agentbridge CLI

agentbridge is the command-line interface for the agentbridge general-purpose agent interconnect. It registers agents (CLI coding tools like Claude Code, Copilot, Codex, Cursor Agent, or any subprocess that speaks the agentbridge JSON protocol) as live A2A services, lists them via DIR discovery, hands off context between them, and coordinates them autonomously toward a shared goal.

Install

cargo install --path crates/agentbridge_cli

Commands

register — start an adapter server

Wrap a CLI tool as an agentbridge adapter and keep it running as an A2A service so remote callers can reach it over SLIMRPC and/or official A2A unicast (gRPC, JSON-RPC, or HTTP+JSON).

Security: register --slim-endpoint and register --a2a-listen refuse to start unless they're running under a SHADI sandbox with network blocked by default — wrap them in shadictl, as shown below, and --read the directory holding certificates ($SHADI_TMP_DIR/shadi-slim-mtls for SLIM, $SHADI_TMP_DIR/shadi-a2a-tls for non-loopback unicast) so the listener can still read its own cert under the sandbox. On macOS, resolve $SHADI_TMP_DIR to its real path first (cd "$SHADI_TMP_DIR" && pwd -P) — /tmp is a symlink to /private/tmp, and Seatbelt's sandbox rules don't match a path reached through the symlink if the rule was generated for the canonicalized form. See Environment variables.

# Wrap any subprocess that speaks the agentbridge JSON protocol
shadictl --net-block --net-allow 127.0.0.1:47357 --read "$SHADI_TMP_DIR" -- \
  agentbridge register \
  --tool generic-stdio \
  --command my-tool \
  --arg --agentbridge-mode \
  --slim-endpoint 127.0.0.1:47357

# Start a Claude Code adapter
shadictl --net-block --net-allow 127.0.0.1:47357 --read "$SHADI_TMP_DIR" -- \
  agentbridge register --tool claude-code --slim-endpoint 127.0.0.1:47357

# Start a Copilot adapter
shadictl --net-block --net-allow 127.0.0.1:47357 --read "$SHADI_TMP_DIR" -- \
  agentbridge register --tool copilot --slim-endpoint 127.0.0.1:47357

# Start a Codex adapter
shadictl --net-block --net-allow 127.0.0.1:47357 --read "$SHADI_TMP_DIR" -- \
  agentbridge register --tool codex --slim-endpoint 127.0.0.1:47357

# Goose / OpenCode use the operator's existing CLI config. Host
# GOOSE_PROVIDER / GOOSE_MODEL are passed as goose run --provider / --model.
# Host GOOSE_*, OPENAI_*, and *_API_KEY env vars are forwarded to goose.
shadictl --net-block --net-allow 127.0.0.1:47357 --read "$SHADI_TMP_DIR" -- \
  agentbridge register --tool goose --slim-endpoint 127.0.0.1:47357
shadictl --net-block --net-allow 127.0.0.1:47357 --read "$SHADI_TMP_DIR" -- \
  agentbridge register --tool opencode --slim-endpoint 127.0.0.1:47357

# Publish an OASF record to the Agent Directory after registering
agentbridge register --tool claude-code --dir-publish

Supported --tool values: generic-stdio, claude-code, copilot, codex, cursor-agent, goose, opencode.

list — discover registered adapters

# Query DIR for registered adapters
agentbridge list

# List listeners this machine started with register --slim-endpoint / --a2a-listen
agentbridge list --local

handoff — transfer context from one tool to another

Snapshot the current session from a source tool and inject it into a destination tool. Bare names (claude-code, copilot, …) open local adapters (JSON protocol / native CLI). slim:<id> is A2A over SLIM: the process proves as SHADI_AGENT_ID and SendMessages to the peer. When SHADI_AGENT_ID matches --from slim:<id>, that agent is the A2A client — snapshot stays local so it does not call its own listener.

# Native tools (same specs as coordinate)
agentbridge handoff --from claude-code --to copilot

# Finishing agent hands off as an A2A client (collab demo)
SHADI_AGENT_ID=claude-code agentbridge handoff \
  --from slim:claude-code --to slim:copilot --slim-endpoint 127.0.0.1:47591

# Basic handoff using subprocess commands directly
agentbridge handoff \
    --from ./my-source-tool \
    --to ./my-dest-tool

# Save the captured ContextPacket for recovery
agentbridge handoff \
    --from ./my-source-tool \
    --to ./my-dest-tool \
    --save /tmp/context.json

# Resume from a previously saved ContextPacket
agentbridge handoff \
    --from-file /tmp/context.json \
    --to ./my-dest-tool

delegate — send a single task to a remote adapter

Dispatch one prompt to a remote agentbridge adapter over A2A and print the response. --to is a DID (did:key:…) or a local alias from list --local. The URL is a locator: delegate looks it up from the current lease (or DIR). --a2a-url overrides the locator only and must be paired with --to did:key:…. Locator URIs (jsonrpc://host:port) carry the binding; a bare http:// still means gRPC unless --a2a-binding is set. https:// gRPC client TLS is not wired yet.

agentbridge delegate "write unit tests for src/parser.rs" \
  --to did:key:z6Mk… \
  --agent-id avatar

# Constrained unicast without a SLIM node (loopback plaintext)
shadictl --net-block --net-allow 127.0.0.1:50051 -- \
  agentbridge register --tool copilot --a2a-listen 127.0.0.1:50051
# JSON-RPC / HTTP+JSON: add --a2a-binding jsonrpc or --a2a-binding http+json

agentbridge list --local
agentbridge delegate "write unit tests for src/parser.rs" \
  --to did:key:z6Mk… \
  --agent-id avatar

coordinate — autonomous multi-round coordination

Run MasRuntime<DevelopmentEngine> across a set of agents until a winning code artifact is produced. Agents propose, vote, and converge without human mediation.

# Local agents (in-process subprocess adapters)
agentbridge coordinate \
  --goal "implement a JSON parser in Rust" \
  --agents claude-code,copilot,codex,cursor-agent \
  --quorum 3 \
  --max-rounds 5 \
  --output result.rs

# Remote agents over SLIM A2A (register each in a separate terminal first)
agentbridge coordinate \
  --goal "implement a JSON parser in Rust" \
  --agents slim:copilot,slim:codex \
  --quorum 1 \
  --max-rounds 2 \
  --output result.rs \
  --slim-endpoint 127.0.0.1:47357

# Require explicit human approval before accepting the result
agentbridge coordinate \
  --goal "refactor auth module" \
  --agents claude-code,copilot \
  --quorum 2 \
  --require-human

Agent spec formats for --agents:

Subprocess protocol

Any process can participate in a handoff if it reads JSON from stdin and writes JSON to stdout:

stdin:   {"cmd":"snapshot"}
stdout:  {"ok":true,"data":{... ContextPacket JSON ...}}

stdin:   {"cmd":"inject","context":{... ContextPacket JSON ...}}
stdout:  {"ok":true}

stdin:   {"cmd":"execute","prompt":"write a parser"}
stdout:  {"ok":true,"data":"fn parse(...) { ... }"}

Environment variables

⚠️ Security: register, delegate, and coordinate (for slim: agent specs) authenticate to the SLIM mesh via DID/keys only — set SHADI_SLIM_AUTH=did, SLIM_HUMAN_SEED, and SLIM_MEMBER_DIDS (see docs/content/demos/demo-env.sh). Shared secrets are not supported: a register listener forwards incoming A2A tasks to the local CLI tool, so admission must be cryptographic, not a symmetric secret compiled into every demo script.

That decides who may send a task. It doesn't constrain what the task can do once it runs, so register --slim-endpoint and register --a2a-listen separately refuse to start unless they're running under a SHADI sandbox with network blocked by default — wrap them in shadictl --net-block --net-allow <listen-addr> --. Kernel sandboxes (Seatbelt/Landlock/AppContainer) are inherited by child processes, so this confines whatever CLI tool the adapter spawns with no extra code in agentbridge itself.

Live SLIM demo

See scripts/agentbridge_shell*.sh (DID admission, sandbox, list --local), the one-command walkthrough in docs/content/demos/did-agent-group.md, or the two-lines-per-turn coding loop in docs/content/demos/collab-rust.md. A harness should load skills/agentbridge and call these commands rather than a new per-CLI adapter.