Anywhere CLI

v2 Connector. Use the source from the same release line as your Server. The Python package version is independent of the 2.0.0 product version.

Local runtime connector for Agents Anywhere. It runs on the machine that owns the workspace and agent runtimes, connects to the server over HTTP/WebSocket, executes connector RPC locally, and uploads normalized runtime/session state back to the backend.

Layout

connector/
  runtime_protocol/  AgentRuntime, RuntimeProvider, RuntimeHostClient contracts
  runtimes/          Codex, Claude and DSH RuntimeProvider/AgentRuntime packages
  server/            Backend auth, ingest, RPC channel, request dispatch, host mapping
  core/              Connector config, JSON-RPC, runtime owner, runtime config storage
  local/             Local filesystem, shell, and terminal backends
  _reference/        Old adapter implementations retained only as migration references
  cli.py             anywhere-cli CLI
  control.py         Local desktop/control JSON-RPC entrypoint
tests/          Connector tests
pyproject.toml  Connector dependencies and console script
run.sh          Local helper for saved-config startup

Run

Run from this repository's connector/ directory. This avoids depending on an unverified public package version when connecting to a v2 Server. Install dependencies:

uv sync

Start with explicit credentials from the web pairing flow:

uv run anywhere-cli start \
  --server-url http://127.0.0.1:8000 \
  --connector-id conn_xxx \
  --connector-token cxt_xxx

Or save the config locally and start without arguments:

uv run anywhere-cli configure \
  --server-url http://127.0.0.1:8000 \
  --connector-id conn_xxx \
  --connector-token cxt_xxx

uv run anywhere-cli start

The default config path is ~/.agents-anywhere/connector.json. Override it with --config or AGENT_CONNECTOR_CONFIG.

Connector configuration, runtime ownership, sync state, and attachments all live under ~/.agents-anywhere by default. Runtime sync cursors are stored as atomic JSON in <connectorId>/<runtimeId>/sync-state.json; runtime message bindings use the adjacent kv.json. Connector prepares these stores before constructing a runtime and retains all opened stores for periodic/shutdown flush, including stopped instances. Existing sync and KV keys and read/write interfaces are unchanged, including runtime source-key isolation. Connector does not use SQLite. On first use, the v2 connector performs a one-time local data migration from the old ~/.agent-server directory into ~/.agents-anywhere and discards obsolete SQLite sync state.

When an instance directory does not exist, startup copies the entire legacy connector-state.json and connector-kv.json into a temporary directory and publishes it after both files validate. Legacy files remain untouched apart from flushing pending committed sync state before copying. Existing instance directories are never recopied or merged; reads and deletes do not fall back to legacy data. Copied files retain their old namespaces, so foreign instance records are not selected by that runtime's normal keys. A failed migration blocks that runtime's startup and can be retried. Agent-native histories and the machine ownership record are not relocated.

Codex recovery uses an optional read-only native history index to avoid loading unchanged message bodies. For standalone paginated histories, it verifies that thread_history_1.sqlite has consumed the current rollout, then compares every turn's metadata and item update ordinals with its committed checkpoint. Unchanged, settled history needs no history RPC after restart. New turns normally require one 20-turn page; the previous tail is rechecked. Changes to older turns read back to the earliest changed turn. Prefix item counts preserve timeline ordering.

The index is an optimization, not another data owner: it is opened read-only and never repaired or migrated by AA. Unknown schemas, lagging projections, inherited histories, compaction, missing sources or invalid checkpoints use the full history RPC. Deletion, reordering and source replacement require full calibration. Changes during a paginated read abort the checkpoint commit and retry on the next scan. Native file identity, size and nanosecond mtime supplement the API's second-resolution change marker. Index validation and checkpoint metadata still scale with history size, but unchanged message bodies are neither fetched nor projected.

Each projected item is compared with its last successfully ingested fingerprint in sync-state.json. Unchanged items are omitted, including after reconnect/restart; new and modified items are sent as a delta. First sync sends all items. Item removals or incompatible checkpoint versions use a session replacement snapshot. Replacement is deferred while a session is active, without committing its checkpoint. Failed ingestion never advances the prepared fingerprint state. Live notifications do not advance this scanner checkpoint, so the latest live items can be safely resent once by the next successful scan. Fingerprints scale with the number of timeline items. Recovery progress is tracked per session, so one failing session does not force all successful sessions to reread their history on every poll. Codex's active-writer conflict is reported as a takeover failure before sending the message; AA cannot silently displace another native client holding the writer lock.

Local startup ownership

All CLI, Desktop and DSH plugin launches use Python's per-user startup check at <OS user home>/.agents-anywhere/connector-runtime.json, independent of private configuration or data paths. Python records the actual Connector PID, its startup source and process start time. A record blocks startup only while its PID still identifies that Connector process. A live unrelated process or a reused PID does not block startup; an inspection permission failure is reported instead of bypassed.

Each accepted configured start appends its Connector ID once to the ordered history, including CLI launches and existing bindings. Python removes only its own runtime record at normal shutdown. A crash leaves a record that the next start checks against the actual process. Stopping the backend connection through connector.stop retains ownership while the RPC process is alive.

RPC callers receive -32009 with data.reason = connector_already_running on a conflict. connector.acquireOwnership supports preflight before credentials exist; a rejected request keeps the RPC channel alive for connector.getState and retry. Direct CLI startup reports the conflict and exits with code 2.

Desktop remains the only installation-metadata writer; the plugin only reads it. Neither host appends shared IDs or checks startup PIDs. See the local machine v2 contract for fields, atomic file transactions and legacy migration.

Runtime Discovery

The default providers are Codex, Claude and DSH. The connector reports attached runtime capabilities to the server. Codex is discovered through the official openai-codex SDK package; the connector does not use a Codex CLI/app-server path or IPC switch as an active runtime surface. If Claude Code is not on PATH, set:

CLAUDE_BIN=/path/to/claude

DSH requires the bridge integration described in DSH Bridge Next. Legacy ACP adapters are not part of the default provider registry.

The connector uses local runtime credentials and local filesystem permissions. Agents Anywhere does not proxy Claude or Codex account credentials.

API Namespace

Connector configuration stores the server origin, for example http://127.0.0.1:8000; do not include /api/v2 in --server-url.

The connector adds the v2 namespace internally and talks to /api/v2/connector/* and /api/v2/health. See ../docs/api/namespace.md for the namespace rules.

Local Operations

The server can ask an online connector to perform local work:

• read/list/write files inside workspace-safe roots
• upload/download file content through the server
• run one-shot shell commands
• start and wait for shell tasks
• create, write, resize, stream, list, and close interactive terminals
• start, interrupt, sync, and approve runtime turns

Environment

Verify

uv run ruff check connector tests
uv run pytest -q

DSH 插件默认将私有数据存放在 ~/.agents-anywhere/dsh-bridge-next/，其托管 Connector 通过 AGENT_CONNECTOR_DATA_DIR 使用其中的 connector/ 子目录。插件首次启动负责迁移旧 .agentsanywhere/dsh-bridge-next/ 数据；通用 Connector 的默认路径和自定义环境变量行为不变。