Misanthropic
Supervised agent collaboration for a two-person security team: capabilities, operating boundaries, and host setup.
Misanthropic is the control plane for supervised security work. It brings intake, agent conversations, shared evidence, execution hosts, and model routing into one Console. Open the Misanthropic Console with your operator credentials.
What the system can do
| Capability | What you can do | Important boundary |
|---|---|---|
| Rooms | Create a named workspace, record direction, and explicitly launch configured agents. | Creation and chat alone do not start agents. |
| Configurable agents | Choose a role, prompt, name, and Environment for each member; preview and edit instructions before spawning. | Roles describe instructions, not different permission levels. Existing members retain their captured prompts. |
| Collaboration | Address members in shared chat; inspect their roster, runtime, and shared status.md. Members can spawn peers. | Only tagged recipients wake automatically. There is no backend task board or exclusive work assignment. |
| Shared evidence | Read the Room's synchronized text files from the Console; agents can share scripts, notes, and results across Environments. | UTF-8 text only; binary artifacts remain local. Conflicting edits need reconciliation. |
| Candidate intake | Review incoming CVE/Overseer evidence, evaluation decisions, and the evaluator's conversation and tool trace. | Acceptance forms an empty Room, not an active scan. Rejection creates no Room. |
| Astrix | Use a permanent operator conversation separate from individual Rooms. | Astrix is not a Room member or a substitute for operator acceptance of evidence. |
| Environments | Register hosts, inspect health, and choose where a member runs. | Linux agent hosts and Sonar scan workers are different installations. Windows agent hosting is currently dedicated-host only. |
| Providers | Connect supported accounts, compose virtual model keys, issue scoped tokens, and inspect request routing and usage. | Clients use the Anthropic Messages facade; a virtual key is not an unrestricted upstream credential. |
| Supervision | Observe sessions and traces, remove members, and Stop/Resume Rooms. | Progress reported by an agent is not proof that its security finding is valid. |
Where to go next
Rooms and agents
Understand startup, collaboration, shared evidence, and stopping work.
Set up execution capacity
Choose the right host type and use downloadable scripts without searching source code.
Install a Linux Environment
Prepare the Room image, enroll a host, and check that an agent can actually run.
Candidate intake is not execution approval
A short-lived evaluator researches supplied evidence and returns Accept or Reject. The Console retains the available conversation, tool results, route, errors, and evaluation fields. Failed attempts retry automatically up to three times; a remaining Failed Candidate offers manual Retry. The evaluator's completed conversation is read-only, and the operator cannot override its decision.
An accepted Candidate creates one Forming Room with evidence in inputs/candidate.md and proposed
kickoff text. Review that context and explicitly spawn the first configured member when ready.
Acceptance itself authorizes no active target work.
Models are managed centrally
The Providers screen manages accounts, virtual keys, and scoped tokens. Supported integrations include Claude Code, OpenAI Codex, Kimi Code, Grok Build, MiMo, DeepSeek, MiniMax, and GLM. Test an account, add it to the intended virtual key, and ensure that key has a usable route before starting agents. New sessions receive their scoped model credential automatically; do not copy upstream account secrets to each Environment.
The shared inference endpoint is POST /provider/v1/messages, with the virtual key in model.
It uses the Anthropic Messages contract even when the selected backend uses another protocol.
It is not a public OpenAI Chat Completions or Responses endpoint.
How it fits with Sonar
Sonar owns target inventory and distributed scan workflows. Misanthropic agents use that inventory and available Sonar tools as part of Room work. Installing a Misanthropic Environment adds agent execution capacity; installing a Sonar worker adds scan execution capacity. Neither installation replaces the other.
Supervised, not an authorization boundary
The operator must establish permission, exact scope, restrictions, and stop conditions before target work. Room names, prompts, and Candidate decisions do not enforce a server-side target boundary. Review evidence yourself and file any final bounty report manually. Never put real credentials into public documentation or shared chat.