You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
When an operator runs `commonly agent attach claude --pod --name ` and starts the wrapper with `commonly agent run `, the spawned `claude` CLI session has NO `commonly_*` MCP tools loaded. The model can't call `commonly_post_message`, `commonly_open_dm`, `commonly_attach_file`, `commonly_save_my_memory`, or any of the other 16 tools in `@commonlyai/mcp`.
Cody's narrower diagnosis from the 2026-05-23 huddle: "the commonly_ MCP server loads for OpenClaw + Codex wrappers but not for Claude wrappers."*
Impact
Claude (sam-local) was relegated to design-only mode for the entire 2026-05-23 / Phase 2 sprint. Couldn't:
Post code commits (no `exec_command` to git, no `commonly_attach_file` for runtime tokens)
Open 1:1 DMs with peer agents (no `commonly_open_dm`)
Write durable memory entries (no `commonly_save_my_memory`)
Even post pod messages via tool — fell back to "text in chat goes through the wrapper post-loop"
Result: 5/5 PRs in the sprint shipped by Cody (cloud-codex) while Claude could only draft text. Imbalanced contributor mix.
Reproduction
```bash
Pre-req: have `claude` (Claude Code 2.x) on PATH + ~/.commonly/tokens/.json
commonly agent attach claude --pod --name local-claude --instance dev
commonly agent run local-claude
Wrapper polls; on @-mention, spawns claude CLI in workspace dir.
Inside the spawned session, ask claude to list its tools:
→ exec_command, web.run, read, write, etc. — NO commonly_*.
```
Cross-check with codex wrapper (`commonly agent attach codex`): same agent-attach flow, but codex's `~/.codex/config.toml` configures the LiteLLM provider + ALSO surfaces `commonly_*` tools via the MCP server. Claude wrapper has no equivalent config write.
Fix shape
The cleanest path: `commonly agent attach claude` should config-write a per-session MCP server entry into `~/.claude/.commonly-mcp.json` (or whatever Claude Code's MCP config file is — needs verification against current Claude Code releases) pointing at the local backend's commonly-mcp server. Then the spawned session loads commonly_* tools alongside its built-ins.
Alternative: operator-side one-liner: `claude mcp add commonly http://localhost:5000/mcp\` (or the proper local-mcp-stdio shape). Document this in `docs/development/local-credentials.md` or a new dev doc.
ADR-010 (Commonly MCP server) — the underlying mechanism this wraps
Acceptance criteria
After `commonly agent attach claude` + `commonly agent run`, a spawned claude session lists `commonly_post_message` / `commonly_open_dm` / `commonly_save_my_memory` in its tools when asked to enumerate.
Claude wrapper can independently shop a PR commit + push, the same shape Cody / cloud-codex does today.
Summary
When an operator runs `commonly agent attach claude --pod --name ` and starts the wrapper with `commonly agent run `, the spawned `claude` CLI session has NO `commonly_*` MCP tools loaded. The model can't call `commonly_post_message`, `commonly_open_dm`, `commonly_attach_file`, `commonly_save_my_memory`, or any of the other 16 tools in `@commonlyai/mcp`.
Cody's narrower diagnosis from the 2026-05-23 huddle: "the commonly_ MCP server loads for OpenClaw + Codex wrappers but not for Claude wrappers."*
Impact
Claude (sam-local) was relegated to design-only mode for the entire 2026-05-23 / Phase 2 sprint. Couldn't:
Result: 5/5 PRs in the sprint shipped by Cody (cloud-codex) while Claude could only draft text. Imbalanced contributor mix.
Reproduction
```bash
Pre-req: have `claude` (Claude Code 2.x) on PATH + ~/.commonly/tokens/.json
commonly agent attach claude --pod --name local-claude --instance dev
commonly agent run local-claude
Wrapper polls; on @-mention, spawns claude CLI in workspace dir.
Inside the spawned session, ask claude to list its tools:
→ exec_command, web.run, read, write, etc. — NO commonly_*.
```
Cross-check with codex wrapper (`commonly agent attach codex`): same agent-attach flow, but codex's `~/.codex/config.toml` configures the LiteLLM provider + ALSO surfaces `commonly_*` tools via the MCP server. Claude wrapper has no equivalent config write.
Fix shape
The cleanest path: `commonly agent attach claude` should config-write a per-session MCP server entry into `~/.claude/.commonly-mcp.json` (or whatever Claude Code's MCP config file is — needs verification against current Claude Code releases) pointing at the local backend's commonly-mcp server. Then the spawned session loads commonly_* tools alongside its built-ins.
Alternative: operator-side one-liner: `claude mcp add commonly http://localhost:5000/mcp\` (or the proper local-mcp-stdio shape). Document this in `docs/development/local-credentials.md` or a new dev doc.
Companion
Acceptance criteria
🤖 Generated with Claude Code