Slack ↔ Claude Code remote control: architectures, recipes, and traps

Decision. If you can live with a hosted experience, use Claude Code in Slack [5] [23] — Anthropic ships it, sessions run on claude.ai/code, no infra to maintain. If you need self-hosted (private repos, custom MCP, on-prem secrets), pick chenhg5/cc-connect ⭐ 6.4k (Apr 2026) [12] for the Socket-Mode + multi-platform path, or wire anthropics/claude-code-action ⭐ 7.3k [28] behind a thin Slack→GitHub-Issue shim like AnandChowdhary’s bot [8] for the indirect / GitHub-Actions path. Build-it-yourself only if the use case is genuinely off-pattern — the Slack 3-second ack rule, prompt-injection from arbitrary chat input, and runaway-cost from open @-mentions [38] are all solved problems but easy to get wrong.

The four architectures, side-by-side

The split is real and load-bearing: Slack transport (Socket Mode vs Events API) and execution location (managed cloud, bot host, GHA runner, local daemon) compose independently. Any pair is achievable; pick the pair that matches the threat model and ops tolerance.

The Slack contract you cannot avoid

Three Slack-side rules force the architecture more than any Claude-side concern:

The Bolt SDKs codify the workaround as lazy listeners — await ack() first, push to queue, do the work in a separate function [33]; ClaudeFluent’s Slack-bot recipe names the same constraint as the single most critical timing rule [34]. Even this leaks: Bolt-JS issue #1548 catalogs operation_timeout errors when ack() runs but the surrounding handler is slow under Lambda cold-start [32]. The canonical pattern is ack synchronously → SQS / Cloudflare Queue / Redis stream → worker calls claude -p [9]. For self-hosted Events API without a static IP, Cloudflare Tunnel is the 2026 default over ngrok because the hostname survives restarts and avoids re-registering the Slack URL [10].

Claude Code’s headless surface — what makes it bot-friendly

The CLI’s non-interactive surface is everything a Slack bot needs and was explicitly designed for it:

The Python claude-agent-sdk skips the spawn altogether: ClaudeSDKClient is an async context manager that auto-tracks the session, and CanUseTool: Callable[..., Awaitable[PermissionResult]] is the in-process equivalent of a permission prompt — it returns PermissionResultAllow (with optional updated_input) or PermissionResultDeny [22]. TypeScript’s query() exposes equivalent continue / resume / forkSession options on its options object [21].

⚠ Session-resume gotcha: transcripts are keyed on <encoded-cwd> — if your worker resumes from a different working directory than the original run, you silently get a fresh session [21]. Pin the cwd or ship the JSONL.

Existing projects, ranked

Anthropic’s first-party paths (Apr 2026)

Bottom line: Anthropic ships the front door (Claude Code in Slack) and the backend primitive (claude-code-action), but no first-class Slack-trigger for self-hosted Claude Code. The community fills that gap.

Counter-argument: maybe don’t ship this

A Slack ↔ Claude Code bot inherits the trust, reach, and tone of the user account it runs as [35], and any workspace member who can @-mention can spend money [38]. If your use case is “trigger one repo’s tests from chat occasionally”, a /claude-fix slash command shelling claude-code-action via repository_dispatch is much smaller surface area than a long-running bot host. Build the bot only when streaming progress, multi-turn threading, or thread-pinned context actually pay for the operational debt.

Reference recipe: ack → queue → CLI → reply

A minimal self-hosted pipeline that satisfies every Slack constraint:

┌──────────────┐  Events API  ┌───────────────┐  enqueue  ┌──────────┐
│ Slack channel│─────HMAC────▶│  Slack worker │──────────▶│  Queue   │
└──────────────┘  (Cf Worker, └───────────────┘ (SQS, Cf  └──────────┘
       ▲          Lambda, etc.)│  ack ≤ 3 s    │ Queue, Redis)   │
       │                       │  add 👀 react │                 ▼
       │                       └───────────────┘          ┌──────────────┐
       │                                                   │   Worker     │
       │            chat.postMessage thread_ts             │ claude -p \  │
       └───────────────────────────────────────────────────│   --output-  │
                                                           │   format     │
                                                           │   stream-json│
                                                           └──────────────┘

Key choices each box:

1. Slack worker — Cloudflare Workers (free tier, persistent hostname [10]) or AWS Lambda. Verify HMAC-SHA256 [3], reject ts > 5 min old, return 200, write payload to queue. Add an eyes reaction synchronously so users see the bot accepted the request even before the queue picks it up — AnandChowdhary’s bot does this [8].
2. Queue — for ≤ 30 min runs, Cloudflare Queue with a 10 s poller works [8]; for longer, a stateful Redis stream or git-as-queue (41fred’s pattern, repo tasks/ directory polled every 5 s [41]).
3. Worker — clones the repo into an ephemeral cwd, runs claude -p --output-format stream-json --verbose --resume <session-mapped-from-thread_ts> --max-budget-usd 5 --allowedTools 'Read,Edit,Bash(git *)' "$prompt". Pipe stdout through a NDJSON parser → chat.postMessage updates on the original thread_ts. Cap --max-turns defensively.
4. Auto-commit & push — the 41fred daemon commits and pushes after each task automatically [41]; MattKilmer/claude-autofix-bot opens a PR with simple-git + the GitHub API and links it back to the Slack thread [42].
5. Inject Slack context for free — takafu/slack-claude-bot’s clever shortcut: pass SLACK_BOT_TOKEN, CHANNEL, THREAD_TS as env into the Claude Code child process so Claude itself can use Bash+curl to post progress and reactions, no SDK plumbing needed [15].

For the SRE / production shape, David Calvert’s “Hive Ix” is <2,000 lines of Go on GCP/Kubernetes via Helm, calling exec.Command("claude", "-p", prompt) against Sonnet 4.5 on Vertex AI, with Atlassian + Datadog MCP servers attached for incident-response context [43].

Pitfalls — and what they cost

When to pick what

→ Hosted, GitHub-connected work → Claude Code in Slack [5]. Done.

→ Self-hosted, simple, 1 repo, behind firewall → chenhg5/cc-connect [12] (Socket Mode, no public IP).

→ GitHub-Actions–centric team, want PRs naturally → AnandChowdhary’s pattern [8]: Slack → GH Issue → claude-code-action → comments mirrored back. The 30-min poll cap is the hard limit.

→ Laptop is the runner, public Slack endpoint → 41fred split-worker pattern [41], git-as-queue.

→ Long-running runs that need human-in-the-loop → tomeraitz/claude-slack-bridge inverted MCP [44]: Claude itself blocks on a Slack reply mid-run.

→ Production SRE bot at scale → Calvert’s Helm/Vertex pattern [43] — minimal Go, MCP-attached observability.

→ Anything else → Routines API [26] for triggered runs; build a thin repository_dispatch bridge to claude-code-action for the chat-driven side.