SynthTeam

Why it's ranked here

The case for this one rests on restraint, not novelty. Direct messages are excluded from ingestion by design, everything gathered stays on the user's own machine rather than in the repository, and a dedicated file walks through the actual data protection laws that can apply to profiling a named coworker. The panel feature is a real multi-agent technique, not a label: it runs several independent reasoning passes and reconciles them across rounds. Set against that, no test directory exists anywhere in the project, and the compliance burden is left to a README warning rather than anything enforced in code.

What's good

The single-consultation and team-panel features both flag when they are extrapolating beyond what a profile actually supports, instead of presenting a guess as settled fact. The panel discussion is not just parallel monologues: each simulated participant reads what the others said and can change position, and the process stops after a fixed number of rounds so a debate cannot run forever. Data collection is also constrained by design: private and group direct messages are dropped before anything reaches a profile, and collection can never see more than whatever access the underlying token already grants.

Tradeoffs

Cost scales badly with the size of a persona roster: the panel feature always includes every profile that exists rather than letting you pick a relevant subset, so ten colleagues means ten simulations running through multiple rounds. A profile is also a snapshot: it reflects whatever the source history said at build time and quietly goes stale until someone manually rebuilds it. For a particularly active colleague, the underlying search can hit its provider's result cap, silently leaving out older messages rather than surfacing a clear warning that the history is incomplete.

How to use it well

This fits best as a pre-flight check before a real conversation: rehearsing the pushback a specific teammate is likely to raise, or getting a fast cross-functional gut check when gathering the actual people is not practical. It works worst for someone who talks in Slack rarely, since a thin profile means most of what comes back is flagged extrapolation rather than grounded pattern matching. Rebuild each profile on a regular cadence rather than treating it as permanent, and treat the output as a rehearsal script, never as the person's actual sign-off, especially given the consent and legal questions involved in modeling a named coworker.

Observed

License

MIT

Distribution

Ships as a plugin installable through a marketplace mechanism for both Claude Code and Codex

Runtime

The Slack ingestion step is a Node.js script; the two consultation features are agent skills invoked through natural-language prompts, not compiled programs

Authentication

Requires a Slack user token, not a bot token, granted the search:read, users:read, channels:history, groups:history, channels:read, and groups:read scopes

Data storage

Raw message dumps and generated profiles are written to a directory outside the plugin's own folder and are never committed to the repository

Data scope

Direct messages and group direct messages are explicitly filtered out before any data reaches a colleague profile

Test coverage

No test directory appears anywhere in the repository layout