and at natural work boundaries. They can join an existing team and exchange messages with every other adapter, but Agent Intercom Orchestrator does not currently spawn or lifecycle-manage Grok or AGY workers.
ctliz: This distribution is maintained independently by ctliz.pi-intercom and the upstream dataforxyz/agent-intercom-* repositories. This project is not officially endorsed by or affiliated with upstream organizations.@ctliz/*. The historical @dataforxyz/* namespace was used up to and including connect.1 and is retained only as provenance and as a migration-detection input; it is never treated as a current or healthy installation. The Agent Intercom branding and the intercom_* API surface are unchanged.Agent Intercom protocol v4 introduces broker-enforced scope routing via AGENT_INTERCOM_SCOPE_ID:
scopeId once in the top-level registration payload.ConnectedSession record and enforces same-scope discovery (intercom_list), naming, and prefix matching.scopeId value never enters SessionInfo, list payloads, lifecycle events, frontend displays, or execution logs.AGENT_INTERCOM_SCOPE_ID is a general shell/IDE/service launcher contract. Agent Intercom works completely standalone in any terminal, tmux window, or script; TmuxDeck is optional visual tooling.Agent Intercom grew from Nico Bailon's original pi-intercom. A sincere thank you to Nico and the original contributors for creating the Pi extension and the foundation this cross-harness family builds on.
This repository contains the Claude Code adapter. It uses the shared strict pi-intercom protocol v4. Any adapter may start the broker first; incompatible legacy brokers fail closed without killing, downgrading, or creating second islands. Sends are retained in a durable per-session outbox and replayed after reconnect, while receiver acknowledgement distinguishes broker acceptance from durable receipt.
Attached Claude sessions support two local delivery transports. native bridges the broker to Claude Code's cross-session Unix socket protocol; mcp preserves the plugin/inbox/Monitor path. The default auto mode selects native only for Claude Code versions in the verified compatibility window (currently 2.1.220–2.1.226) and otherwise selects MCP. Explicit native selection fails closed outside that window.
When running cci or ccim in an attached terminal, press Alt+M to choose a connected session and send it a message, or Alt+I to copy that worker's intercom contact target. The MCP plugin cannot register native Claude Code keyboard shortcuts because Claude Code does not expose plugin keybinding registration; the plugin instead provides /claude-intercom:intercom and /claude-intercom:intercom-id. Detached worker-daemon mode has no terminal shortcuts.
Claude Intercom adds local messaging between Claude Code, Codex, Pi, OpenCode,
and other coding-agent sessions on the same machine. It speaks the same local broker
protocol as pi-intercom and
codex-intercom, so sessions
can discover each other, send updates, ask blocking questions, read pending
messages, and reply to asks across all four supported harnesses.
The project has two related pieces:
claude-intercom-mcp: an MCP server that exposes intercom tools inside a
normal Claude Code session.cci / claude-intercom-worker: a wakeable Claude worker. It registers
an intercom identity, and when another session sends it work, it starts a
fresh headless claude -p turn that resumes the worker's own conversation —
so the worker can read files, run commands, edit code, and reply on its own.Use plain MCP when you only need tools inside an already-active Claude turn. Use a wakeable worker when you want another session to wake Claude automatically and have it act with real system access.
Preview. This is the Claude-side adapter, built alongside pi-intercom and
codex-intercom.
A plain Claude Code MCP session does not receive unsolicited visible turns.
Incoming messages are queued while the MCP server is running; call
intercom_pending to read them. Wake-on-message workflows use cci /
claude-intercom-worker.
Claude Code has no long-lived programmatic "app-server" the way Codex does, so the worker uses the most robust primitive available: the headless CLI.
claude -p --output-format json --resume <session-id> ..., feeding the
message text on stdin. Normal cci workers automatically receive the
packaged Intercom MCP server, even under an isolated CLAUDE_CONFIG_DIR or
custom ANTHROPIC_BASE_URL.session_id.session_id so the next message resumes the same
conversation, and (for blocking asks) sends the final assistant message back
to the asker as the reply.This gives a woken worker genuine access to the system while keeping each worker
a continuous, resumable conversation. You can attach to a worker's conversation
at any time with claude --resume <session-id>.
Install via npm using the connect dist-tag:
npm install -g @ctliz/agent-intercom-claude@connect
# or by exact prerelease version
npm install -g @ctliz/[email protected]
Or install from GitHub source at the exact tag so the command-line entry points are on PATH:
git clone --depth 1 --branch v0.13.1 https://github.com/ctliz/agent-intercom-claude.git
cd agent-intercom-claude && npm ci && npm link
This provides:
claude-intercom-mcpclaude-intercom-workercci — start a normal wakeable workerccim — start a minimal wakeable worker (cci --minimal)To let a Pi manager create Claude workers with owned systemd cgroups, leases, model/effort selection, logs, and verified cleanup, install the companion Pi packages:
pi install git:github.com/ctliz/[email protected]
pi install git:github.com/ctliz/[email protected]
Restart Pi or run /reload, then call agent_fleet({ action: "doctor" }). The orchestrator invokes the installed cci/ccim commands; it does not replace this Claude adapter.
cci and ccim are the recommended entry points when you want an attached,
wakeable Claude session. Unlike a plain MCP session or a detached headless
worker, the attached wrappers provide the Alt+M session picker/message
composer and the Alt+I contact-copy shortcut; they also keep an intercom
identity online so another agent can wake the worker.
If you use the same worker profiles repeatedly, add memorable shell aliases
with your own portable project paths and stable IDs:
alias claude-reviewer='cci --cwd "$HOME/src/my-project" --name reviewer --id reviewer'
alias claude-reviewer-min='ccim --cwd "$HOME/src/my-project" --name reviewer-min --id reviewer-min'
Put aliases in your shell startup file (for example ~/.bashrc or ~/.zshrc).
They are optional convenience shortcuts: the installed cci and ccim
commands work directly, but aliases make stable identities and project-specific
defaults easier to reuse without copying a long command.
For a plain, already-active Claude Code session, add the MCP server explicitly:
claude mcp add claude-intercom -- claude-intercom-mcp
With --transport mcp, cci does this automatically for each normal headless worker. Native headless workers are still woken and replied through the worker daemon's broker connection, but omit the packaged MCP server from the Claude turn. ccim intentionally uses Claude's --safe-mode, which disables MCP servers along with plugins, hooks, and skills.
The claude-intercom-mcp executable is a standard stdio MCP server, but Grok Build and AGY should normally use the dedicated @ctliz/agent-intercom-grok and @ctliz/agent-intercom-agy packages. Those packages install agent-intercom-grok-mcp and agent-intercom-agy-mcp, depend on this package, and load its runtime internally. A separate global claude-intercom-mcp installation is not required. For a source-only checkout, configure node with the absolute path to dist/claude-server.mjs. Node 22.19+ is required.
Install the matching host package globally, then install its GitHub plugin at the matching release tag. Companion launchers may await the exported runtimeReady promise from dist/claude-server.mjs and call syncSession({ ...runtime.getIdentity(), name }) on that same runtime. This entry starts MCP stdio; do not create a second broker client to synchronize a host title.
Use AGENT_INTERCOM_SESSION_ID and AGENT_INTERCOM_SESSION_NAME for a host-neutral identity; CLAUDE_INTERCOM_SESSION_ID and CLAUDE_INTERCOM_NAME remain higher-priority compatibility aliases. Do not share a fixed session ID between concurrent host sessions.
A plain MCP host is connected only while its MCP process is alive. Incoming Intercom messages are durable and appear through intercom_pending, but this server cannot itself inject a new prompt into Grok or AGY. Use intercom_send plus intercom_pending for manual-polling workflows; a host-specific relay is required for wake-on-message.
For Grok Build, register the dedicated launcher without a fixed identity. It uses Grok's native GROK_SESSION_ID when available and synchronizes the exact session's persisted title:
grok mcp add agent-intercom \
-e CLAUDE_INTERCOM_MODEL=grok-build \
-- agent-intercom-grok-mcp
For AGY, add the equivalent entry to ~/.gemini/config/mcp_config.json:
{
"mcpServers": {
"agent-intercom": {
"command": "agent-intercom-agy-mcp",
"env": {
"CLAUDE_INTERCOM_MODEL": "agy"
}
}
}
}
Both hosts must use the same AGENT_INTERCOM_SCOPE_ID as their intended peers, or leave it unset for the default local scope. Their persistent MCP server registers eagerly, without a first prompt or tool call. AGY's optional native metadata publisher is described in its package README. Use intercom_whoami, intercom_list, intercom_send, and intercom_pending to verify identity and delivery.
Optional identity variables can be attached at registration time:
claude mcp add claude-planner \
--env CLAUDE_INTERCOM_NAME=planner \
--env CLAUDE_INTERCOM_SESSION_ID=claude-planner \
--env CLAUDE_INTERCOM_MODEL=opus \
-- claude-intercom-mcp
The repo also ships Claude Code plugin metadata:
.claude-plugin/plugin.json.mcp.jsonskills/claude-intercom/SKILL.mdcommands/intercom.md and commands/intercom-id.mdThe plugin packages the MCP server and the bundled claude-intercom skill (which
gives Claude copy-paste coordination patterns). It also installs these Claude
Code slash commands:
/claude-intercom:intercom [target and message] — list sessions and send a
message. Without arguments, Claude asks which peer to contact and what to send./claude-intercom:intercom-id — print this session's stable, copyable
intercom target.Claude custom commands are model-driven prompt commands, not native modal UI.
Claude namespaces plugin commands by plugin name, so an installed plugin cannot
claim the unqualified /intercom command globally.
They call the same MCP tools and work in a normal Claude Code session, but only
the attached cci/ccim wrappers can own the terminal and provide an immediate
Alt+M picker. Load the plugin for a single session with --plugin-dir:
claude --plugin-dir /path/to/agent-intercom-claude # this session only
For the minimal tool surface, prefer plain MCP registration
(claude mcp add claude-intercom -- claude-intercom-mcp) so you get the intercom
tools without the skill.
intercom_whoami: show this session's intercom ID, name, cwd, and model.intercom_team({ team? }): show all named task memberships, inspect one team, or fall back to managed-team discovery.intercom_join({ name?, create?, members?, work? }): add task-team membership without leaving earlier teams; managers add connected peers in one call.intercom_status: show connection status and pending message counts.intercom_list: list local Pi, Codex, and Claude sessions in your scope (protocol v4 is same-scope; cross-scope contact requires an exact full session ID).intercom_set_summary: publish a short discoverable status.intercom_send: send a non-blocking message.intercom_ask: send a question and wait for the target's reply.intercom_pending: read queued inbound messages and unresolved asks.intercom_reply: reply to an inbound ask or ordinary message via askId or contextId from intercom_pending; replies inherit the original team. Within one task, to plus which: "oldest" | "latest" also selects an ask.When the user delegates to named peers, the agent asks once whether to form a team for this task and waits for approval. An explicit create/join request is already approval; inbound team messages and approved tasks do not prompt again. After discovery, one call adds everyone:
intercom_join({ name: "launch", create: true, members: ["front", "writer"], work: "Launch page" })
intercom_send({ to: "front", team: "launch", message: "Implement the header." })
Membership is additive, with a separate manager/member role in each team. Reuse a team for the same task and append approved peers with intercom_join({ name: "launch", members: ["reviewer"] }); a different task may need a different team. Registration scope and earlier memberships stay unchanged. Without a shared team, omit team for initial ungrouped contact—even if either session belongs to unrelated teams. When several teams are shared, sends/asks require team; exact reply selectors retain the source team and cannot override it. AGY and Grok inherit these rules through the same MCP runtime.
Pending output never exposes protocol message IDs. Keep at most one unresolved intercom_ask to the same recipient; the broker rejects a second ask and recommends intercom_send for a non-blocking follow-up. Use intercom_send—not intercom_ask—for assignments and progress/status checkpoints.
Persistent Claude workers and plain MCP runtimes automatically reconnect their stable Intercom identity after a broker restart, so a live worker does not need to be respawned merely to become reachable again.
Example:
intercom_team({})
// Manager: manager-id [connected]
// You: worker-a
// Coworkers: reviewer target=reviewer (codex, reviewer, running) [connected]
intercom_ask({
to: "worker-a",
message: "Please inspect the failing test and reply with the likely cause.",
timeout_ms: 45000
})
Blocking asks default to a short bounded wait and reject waits over 120 seconds.
For longer work, use intercom_send and check later with intercom_pending.
ccicci (Claude Code Intercom) starts a single wakeable worker in the foreground.
It registers the worker on the broker. For every inbound message, the attached
terminal visibly prints the sender and message, a working indicator, and the
final Claude result or error. Blocking asks still receive that final result as
their automatic intercom reply. Press Alt+M for a numbered list of connected
peers, then choose one and enter a message. Press Alt+I to copy the worker's
contact target.
This is an attached worker console, not Claude Code's interactive TUI: woken
turns run through claude -p, and their final output is mirrored into the
console. To continue or inspect the full Claude conversation, run claude --resume <session-id> using the session ID printed with the completed turn.
ccim has the same visible wake behavior and shortcuts.
Start a named worker:
cci --name worker-a --id worker-a
Flags (all optional; ccim accepts the same set):
cci --cwd /path/to/project --instructions "Reply tersely. Ask before destructive changes."
cci --model opus --effort max --name reviewer --id reviewer
cci --yolo --name trusted-worker --id trusted-worker # explicit opt-in only
cci --add-dir ../shared-lib --name worker-a --id worker-a
By default cci passes the standard --permission-mode manual; it never adds
--dangerously-skip-permissions on the user's behalf. Headless turns cannot
answer an interactive permission prompt, so choose a validated explicit mode
when a different non-interactive posture is required. --yolo remains an
explicit trusted-user opt-in outside hardened roles.
cci --tui)Run cci --tui --channel --name reviewer --id reviewer to receive inbox notifications through Claude Code Channels. From a built local checkout, use node dist/cci.mjs --tui --channel --name reviewer --id reviewer. This is an explicit interactive MCP mode; it launches one bare claude-intercom server with CLAUDE_INTERCOM_CHANNEL=1 and --dangerously-load-development-channels server:claude-intercom. It does not load the plugin's Monitor path or enable permission relay. Existing Claude settings still apply; do not configure another Intercom MCP owner for the same identity.
Claude shows its local-development confirmation at startup. Channels are a research preview and require supported Anthropic authentication and any applicable organization policy. The launcher preserves normal tool permissions; the development flag only selects this local channel. See the official Channels reference.
After MCP initialization, unread messages trigger one fixed automated notification per new burst. Claude reads the full messages with intercom_pending({mark_read:true}) and replies with intercom_reply, preserving context and team. The notifier does not consume messages, put peer text into the notification, create a second identity, or replay inbox history. Messages received before MCP initialization are included. Host notification admission is not proof of model handling; unread memory lost on an MCP restart is not recovered.
Local Claude 2.1.296 was checked with an untouched idle session using the CPA gateway and gpt-6.1-sol: a broker message triggered a channel event, the model called intercom_pending, and intercom_reply returned CHANNEL_CPA_OK with the original ask correlation. Only those two Intercom tools were explicitly allowed for that test; ordinary tool permissions still apply. That host can display no MCP server configured with that name at startup for a server provided through --mcp-config, despite connecting and receiving channel events; use /mcp and the actual inbound event to inspect the connection.
Omit --channel to use the existing live transport selection below.
Default cci is a headless worker: each message spawns a claude -p turn. With
--tui, cci instead opens a live interactive Claude session that you sit in
and that is woken in place — the Codex coi experience. Inbound intercom
messages are injected into the running session and it replies over the broker;
you see everything and can type alongside it.
cci --tui --name worker-a --id worker-a
Claude Code has no Codex-style app-server. cci --tui therefore resolves one of two local transports before launch:
SendMessage tool so the bridge can preserve blocking ask/reply correlation. Native launches enable Claude's cross-session feature flag automatically. auto uses this only for the verified Claude Code compatibility window (currently 2.1.220–2.1.226). If native attachment fails under auto, cci restarts once with MCP; explicit --transport native fails closed instead.monitors/monitors.json to inject them with Claude Code's local Monitor mechanism. Blocking asks are answered with intercom_reply.Choose explicitly with --transport native or --transport mcp, or set CLAUDE_INTERCOM_TRANSPORT. Worker JSON entries also accept "transport": "auto" | "native" | "mcp". The Claude executable is probed with claude --version; unreadable or out-of-window versions never silently enable native mode.
--minimal is ignored in live TUI mode. The native path requires an interactive Claude process that publishes its local messaging socket. The MCP path additionally needs a built checkout (npm run build) and an available Monitor feature; Monitor is unavailable when DISABLE_TELEMETRY or CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC is set, or on Bedrock/Vertex/Foundry. Both paths remain local and work behind a custom ANTHROPIC_BASE_URL/proxy. See docs/wake-mechanisms.md.
The MCP server now registers as soon as it starts, without waiting for a prompt or an Intercom tool call. Bare MCP configuration provides discovery and queued messages; load the packaged plugin for hooks and automatic Monitor delivery:
claude --plugin-dir /path/to/agent-intercom-claude --name reviewer
On macOS/Linux, the plugin's SessionStart/SessionEnd hooks publish private
metadata only. The persistent MCP server is the only broker connection.
Without a launcher-provided ID, Intercom uses claude-<native-session-id>, so
resuming a conversation preserves its target and /clear releases the old
conversation before registering the new one. Explicit Claude/generic launcher
IDs retain their existing precedence.
cci --tui --name reviewer also passes Claude's native --name flag. /rename
updates broker presence without changing the stable ID. MCP mode incrementally
tails only native custom-title transcript records; it never replaces your
statusline configuration. Native mode follows the verified socket registry's
name. The native transport version gate remains 2.1.220–2.1.226; Claude
2.1.285 was smoke-tested through MCP, not added to that gate.
The Monitor skips historical inbox entries but retains messages that arrived
after SessionStart, even if they preceded Monitor startup. Unsupported host
process discovery falls back to eager MCP registration and explicit launcher
identity/inbox settings, as used by cci.
Enable Intercom once per Claude process: loading the plugin and a second manual
MCP entry for the same identity creates competing runtimes. SESSION_ID_IN_USE
leaves the incumbent authoritative, pauses automatic retries, and is exposed by
intercom_status. Do not share a fixed launcher ID between concurrent sessions.
Like Codex's coi (normal) and coim (minimal), cci has a minimal mode. Codex
needs a dedicated CODEX_HOME and a hand-written config.toml to strip
memories, plugins, skills, and browser surfaces (while keeping multi_agent).
Claude Code has this built in: cci --minimal runs every woken turn with Claude
Code's --safe-mode, which disables CLAUDE.md, skills, plugins, hooks, and MCP
servers while keeping auth, built-in tools (Bash/Read/Edit/…), and permissions
working normally. It is the focused-worker profile: less prompt and tool surface,
same coding ability.
Subagents are retained in minimal mode. --safe-mode only disables custom
agent-type definitions (.claude/agents/), not the built-in Task tool — so a
minimal worker can still delegate to general-purpose subagents, matching Codex
minimal's multi_agent = true. This is verified end-to-end
(test/e2e/minimal-subagent.sh): a minimal worker spawns a subagent that runs a
shell command and reports back.
cci and ccim are installed as a matched pair (like Codex's coi and coim):
ccim is exactly cci --minimal — same flags, same identity handling, minimal
by default. You do not need an alias to enable minimal mode; aliases are useful
only for reusable names, IDs, paths, or permission settings.
cci --name reviewer --id reviewer # normal: full config, CLAUDE.md, skills, MCP
ccim --name lean-worker --id lean-worker # minimal: --safe-mode woken turns
ccim --safe --name lean-safe --id lean-safe # minimal + standard permission prompts
cci --minimal --name worker-a --id worker-a # equivalent to `ccim ...`
Because minimal mode disables MCP in the woken turn, a minimal worker cannot use the intercom tools to message other sessions itself — it still receives work and replies normally (the worker daemon captures its final message and sends the reply). Use a normal worker when you want the woken turn to reach out to peers on its own.
Use one Claude Code session as the manager and one or more cci workers.
Launch a worker in tmux:
tmux new-session -d -s worker-a 'cd /path/to/project && cci --name worker-a --id worker-a'
Then, from the manager session, delegate through the intercom tools:
intercom_ask({
to: "worker-a",
message: "Create a plan for adding retries to src/api/client.ts, then report your first step.",
timeout_ms: 60000
})
For non-blocking delegation, use intercom_send and check back with
intercom_pending. For a decision you need before continuing, use
intercom_ask.
Use claude-intercom-worker when you want one process to publish several
configured workers without a launcher per worker.
Create a config:
{
"statePath": "/path/to/intercom/claude-worker-state.json",
"claudeCommand": "claude",
"agents": [
{
"id": "claude-worker",
"name": "claude-worker",
"cwd": "/path/to/project",
"model": "sonnet",
"instructions": "Reply concisely. Ask before making destructive changes.",
"permissionMode": "manual"
}
]
}
Worker configuration validates permission modes and rejects permission flags
hidden in claudeArgs. A tightening-only bossRole hint of adversary or
council forces --bare, permissionMode: "plan", and a read-only ceiling;
permission-granting settings, agents, plugins, and appended argv cannot widen
it. This local hint does not enroll a Boss participant or expose a reviewer
tool; those surfaces stay unavailable until a protected Controller supplies the
binding, transport, and durable dispatch path.
Start it:
claude-intercom-worker --config "$HOME/.pi/agent/intercom/claude-worker.json"
Each worker's session_id is persisted in statePath, so later messages resume
the same Claude conversation. The daemon reads a single worker's config from the
environment when no config file is given (CLAUDE_INTERCOM_WORKER_ID, …_NAME,
…_CWD, …_MODEL, …_INSTRUCTIONS, …_STATE).
The PI_* names are shared with the Pi, Codex, and OpenCode adapters on purpose —
all four read the same broker location and ask-timeout so they interoperate.
git clone https://github.com/ctliz/agent-intercom-claude.git
cd agent-intercom-claude
npm install
npm run build
npm test
For MCP development, register the TypeScript source directly:
claude mcp add claude-intercom-dev -- npx --no-install tsx ./claude/server.ts
agent-intercom-pi is the Pi-native adapter with overlays and inline rendering.
agent-intercom-codex is the Codex MCP/plugin adapter plus wake-on-message Codex
app-server sidecars. This repository, agent-intercom-claude, is the Claude Code
MCP/plugin adapter plus wake-on-message headless claude -p workers.
agent-intercom-opencode provides the native OpenCode plugin.
All four vendor the compatible local broker/client protocol and share one broker socket, so a single session list spans Pi, Codex, Claude Code, and OpenCode.
Releases are automated from version tags. Update package.json, the lockfile when
present, and CHANGELOG.md on main, then push an annotated tag that exactly
matches the package version:
git tag -a vX.Y.Z -m "vX.Y.Z"
git push origin vX.Y.Z
The release workflow verifies that the tag points into main, runs typecheck,
tests, and the build, publishes the public npm package with trusted OIDC
provenance, and creates the GitHub Release. Existing npm versions and GitHub
Releases are skipped safely when a workflow is rerun.
pi, claude, codex, and opencode — connect to one local broker over a Unix domain socket (~/.pi/agent/intercom/broker.sock or $PI_CODING_AGENT_DIR/intercom/broker.sock).@ctliz/agent-intercom-core is an internal dependency that arrives with the adapters and is never installed or upgraded on its own.agent-intercom-orchestrator is an optional Linux/systemd lifecycle component. It does not implement or start a Broker and is not part of the Broker compatibility set. Omitting it — for example on macOS, or when using TmuxDeck — is a fully supported configuration and is not a mixed or unsupported state. If it is installed on a supported Linux host, or on WSL with a systemd user manager enabled, update it together with the adapters it manages.ctliz, so a pre-upgrade backup of the exact installed specs/locks is the supported rollback material. Leaving some installed broker-capable adapters on the old protocol while others are upgraded is an unsupported mixed state.The current project is licensed under the GNU Affero General Public License
v3.0 or later (AGPL-3.0-or-later). If you modify this software and
make the modified version available to users over a network, the AGPL requires
you to offer those users the corresponding source code.
Portions derived from the original MIT-licensed pi-intercom project retain
their original notices. See THIRD_PARTY_NOTICES.md and
licenses/MIT-pi-intercom.txt. Versions already
published under MIT remain available under their original terms. See
LICENSE_TRANSITION.md for the exact commit and tag boundary.
The v4 release line renames the package namespace from @dataforxyz/* to @ctliz/*. The two namespaces are different packages to npm. For Claude Code, 0.12.0-connect.3 packages monitors/monitors.json alongside the plugin; other installed adapters update to their compatible v4 release (e.g. connect.2). Pi Git package installations deduplicate by repository URL without ref, but running agent sessions continue to execute legacy code in memory, and npm or global installs along with binary links can coexist and conflict. Operators must stop active sessions, clean active install surfaces, and follow remove-before-install — side-by-side installation is not supported.
@dataforxyz/* specs, packages, and binary links that are actually installed.node_modules installations, and conflicting binary links that the current PATH would resolve. Do not scan or delete unrelated source checkouts, historical documentation, or other users' files — a @dataforxyz/* string in an unrelated development clone is not an installation.@ctliz/* packages for the components you actually use (e.g. npm install -g @ctliz/agent-intercom-claude@connect / v0.12.0-connect.3, and companion connect.2 releases for other harnesses).Classification rule. Migration-aware setup and update tooling classifies an old-namespace-only install surface as MIGRATION_REQUIRED, and the simultaneous presence of both namespaces as a duplicate/dual-load hard error that refuses setup, update, and further installation. This tooling does not exist for every platform and adapter combination; where it is not available, apply the same two rules manually against the surfaces in step 4. Do not assume every adapter emits this code automatically.
Rollback reverses this and covers only the components that were installed on this machine before the upgrade: remove the @ctliz/* packages, then restore the backed-up exact @dataforxyz/* specs and locks. Roll Orchestrator back only if it was installed to begin with.
The connect.1 tags, source commits, and published release assets are immutable and are not modified by this migration. Release notes may carry an explicit erratum, which corrects the description only and never moves a tag or replaces an asset.
Claude Code MCP plugin and wakeable worker for local messaging with Pi, Codex, OpenCode, Grok Build, and AGY agents.
{ "mcpServers": { "ctliz-agent-intercom-claude": { "command": "npx", "args": [ "-y", "@ctliz/[email protected]" ] } }}工具
暂无工具列表。远程服务器在首次健康检查后会获得工具列表。
常见问题
在本页复制适用于 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 配置。它使用发布者的安装方式，而不是 PlugMCP 代理。
此记录未列出必需的环境变量。
@ctliz/agent Intercom Claude 通过 stdio 在本地运行。
上次检查时间：2026-11-08。 Claude Code MCP plugin and wakeable worker for local messaging with Pi, Codex, OpenCode, Grok Build, and AGY agents.