Right now, with zero provider keys configured, /claude-council:ask exits
immediately with Error: No providers configured. That's correct behaviour
for the cross-vendor pitch, but it leaves Claude Code users who only have
their Claude subscription (no Gemini / OpenAI / xAI / Perplexity keys) unable
to use the plugin at all.
Suggestion
Add an opt-in "local council" mode that spawns N Claude subagents in
parallel, each pinned to a different role from config/roles.json (security
auditor, devil's advocate, simplicity champion, scalability architect, etc.)
and synthesises their responses the same way the cross-vendor flow does.
The diversity is weaker than cross-model — same priors, same training data —
but role-prompting still produces meaningfully different angles, especially
for "what am I missing" / "poke holes in this" use cases. And it lets users
try the plugin before deciding whether to wire up paid API keys.
Rough shape
- New flag, e.g.
--local or auto-fallback when no provider keys exist
- Reuse the existing role presets (
balanced, security-focused,
architecture, review)
- Reuse
skills/deep-execution plumbing for parallel subagent dispatch
- Synthesis step is unchanged
Why it's worth doing
- Lowers the activation cost — install and try, configure later
- Makes the
roles.json work pull double duty
- Plays well with
/claude-council:status showing 0/4 vendors but
"local council available"
Right now, with zero provider keys configured,
/claude-council:askexitsimmediately with
Error: No providers configured. That's correct behaviourfor the cross-vendor pitch, but it leaves Claude Code users who only have
their Claude subscription (no Gemini / OpenAI / xAI / Perplexity keys) unable
to use the plugin at all.
Suggestion
Add an opt-in "local council" mode that spawns N Claude subagents in
parallel, each pinned to a different role from
config/roles.json(securityauditor, devil's advocate, simplicity champion, scalability architect, etc.)
and synthesises their responses the same way the cross-vendor flow does.
The diversity is weaker than cross-model — same priors, same training data —
but role-prompting still produces meaningfully different angles, especially
for "what am I missing" / "poke holes in this" use cases. And it lets users
try the plugin before deciding whether to wire up paid API keys.
Rough shape
--localor auto-fallback when no provider keys existbalanced,security-focused,architecture,review)skills/deep-executionplumbing for parallel subagent dispatchWhy it's worth doing
roles.jsonwork pull double duty/claude-council:statusshowing 0/4 vendors but"local council available"