agent-relay-tool

Agent Relay

Agent Relay is a local-first interactive layer for working with multiple coding agents in one durable project session.

It is built for the moment when one agent needs to stop because of rate limits, tool limits, or a manual handoff. Start inside Relay, talk to Claude, Codex, Gemini, Kimi, or OpenCode from the same shell, and Relay keeps the observable session context, decisions, validation state, and handoff packets on disk so the next agent can continue from the same point.

Built-in agent adapters currently support Claude Code, Codex, Gemini, Kimi Code, and OpenCode.

Why use it

• keep one durable session history across multiple agent turns and handoffs
• capture checkpoints with decisions, blockers, touched files, and validation state
• render agent-specific resume packets from the latest checkpoint
• switch agents manually or automatically when a rate limit is detected
• recover and inspect session state from on-disk journal data

Installation

Python 3.11 or newer is required.

macOS / Linux

curl -fsSL https://agent-relay.dev/install.sh | sh

Homebrew

brew tap bethvourc/tap
brew install agent-relay
relay install

Windows (PowerShell)

irm https://agent-relay.dev/install.ps1 | iex

Already using uv or pipx?

uv tool install agent-relay-tool   # or: pipx install agent-relay-tool

Full per-OS walkthroughs and verification steps: agent-relay.dev/installation.

Quick start

Run Relay inside the repository whose work you want to hand off:

cd /path/to/your/repo

# First-time local wiring: hooks, daemon, and fallback order
relay install

# Start the interactive layer
relay

Inside the REPL, bare text goes to the active agent and slash commands control Relay:

/use claude
Fix the failing tests

/use codex
Continue the release prep from the current Relay session

/handoff-order claude codex gemini kimi opencode
/status
/metrics

Scriptable commands are still available when you need a one-off or automation-friendly flow:

# One-command relay: hand off to another agent from outside the REPL
agent-relay codex --task "Continue the release prep"

# Single-agent managed run
agent-relay run c "Fix the failing tests"

# Turn-based conversation between agents
agent-relay chat c x "Fix the failing tests"

# Concurrent agents working simultaneously (tmux)
agent-relay race c x "Build the auth module"

# Resume the latest unresolved concurrent conflict
agent-relay resolve --latest

# Inspect saved conflict artifacts
agent-relay inspect-conflicts <session-id>

# See what agents are available
agent-relay discover

# View sessions
agent-relay status

# Live view of an in-progress session (auto-picks newest active)
agent-relay watch

# Token / cost / duration rollup
agent-relay metrics

Agent aliases: c = Claude, x = Codex, g = Gemini, k = Kimi, o = OpenCode. Use agent-relay discover to see all available agents and aliases.

Interactive REPL

Run relay with no arguments to drop into the persistent slash-shell:

relay                       # interactive REPL (the default when no subcommand is given)
relay --tmux                # wrap the REPL in a tmux session for crash-safety
relay --attach              # reattach to this project's relay tmux session
relay --no-tmux             # skip the first-run tmux prompt for this run
relay --no-onboarding       # skip the first-run wizard (useful in CI/scripts)

Inside the shell:

• Type / to live-filter the slash menu (every CLI subcommand is available as a slash command). Tab accepts, Enter submits.
• Type ? on an empty input to open a shortcut overlay; Esc closes overlays.
• Bare text (no leading /) is forwarded to the active agent. Relay persists the prompt/output artifacts into the active repo-local session and uses that same session lineage for later handoffs.
• Switch agents with /use claude, /use codex, or /use gemini. If there is live context, Relay routes the switch through the handoff machinery so the target starts with a resume packet.
• Pick the active agent's model with /model. The picker lists the models your signed-in account can actually use, asking the agent's own tooling where it can (Codex's model cache, agy models, opencode models) and falling back to Relay's suggestions otherwise. /model <name> sets one directly and /model default clears it; an unrecognised name is still accepted, since providers ship models faster than Relay tracks them. Results are cached for a few minutes, so use /model --refresh after changing your account's models.
• Keep typing while the agent works. What you type appears on the line below the spinner and Enter queues it; what is waiting is listed above, and ↑ on an empty line pulls the newest one back to edit before it is sent. Queued messages run in order once the turn finishes. Only submitted lines run — a half-typed line carries into the next prompt instead of being sent, and a multi-line paste stays one message. If a turn fails or you press Ctrl+C, the queue is discarded and printed back so nothing is silently lost.
• Prefix a line with ! to run it as a shell command in the repo root without leaving the session (!git status). Each ! is a fresh shell; output lands in the transcript, so it feeds handoffs. Typed mid-turn, a ! command runs between turns.
• Run /help for every command, each with an example, or <command> --help (also -h, and aliases like /c -h) for one command's arguments and flags.
• Ctrl+C clears the input on first press; a second press within ~2 s exits.
• Ctrl+D exits immediately.

On first launch Relay runs a brief wizard (detects agents, asks about tmux), persists your choices to ~/.config/relay/config.toml, and never asks again. Re-run anytime with /setup. State and logs live under ~/.config/relay/:

For full details:

• docs/repl.md — every slash command (auto-generated from the registry; CI ensures this never drifts).
• docs/architecture.md — REPL internals: shell loop, agent layer, env contract, observability.
• docs/security.md — input validation, env allowlist, redaction pipeline.
• docs/performance.md — keystroke-to-render budgets and how to run the micro-benches.
• docs/troubleshooting.md — orphan reaper, log location, common gotchas.

Best practices

Start inside the interactive Relay shell for the strongest handoffs

The single most important thing you can do for better handoffs is start the work inside Relay instead of only calling Relay after an agent stops.

When Relay manages the session from the beginning, the REPL owns the live agent process, forwards your bare prompts to the active agent, captures the observable output stream, and persists turn artifacts under .agent-relay/sessions/. This means the handoff packet contains everything Relay can observe or export: recent conversation history, the current plan, blockers, remaining work, intended edits, and any provider-exported session state.

When Relay joins after the fact, it can only hand off what is still observable: the working tree changes, any notes you provide, and whatever the provider can export at that moment. Private reasoning, in-progress drafts, and tool-call history from an unmanaged session are lost.

In simple terms: if Relay was there while the work was happening, handoffs are much stronger. If Relay joins later, it can only hand off what it can still see.

Use relay as the default starting point

relay

relay is the recommended way to start day-to-day work. It:

• keeps a long-lived active agent process for the current repo
• stores each prompt/output turn in the active Relay session
• restores visible session transcript when you resume a prior session
• hands off through resume packets when you switch agents with /use
• listens for daemon rate-limit handoff events when relay install has wired hooks
• auto-compacts older saved context when the session grows large

Use run when you need a scriptable single-agent command

agent-relay run c "Fix the failing tests"

run is still useful for CI, scripts, or one-off single-agent tasks. It:

• runs the agent with live output capture (reasoning, tool calls, decisions)
• extracts structured state from each turn (status, remaining work, blockers)
• stores full turn artifacts (prompt, output, stderr, state) for later recovery
• builds continuation context automatically so the next agent picks up exactly where this one left off

If the agent finishes the work, great. If it hits a rate limit or needs to hand off, the session already has everything Relay needs to produce a strong resume packet.

Use chat for multi-agent collaboration

agent-relay chat c x "Fix the failing tests"

chat runs agents in alternating turns. Each agent sees the full conversation history — what every other agent said, decided, and proposed. Use this when agents need to build on each other's work iteratively, like one agent investigating and another implementing.

Use race for parallel work with enforced delegation

agent-relay race c x "Build the auth module"

race is a phased concurrent workflow, not just two agents launched side by side. It enforces a structured process:

1. Planning: every agent must claim a concrete slice of work before implementation begins.
2. Implementation: each agent works inside its own isolated git worktree, so changes cannot collide during execution.
3. Merge and review: Relay only merges in-scope work back to the main repo.
4. Conflict handling: Relay saves conflict artifacts, can run an automatic resolver/reviewer pass, and hands off to resolve when human judgment is still needed.

Fall back to one-command relay when Relay was not managing the session

agent-relay codex --task "Claude hit its limit; continue from the current state"

If an agent was working outside of Relay and needs to hand off, open a new terminal in the same repo and run a one-command handoff. Relay will capture what it can (git changes, any notes you provide) and generate a packet for the target agent. This is less complete than a managed session but still far better than starting from scratch.

What Relay can recover

When Relay managed the session (REPL, run, chat, race)

• full conversation turn history (prompts, outputs, reasoning)
• structured state from each turn (status, plan, blockers, remaining work)
• intended and proposed edits (even if not yet applied)
• provider-exported session state when available
• current working tree changes
• verification items and validation results

When Relay joins after the fact (one-command handoff)

• current working tree changes (git diff)
• planning notes you provide via --planning-note-file
• proposed edits you provide via --proposed-edits-file
• anything the provider can export at handoff time

What Relay cannot recover

• private hidden reasoning from an unmanaged session
• UI-only drafts that were never saved anywhere
• proposed edits shown in a tool UI but never accepted, exported, or passed to Relay
• hidden OpenCode state unless OpenCode exposes it through opencode export

Commands

Primary commands

Conflict resolution

Session management

Options

Recommended workflows

Single-agent task with potential handoff

# Start inside Relay for the strongest handoff later
agent-relay run c "Fix the failing tests"

# If the agent hits a limit, the session is already captured.
# Hand off to another agent:
agent-relay codex --task "Continue from the saved session"

Turn-based collaboration

# Two agents alternating
agent-relay chat c x "Fix the failing tests"

# Three agents, 6 turns max
agent-relay chat c x c "Review and fix" -n 6

Live monitoring of a running session

While a session is in flight, open a second terminal in the same repo and run:

# Auto-pick the newest active session
agent-relay watch

# Or pin to a specific session
agent-relay watch <session-id>

# Stream events as JSONL — pipe into other tools
agent-relay --json watch | jq -c '{ts: .timestamp, kind: .kind}'

# Print a single snapshot of current state and exit
agent-relay watch --no-follow

The live TUI surfaces journal events, workspace activity, the current turn's elapsed time and progressive agent output, and the latest turn state — all in real time. The watcher exits cleanly when the session reaches a terminal status (completed, ready_for_handoff) or on Ctrl-C.

Add --metrics for a token / cost / duration panel that refreshes after every turn:

agent-relay watch --metrics

Cost & performance metrics

Relay computes metrics on read — no extra files in .agent-relay/, no cache invalidation. The same data backs four surfaces:

# Per-session table (auto-picks the most recent session)
agent-relay metrics

# Cross-session aggregates: by agent, by day, totals
agent-relay metrics --all
agent-relay metrics --all --since 2026-05-01 --agent claude

# Machine-readable
agent-relay --json metrics
agent-relay --quiet metrics      # minimal output

# Live JSONL stream — one line per turn_completed plus a final session rollup
agent-relay metrics-tail
agent-relay metrics-tail --webhook https://hooks.example.com/relay

Cost is best-effort. If an agent output includes an actual total_cost_usd style field, Relay uses it. Codex exec --json normally emits token usage but not a billed cost, so Relay estimates Codex cost from the captured model name. Managed Codex turns save the model from the JSON stream when available, then fall back to AGENT_RELAY_CODEX_MODEL, CODEX_MODEL, OPENAI_MODEL, or $CODEX_HOME/config.toml. Uncached input tokens, cached input tokens, and output tokens are priced separately. Estimates do not include account discounts, Batch/Flex or priority processing, data residency uplift, long-context uplift, or separately billed tool fees.

For dashboards and long-running collection, metrics-serve runs an exporter. At least one of --prometheus or --otlp is required — running it with no exporter flag exits with an error:

# Prometheus pull-based scrape endpoint (stdlib only)
agent-relay metrics-serve --prometheus :9464
# → curl http://localhost:9464/metrics

# OTLP push to a collector (HTTP/JSON, every 30s by default)
agent-relay metrics-serve --otlp http://localhost:4318/v1/metrics

# Both can run together
agent-relay metrics-serve --prometheus :9464 --otlp http://localhost:4318/v1/metrics

Metrics emitted (Prometheus naming; OTLP uses dotted equivalents):

Alerts

Threshold-based alerts ride the metrics-tail channel. They are opt-in — no config file, no alerts. Drop a .agent-relay/config/alerts.toml:

cost_per_turn_usd = 0.50
cost_per_session_usd = 5.00
duration_per_turn_ms = 300000
tokens_per_turn = 200000
error_rate_threshold = 0.4    # ratio 0..1; gated by error_rate_min_turns
error_rate_min_turns = 5

When a turn breaches a threshold, metrics-tail writes a colored line to stderr and emits an extra JSONL line on stdout:

{
  "kind": "metrics.alert",
  "rule": "cost_per_turn",
  "severity": "warning",
  "session_id": "...",
  "turn_number": 3,
  "threshold": 0.5,
  "observed": 0.71,
  "message": "turn 3 cost $0.7100 exceeds threshold $0.5000",
  "timestamp": "..."
}

Severity is critical at ≥ 2× threshold, warning otherwise. Webhook delivery (when --webhook is set) carries alert lines too, so external systems can react without polling.

Concurrent work with conflict recovery

# Start a concurrent run
agent-relay race c x "Build the auth module"

# Continue an interrupted or timed-out concurrent session
agent-relay race --continue <session-id> c x "Continue the task"

# Inspect saved conflict artifacts and versions
agent-relay inspect-conflicts <session-id>

# Resume unresolved conflict resolution
agent-relay resolve <session-id>
agent-relay resolve --latest

Claim roles in race:

• owner: exclusive editor for that path or directory
• shared: multiple agents may edit that scope intentionally
• reviewer: review-only overlap; edits in reviewer-only scope are blocked

Notes:

• On macOS, Relay can auto-open one terminal window or tab per tmux session. Use --open-terminals or --no-open-terminals to control that behavior explicitly.
• If a concurrent run ends in manual_resolution_required, use inspect-conflicts first to see the saved versions, then resolve to continue the resolution workflow.
• If a concurrent run ends in max_time, interrupted, incomplete, or agent_error, use race --continue <session-id> ... to continue the broader task.

REPL permission backends

Relay now keeps repo-local permission backend config in:

.agent-relay/permissions.toml

From the REPL:

/permissions
/permissions set claude mode bypassPermissions
/permissions set codex approval_policy on-request
/permissions set codex sandbox_mode danger-full-access

Important limitation: this currently controls launch-time permission behavior for turn-mode agents like Claude and Codex. It does not yet provide generic cross-agent per-tool approval unless the underlying CLI exposes pending tool requests in a machine-readable way.

Preserve planning-only work

If there are no file changes yet, pass the planning and proposed-edit context explicitly:

agent-relay codex \
  --task "Continue from the saved plan" \
  --planning-note-file handoff-notes/planning.md \
  --proposed-edits-file handoff-notes/proposed.diff \
  --no-launch

This is the best fallback when an agent did useful planning but did not write code to disk.

Configuration

Capture helpers

Launch template overrides

Permission backend overrides

Capture hook templates

Relay-managed OpenCode turns always carry Relay's observable artifacts (prompts, outputs, and turn state) into handoff packets. When Relay can determine the native OpenCode session id, it also runs the default sanitized export command opencode export {session_id} --sanitize and cites the result as a provider session export. If OpenCode export fails or the native session id is unavailable, handoff continues with Relay's observable artifacts.

Template placeholders

Available in both launch and capture templates:

Built-in defaults

claude: cd {repo_root} && claude -p "$(cat {resume_path})"
codex:  cd {repo_root} && codex "$(cat {resume_path})"

Custom launch templates must include {resume_path} or {resume_path_path}. If a template omits the packet input, launch --execute refuses to run it.

Capture templates are optional. If set, they should print JSON to stdout with any of these fields:

• resumable_state
• planning_snapshot
• proposed_edits
• transcript
• session_metadata
• warnings

Storage model

Agent Relay writes repo-local state under .agent-relay/.

Each session lives under .agent-relay/sessions/<session-id>/ and uses a journal-plus-objects layout:

• session.json: immutable session manifest
• journal/: append-only event log
• objects/checkpoints/<checkpoint-id>/: checkpoint manifests, summaries, repo-state captures, and related artifacts
• objects/handoffs/<handoff-id>/: resume packet, packet hash, and launch specification
• objects/launches/<launch-id>/: launch receipts and captured stdout/stderr
• refs/head.json: latest derived head pointer
• derived/view.json: current materialized session view
• recovery/: pending transactions, quarantine, and repair reports
• turns/<turn-number>/: turn artifacts from run/chat sessions (prompt, output, stderr, state)

This keeps the session state inspectable and vendor-independent.

What not to publish

Do not commit or publish .agent-relay/. Session artifacts can contain:

• absolute local paths
• research notes and implementation notes
• validation summaries
• captured Git status, workspace patches, and untracked-file manifests
• generated resume packets
• rendered launch commands and launch-template text
• full agent conversation output (from run/chat sessions)

Two practical rules:

• add .agent-relay/ to .gitignore
• do not put secrets or tokens into AGENT_RELAY_CLAUDE_LAUNCH_TEMPLATE or AGENT_RELAY_CODEX_LAUNCH_TEMPLATE, because the rendered command and template are recorded in handoff metadata