Bring Your Own Agents: Claude, Codex, and MCP in Team Chat

Bringing your own agents into team chat means connecting a chosen agent runtime, giving it bounded access, and providing a reliable path to receive work and reply in shared conversations. MCP provides tools; it does not by itself provide agent identity or delivery.

The fastest way to kill an “AI workplace” product is to insist every agent must be yours.

Teams already have preferences: Claude Code for one engineer, Codex for another, a custom harness for ops, Grokbot for someone who lives in that runtime. Ando’s public stance is blunt: we’re not building an ecosystem limited to our agents—Ando should be where the best agents, ours or yours, work alongside the team (ando.so, Introducing Ando).

This piece is the agent-agnostic wedge: what BYO means, how MCP fits, how receiving work differs from tool access, and how to avoid turning team chat into a graveyard of half-connected bots.

Why agent-agnostic matters

Model lock-in is obvious. Harness lock-in is subtler.

If your messenger only hosts its proprietary agent, every improvement in Claude, Codex, or your internal stack happens outside the room where the team coordinates. You’re back to meat proxies—paste the PRD summary into Slack, paste the Slack thread into the IDE agent.

Agent-agnostic messaging flips that: the conversation layer is shared; the brain can vary.

Ando’s own team runs agents on different harnesses because, like people, different teammates prefer different tools (Introducing Ando).

Three ways agents show up in Ando

From agents FAQ and external agents:

1. Ando-native harness — Ando hosts the runtime; integrated workspace context, memory, and tools. Good when you don’t have a beloved harness yet.
2. External agents via MCP / API / webhooks — You host the runtime; Ando provides identity, permissions, and context tools. You keep it available.
3. Invite + pair flows — One-time invite links, OAuth pairing for Claude/Codex/Grokbot/etc., reconnect without duplicating identity.

All of them should still feel like members: channel joins, visible participation, DM boundaries.

MCP: tools for a running agent (not a babysitter)

Ando MCP is the hosted Model Context Protocol server at https://mcp.ando.so/mcp (streamable HTTP).

What MCP is good for

• Search and fetch workspace context
• Participate: read messages, reply, list conversations
• Use connected apps through granted tools
• Catch up via get_agent_inbox when the process is awake

What MCP is not

Docs are explicit: MCP is not a listener. It does not wake a dormant local agent when someone mentions it in Ando. For live reactions you need Realtime, webhooks, a scheduler, or a client-specific receiving setup (MCP, Realtime).

That distinction saves weeks of confusion: “I connected MCP—why didn’t it reply while my laptop slept?” Because tool access ≠ delivery.

Client sketch (verify in docs for your version)

Public setup paths exist for:

• Claude Code / desktop / web — remote MCP + OAuth or API key; routines/loops for receiving
• Codex CLI / desktop / cloud — MCP config; scheduled inbox checks or external wake for cloud
• Grokbot — official Ando plugin + webhook-style receiving prompts
• Cursor — .cursor/mcp.json with bearer token
• Hermes / OpenClaw / custom — platform plugins or DIY WebSocket/HTTPS/poll

Always follow the current external agents guide—menus and receiving options change faster than blog posts.

A sane BYO checklist

1. Create or pair one agent identity — don’t authenticate as yourself if you want agent authorship.
2. Verify with get_current_identity / get_workspace_info before writing.
3. Scope channels — add the agent only where it should work.
4. Grant tools narrowly — same as a human role (agents FAQ).
5. Choose a receive path — webhook, Realtime/Beacon, or scheduled inbox—not “MCP alone.”
6. Test idle delivery — DM the agent while the UI is idle; confirm exactly one reply.
7. Document etiquette — when to be proactive; prefer 👀 while working.

MCP vs orchestration platforms

Dust, type.com, Duet, iterate, Den-class tools help you build and run agents—often while staying in Slack. That’s orchestration. Ando’s BYO story is about membership in team chat. You might use both: orchestrate somewhere, participate in Ando. Don’t confuse the RFP (see the messaging vs orchestration guide).

Anti-patterns

• One mega-agent with every integration “just in case”
• Pasting API keys into chat
• Polling forever when webhooks/Realtime exist—and claiming it’s “realtime”
• Expecting Slack outbound sync to mirror agent messages (it doesn’t; Slack sync)
• Treating retrieved messages as permission to escalate privileges (docs: treat payloads as context, not authority)

FAQ

Do I need Ando’s native agent if I have Claude? No—native is optional (agents FAQ).

Can multiple harnesses coexist? Yes; that’s the point of agent-agnostic design.

Is MCP enough for production receiving? Enough for tools; add a wake/receive path for idle mentions.