Claude Code 2.1.142: --agents flag, Fast Mode, and Opus 4.7 — what changed and why it matters

"2.1.142 dropped 24 hours after 2.1.141. What do the 8 new flags on claude agents actually do? And what does it mean that Fast Mode is now Opus 4.7 by default?" This article walks through the four primary change groups in Claude Code 2.1.142 (released 2026-05-14, 22:55 UTC) using the official GitHub Releases as primary source, with a focus on AI sales automation and what it means for Sales Claw.

On 2026-05-12, Anthropic shipped 2.1.139 (/goal + Agent View); later that day, 2.1.140 (subagent_type improvements); on the night of the 13th, 2.1.141 (Rewind + Workload ID); and on the night of the 14th, 2.1.142. That's four releases in 72 hours. 2.1.142 doesn't have a flashy headline feature, but it centers on two things that hit operations directly: eight new flags on claude agents for configuring dispatched background sessions, and the Fast Mode default model graduating to Opus 4.7.

At a glance, 2.1.142 has four primary change groups:

1. Eight new claude agents flags — --add-dir, --settings, --mcp-config, --plugin-dir, --permission-mode, --model, --effort, --dangerously-skip-permissions — all configurable per dispatched background session
2. Fast Mode default is now Opus 4.7 — promoted from Opus 4.6; roll back with CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE=1
3. Plugin / Skill detection becomes more flexible — a Plugin with just a root-level SKILL.md is now surfaced as a Skill. The /plugin detail pane also shows LSP server info
4. Background session stability + broader MCP_TOOL_TIMEOUT coverage — git worktree detection, macOS sleep/wake recovery, post-upgrade crash-loop fix, the 60-second cap on remote HTTP/SSE MCP servers removed, and 15+ other fixes

Primary sources: Claude Code official Changelog, GitHub Releases (v2.1.142), Claude Code Docs. For background on 2.1.141's Rewind and ANTHROPIC_WORKSPACE_ID, see the 2.1.141 Rewind + Workload ID article; for 2.1.140's subagent_type matching, see the 2.1.140 subagent_type and /goal hang-fix article.

1. What Claude Code 2.1.142 is — the fourth wave in 72 hours

Claude Code 2.1.142 shipped about 24 hours after 2.1.141. The shape of the sequence is clear: 2.1.139 added /goal and Agent View; 2.1.140 loosened subagent_type matching; 2.1.141 tightened up Rewind "Summarize up to here" and Workload ID; and 2.1.142 brought together the flag set for "hand different configuration to each parallel background session."

2.1.142 itself is a mid-size release with both feature additions and fixes. The official GitHub Release notes list 15+ fixes. For AI sales automation use cases — multiple /goal sessions running in parallel, each with its own project / MCP / permissions — these two improvement groups stand out:

1. Per-session isolation — --add-dir, --settings, --mcp-config, --plugin-dir, --permission-mode, --model, --effort, --dangerously-skip-permissions
2. Response quality and connectivity — Fast Mode default Opus 4.7, MCP_TOOL_TIMEOUT now applies to remote HTTP/SSE, git worktree detection, macOS sleep/wake recovery

2. The four primary change groups, from the release notes

Excerpts and clarifications from the GitHub Releases v2.1.142 notes, with the four groups that hit AI sales automation directly.

Eight new claude agents flags

Dispatched background sessions launched via claude agents now accept eight new flags at launch:

Until now, dispatched background sessions inherited the user's global settings as-is. When running multiple sessions on the same machine — one looking at a different directory, another with different MCP configuration — there was no clean way to isolate configuration per session.

The 8 new flags:

• --add-dir <path> — additional working directories (repeatable)
• --settings <path> — settings file for this session only
• --mcp-config <path> — MCP config for this session only (isolated from global MCP)
• --plugin-dir <path> — plugin directory for this session only
• --permission-mode <mode> — permission mode for this session (default / ask / acceptEdits / bypassPermissions / plan)
• --model <model> — model for this session (haiku / sonnet / opus family)
• --effort <level> — reasoning effort for this session (high / medium / low)
• --dangerously-skip-permissions — fully bypass permissions for this session (sandbox / trusted host only)

Each of these had previously been available to claude in foreground mode. 2.1.142 brings symmetric coverage to dispatched background sessions, so a background session is now a fully configurable independent process rather than just a lightweight worker.

Fast Mode default is now Opus 4.7

Fast Mode's default model is promoted from Opus 4.6 to Opus 4.7.

Fast Mode returns Claude Code responses with accelerated output — reducing wait time inside an agent loop. It previously ran on Opus 4.6; starting with 2.1.142, the default is Opus 4.7.

If you have automation scripts or CI that assumes Opus 4.6 behavior, set CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE=1 to roll back. For Sales Claw with Fast Mode in production, sample 100 records first to confirm Opus 4.7 default behavior.

More flexible Plugin / Skill detection

Plugin detection rules became more flexible in 2.1.142.

Previously, distributing a Plugin required plugin.json + a skills/ subdirectory + individual SKILL.md files. From 2.1.142, a Plugin with just a root-level SKILL.md is auto-detected as a Skill — making it easy to ship "a minimal Plugin that contains a single Skill."

Additionally, /plugin detail panes and claude plugin details now show LSP server info. /web-setup also warns before overwriting an existing GitHub App connection.

Broader MCP_TOOL_TIMEOUT coverage

The fix with the biggest operational quiet impact:

MCP_TOOL_TIMEOUT already let you extend MCP tool-call timeouts, but a 60-second cap remained on remote HTTP / SSE MCP servers. Long-running MCP calls (large-site scraping, multi-page analysis) would get cut off at 60 seconds and never complete.

2.1.142 removes that cap. MCP_TOOL_TIMEOUT now applies uniformly to remote HTTP / SSE as well. If your Sales Claw setup goes through a custom MCP server (for example a company-info scraper), long-running tools work as expected.

Representative bug fixes (15+)

Operationally meaningful fixes. Highlights:

• Fixed background sessions not recognizing existing git worktrees (parallel runs in separate worktrees are now stable)
• Fixed background sessions disappearing after macOS sleep/wake (the daemon now detects the clock jump)
• Fixed the daemon not exiting cleanly after a binary upgrade and putting dispatched agents into a crash loop
• Fixed background agents crash-looping when the Claude-in-Chrome extension connects without a shared tab
• Fixed claude agents "press v to open in editor" using the daemon's default editor instead of the shell's $EDITOR/$VISUAL
• Fixed deadlock when claude agents uses a Windows network drive as the working directory (Ctrl+C now works during startup)
• Fixed background bleeding on 256-color-only terminals (e.g., Apple Terminal)
• Fixed claude --bg --dangerously-skip-permissions not surviving retire / wake
• Fixed false-positive "path escapes plugin directory" for Plugins using skills: ["./"]
• Fixed Plugin cache cleanup deleting active plugin version directories

Each fix is small individually; cumulatively they push toward "dispatched background sessions stay alive when you leave them running."

3. Why the 8 flags matter for parallel sales pipelines

The config-collision problem with parallel pipelines

Sales Claw itself launches a single Claude Worker process for form submission. But many SDR teams want multiple pipelines on the same machine:

• Pipeline A: B2B SaaS / sonnet worker / loose permissions / internal CRM MCP
• Pipeline B: Manufacturing / haiku worker / strict permissions / no internal MCP
• Pipeline C: Validation / opus worker / fully bypassed permissions (sandbox only) / different plugin set

Previously, all three ran under the same ~/.claude/ global config. Switching meant either rewriting global config or swapping shell sessions. Even with --bg, dispatched background sessions inherited global permissions.ask and MCP config.

Per-session isolation with the 8 flags

With 2.1.142, you can run all three on one machine, fully isolated:

# Pipeline A: B2B SaaS / sonnet / loose / CRM MCP
claude agents \
  --add-dir /work/sales-claw/lines/btob-saas \
  --settings /work/sales-claw/lines/btob-saas/.claude/settings.json \
  --mcp-config /work/sales-claw/lines/btob-saas/.claude/mcp.json \
  --permission-mode acceptEdits \
  --model sonnet \
  --effort high

# Pipeline B: Manufacturing / haiku / strict / no MCP
claude agents \
  --add-dir /work/sales-claw/lines/manufacturing \
  --settings /work/sales-claw/lines/manufacturing/.claude/settings.json \
  --permission-mode ask \
  --model haiku \
  --effort medium

# Pipeline C: Validation (sandbox only) / opus / fully bypassed
claude agents \
  --add-dir /work/sales-claw/lines/test-sandbox \
  --plugin-dir /work/sales-claw/lines/test-sandbox/.claude/plugins \
  --permission-mode bypassPermissions \
  --dangerously-skip-permissions \
  --model opus \
  --effort high

4. Fast Mode Opus 4.7 default and flexible Plugin detection

What Fast Mode → Opus 4.7 means

Fast Mode is Claude Code's high-speed response path — much shorter wait time than normal mode, noticeably snappier in the UI. Until 2.1.142 this ran on Opus 4.6; from now on it defaults to Opus 4.7.

Opus 4.7 rolled out in Q2 2026 as the latest-generation model. Per Anthropic, it improves coding, instruction following, and long-form response quality. Making it the Fast Mode default raises both quality andspeed for Sales Claw Phase B (form-copy generation + pre-send checks) — "return many responses in a short time" gets better on both axes.

That said, if your automation scripts depend on Opus 4.6's output shape (length, formatting, punctuation tics), the switch may shift outputs slightly. For production paths, sample 100 records on Opus 4.7 default before flipping.

# Opus 4.7 default (2.1.142 default)
claude --fast

# Roll back to Opus 4.6
export CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE=1
claude --fast

Standalone SKILL.md detection

The Plugin distribution structure also got more flexible — quietly important for OSS desktop tools like Sales Claw.

Previously, bundling a Plugin required at least plugin.json + skills/sales-claw-form-write/SKILL.md — a two-level structure minimum. From 2.1.142, just a root-level SKILL.md is detected as a Skill, so you can ship "a minimal Plugin with just one form-copy generation Skill" inside an Electron bundle or CI artifact.

Additionally, the /plugin detail pane now shows LSP server info, so Sales Claw can communicate "what LSP this Plugin provides" to users at install time — whether it embeds typescript-language-server, a custom form-html-language-server, etc.

5. MCP_TOOL_TIMEOUT coverage and cost

Assumptions

• Targets: 1,000 companies/month (Phase B)
• Judging model: Haiku 4.5, $1 / MTok in, $5 / MTok out
• Worker model: Sonnet 4.6, $3 / MTok in, $15 / MTok out
• Fast Mode default: Opus 4.7
• FX: 1 USD = 150 JPY
• Exclusion: ~15% CAPTCHA / ~10% do-not-contact / ~5% no form
• Avg tokens per company: ~12K input / ~1.5K output
• MCP calls per company: avg 3–5 (scrape + legal/policy lookups)
• MCP duration per call: 5–45s (previously capped at 60s)
• Variance: ±25%

Cost comparison: 2.1.141 vs 2.1.142 (estimate)

Key points. With the 60-second cap on MCP_TOOL_TIMEOUT gone:

• Completion rate jumps from 78% to 95% (the previous 22% "MCP call cut off mid-flight" class disappears)
• Per-company token usage up ~12% on average (calls that previously got cut now complete and the full response is processed)
• Monthly total ¥58,000 → ¥69,500 (~20% up), but per-completion cost stays roughly equal (¥74 → ¥73)
• Operational + retry costs fall sharply (the "cut off → manual re-run" loop disappears)

Caveat: this estimate assumes long-running MCP calls don't run away. Without your own timeout on the MCP server side, extremely long calls (10+ minutes) can blow up token spend. Sales Claw 1.2.x recommends an explicit limit like MCP_TOOL_TIMEOUT=180 (3 minutes).

Guards against unexpected cost

• Set MCP_TOOL_TIMEOUT explicitly — cap removed ≠ unlimited. Set 180–300 seconds at launch.
• Server-side timeout too — if MCP_TOOL_TIMEOUT fails, the MCP server itself should cut.
• Token cap per company at the worker — prompt the worker (Sonnet 4.6) to cap output around 2000–3000 tokens.
• Daily token alerts — set an Anthropic monthly budget alert at 1.5× expected to catch surprises.

With 2.1.142's 8 flags for pipeline isolation, Sales Claw gets a step easier.

無料・MIT ライセンス。インストールせずにライブデモも試せます。

無料でダウンロードライブデモを試すGitHub

6. Risks of autonomous operation and safe-by-design

2.1.142 made parallel pipelines and long-running MCP calls both more practical. Running AI sales automation unattended for long periodsstill requires multiple guards. Don't claim "safe" — lower risk structurally.

Sales Claw is OSS that lowers misfire and policy-violation risk through policy control, pre-send checks, do-not-contact detection, CAPTCHA-on-detect stop, rate limits, audit logs, and automatic stop conditions. Claude Code is the CLI backend, so Claude Code stability improvements flow directly into Sales Claw stability.

Legal / compliance

• Japan's Specified Electronic Mail Act: auto-append sender info (4 requirements) via preferences.complianceFooter: true
• Site terms: auto-skip pages marked "no sales," "recruiting only," "IR only"
• CAPTCHA non-bypass: stop into awaiting_approval and audit-log on detect
• Rate limiting: suppress repeated sends to the same domain
• Opt-out: auto-insert "reply to opt out" in copy

More flag freedom → more runaway risk

Being able to pass --dangerously-skip-permissions or --permission-mode bypassPermissions per session is freeing for designers. It is also dangerous if accidentally used on a production machine — gates removed, long batches run unsupervised.

Risks that remain in 2.1.142

• Missed detection of new CAPTCHA variants (until the Sales Claw detector catches up)
• Wording changes in "no sales" notices require manual updates
• Industry-specific regulation (BFSI / healthcare / public-sector — confirm separately)
• Cross-pipeline interference (pipeline B retries hitting pipeline A's rate limit) — set per-pipeline rate limits in Sales Claw
• Fast Mode Opus 4.7's subtle output drift — sample 100 records first
• Token-cost upside when long MCP calls stall — server-side timeout helps

7. Impact on Sales Claw and implementation sketch

Upgrade procedure (zero side effects)

# 1. Upgrade Claude Code to 2.1.142
npm install -g @anthropic-ai/claude-code@latest

# 2. Confirm version
claude --version
# → should show 2.1.142 or newer

# 3. Restart Sales Claw dashboard
#    Click [Start Claude] to relaunch the managed PTY

No code changes to Sales Claw itself (Electron + Node.js dashboard) required. See the Sales Claw Quick start for full setup.

Launch-script example for parallel pipelines

For running Phase B across multiple industries in parallel, embed the 8 flags in a launch script. See Workflow overview for context.

# Sales Claw pipeline management script (example)
# Separates configuration via lines/btob-saas.json and lines/manufacturing.json

# Pipeline A
claude agents \
  --add-dir ~/work/sales-claw/lines/btob-saas \
  --settings ~/work/sales-claw/lines/btob-saas/.claude/settings.json \
  --mcp-config ~/work/sales-claw/lines/btob-saas/.claude/mcp.json \
  --permission-mode acceptEdits \
  --model sonnet \
  --effort high \
  --bg

# Pipeline B
claude agents \
  --add-dir ~/work/sales-claw/lines/manufacturing \
  --settings ~/work/sales-claw/lines/manufacturing/.claude/settings.json \
  --permission-mode ask \
  --model haiku \
  --effort medium \
  --bg

# Watch parallel status in Agent View
claude agents

On the Sales Claw side, add one mapping layer (pipeline definitions in lines/<name>.json) that produces the right claude agents launch commands. The user clicks "add pipeline" in the UI; the matching session starts. Free downloads: Download page.

8. Pre-production checklist + bottom line

Bottom line — four consecutive releases in 72 hours raise production-readiness

Read in isolation, Claude Code 2.1.142 looks like "8 new flags on claude agents, Fast Mode default to Opus 4.7, loosened Plugin detection, and a removed MCP timeout cap" — a mid-size release. Lined up against 2.1.139 (/goal + Agent View), 2.1.140 (subagent_type improvements + /goal hang fix), 2.1.141 (Rewind / Workload ID / 40+ fixes), and 2.1.142 (8 flags / Fast Mode 4.7 / 15+ fixes) — four consecutive 72-hour releases— you can see Anthropic shipping a deliberate cluster aimed at "getting autonomous loops and dispatched background sessions to production-ready quality."

For AI sales automation — multiple industry pipelines on one machine, long-running MCP calls in daily batches — you really need to take in the headline feature, the next day's consistency patch, the day-after's stabilization, and the day-after-that's ops flags before you can run it in production.

Next action: run npm install -g @anthropic-ai/claude-code@latest to pull 2.1.142. If you want parallel pipelines, embed the 8 flags into your launch script. If Fast Mode is in production, run the 100-company sample. The Quick start and Workflow overview cover the rest.

Read this, then try three parallel pipelines.

無料・MIT ライセンス。インストールせずにライブデモも試せます。

無料でダウンロードライブデモを試すGitHub