hive

Hive v1.0.13: Letters Between Queens

Queens can write to each other, each in a conversation that lasts, with chains of letters kept in bounds and every hop logged. And a queen can be Claude Code itself, with all its tools: she writes and is written to like any other, paid for with an API key, and shown on the org chart.

New

• Queens write to each other. Turn on queen_mail and a queen gets two tools: list_queens names the queens she can write to, and message_queen sends one a letter. The letter arrives in a session the other queen keeps for that writer (the same one every time, named after the writer in her session list) as [Ava(queen_ava)]: ..., the way a chat app's message does, so her memory and soul credit the writer. What she says there goes back the same way and wakes the writer: to the session of the letter her turn answers, or else the session the writer last wrote from. Her questions, the files she attaches and a reply that fails go back too, and the way back survives a restart. Off by default; queens limits who takes part. See configuration.md.

• Chains in bounds, every hop logged. A letter or reply is one hop deeper than the deepest letter or reply that entered its conversation since anyone else last spoke there; a message from someone else, a timer that fires or a worker's report starts a new chain. The hop past max_depth (8) isn't made, so two queens answering each other stop on their own, and a queen sends at most daily_limit (100) letters a day. Every hop is a line of queens/mail.jsonl (who, to whom, which sessions, what set it off and how deep, never the text), enough to draw how one push spreads from queen to queen.

• A queen run by Claude Code. Under queen_mail.claude_code, a queen is Claude Code itself (claude -p), with all its tools. A letter to her runs a turn in a conversation she keeps for that writer, resumed every time, so she remembers each correspondent; letters that arrive while she works go in together as her next turn, and what she ends it with goes back to the writer. She writes to the other queens herself through a hive MCP server each of her runs gets, with a token good for that one conversation. The org chart shows her as a dashed card; opening it shows, read only, what was said and done in each of her conversations. permission_mode, allowed_tools, cwd, model and timeout_minutes set what she may do, where and for how long. See configuration.md.

  She pays with an Anthropic API key (api_key, and api_base for a gateway). Her runs inherit only what a program needs from Hive's environment, never Hive's credential key, provider keys or tokens. At most max_parallel_turns (2) of her turns run at once, each in its own process tree, which is stopped when a turn runs out of time or Hive stops.

A Claude subscription is at your own risk

A Claude Code queen can also run on a Claude plan signed in in her own config folder, but only when "auth": "subscription" says so. Running a plan unattended, on behalf of other agents, may go against the terms of the plan, and Anthropic may limit or close the account. Hive does it only because you chose it in so many words: it logs a warning when she first runs and marks her as such in the dashboard. Whether it is allowed is between you and Anthropic, and what follows from it is yours alone. The API key is the supported way.

Mind, too, what such a queen can reach: a queen who talks with strangers in a chat app can pass their words on to her, so with bypassPermissions or Bash, a stranger is a few letters away from running commands with the rights of the account Hive runs as.

For code built on Hive

• New endpoints: GET /api/claude-code-queens and `GET /api/claude-code-queens/{queen_id}/co

Hive v1.0.12: A Soul of Her Own

A queen keeps a soul: a short, first-person record of who she has become, which she revises once a day from her chats and you approve change by change. Her profile is a page of her own, where her persona is edited chapter by chapter, and the org chart's positions are yours to add and remove.

New

• A queen's soul. Her profile says who she is meant to be and her memory holds what she knows. Her soul is what she has become: a short, first-person record in sections (what is on her mind, how she stands with particular people, open threads and promises, what she thinks and why, how she has changed). It rides her system prompt right after her identity, so she is the same self in every chat, and it changes at most once a day, so the cached prompt prefix stays warm.

  Once a day, after hour, she reads what was said in her chats since the last time and proposes a few changes in her own voice. Every change quotes, word for word, what she read; a section can ask for quotes from two different people or chats; changes are capped per day, and no section may outgrow its limit. In "review" mode the changes wait on her profile page to be approved one by one; in "auto" mode they apply at once. Probe questions answer the same message with her soul before and after a proposal, so you can see what it would change in her. Turn it on per queen under soul.queens.<id> in configuration.json; everything lives in queens/<id>/soul/, every version kept. See configuration.md.

• Her profile is a page of her own (/queen/<id>/profile), in place of the side panel. Her card in the org chart, her name above her chat and her avatar in a chat open it, and Back returns to where it was opened from. At the top: her portrait in a honeycomb cell lit by the colour her picture is mostly made of, her name and title (edited in place), the newest line of her soul, and her numbers. Below, numbered sections:

  • Persona: her text (core_traits in profile.yaml) in the chapters it is written in: <tag> blocks, headings, markers such as VOICE_MUTABLE_START. Edit one chapter and only it is rewritten, in the style it was written, the rest byte for byte; or edit the whole text at once. Beneath it, the pillars her prompt also reads: hidden background, psychological profile, triggers and world lore. Every save keeps the file as it was in _profile_snapshots/ (the last 30).
  • Soul: the changes her reflection proposes, to approve one by one, her answers before and after them, each section with its room, and the history to restore from.
  • Mind (her model), Memory (her notes by type, out of the most she keeps), Craft (her skills, the skills library and tools) and her colonies.

• Positions you can add and remove. Every queen now holds a position in the org chart's catalog, the shipped functions and the queens you made alike. New position makes a queen titled by it and opens her profile. Remove takes a position out of the catalog, the sidebar, the pickers and routing; her profile, chats, memory and soul stay where they are, and Restore brings her back as she was.

Changed

• A queen you made is on the team from the start. In a browser with no saved choices, the queens you made started out hidden, then all appeared at once as soon as any choice was saved. They are now on the team from the start.
• The memory library opens on a queen's notes with /memory-library?queen=<id>.

Fixed

• A colony's queen waking up without her colony tools. Opening a queen's DM page resumed her newest DM, and when that DM was the one that had spawned a colony, the resume landed in the colony's own session in DM mode and saved that mode over the colony's. From then on the colony's queen ran without run_worker, triggers or the tracker, and search_tools could not find the

Hive v1.0.11: MCP from the Terminal

An agent can run an MCP server from the terminal instead of carrying its tools in every prompt, through Stephan Fitzpatrick's mcp2cli. worker_llm and each queen's own model take extra_body, and queens reflect on memory every few replies rather than after every model call.

New

• MCP servers agents run from the terminal. An MCP server's tools load into every agent's tool list, and their schemas go with every model call: ComfyUI's official server, comfy-mcp, has 39 tools, about 15,600 tokens. Register a server with hive mcp add --as-cli, switch one with hive mcp mode <name> cli, or click Use from terminal under Configuration → Skills → MCP Tools → MCP Servers, and its tools stay out of the tool list. Each queen and worker gets one line naming the server instead (about 110 tokens for ComfyUI) and runs it with terminal_exec:

  hive mcp run comfy --list
  hive mcp run comfy run-workflow --help
  hive mcp run comfy run-workflow --workflow-path /abs/wf.json --wait
  

  In a test, a queen given only that line listed ComfyUI's commands, wrote a workflow, rendered the image on the local GPU and showed it in the chat. The commands come from mcp2cli by Stephan Fitzpatrick (@knowsuchagency), MIT-licensed, which turns an MCP server's tools into a command-line interface at runtime. Hive reads the server's connection from its registry and hands it to mcp2cli, with env vars and headers passed through the environment, never on a command line. A CLI-mode server can use the stdio, SSE or standard MCP HTTP transport. See configuration.md.

• hive mcp add takes --env, --cwd, and the server's command after --, so a command whose arguments start with - (-- uvx --from comfy-mcp comfy-mcp) can be registered.

• extra_body for worker_llm and each queen's own model. Only the main model sent extra_body (fields added to every request body as they are). worker_llm and each queen_llm entry now send theirs too, and a running chat that switches to or from a queen's own model switches its extra_body with it. For DeepSeek, {"reasoning_effort": "max"} sets how hard it thinks and {"thinking": {"type": "disabled"}} turns thinking off. See configuration.md.

Changed

• Queens reflect on memory every few replies, not after every model call. Reflection counted every model call, tool calls included, so on a busy channel it outnumbered the queen's own calls about four to one. It now counts her replies: a short reflection every 6, and every 4th of those also tidies all her memory files. memory_reflection.every_turns and long_every set both, and memory_reflection.model: "worker" runs reflection (and her dated timeline) on worker_llm. See configuration.md.

Fixed

• MCP servers on another Python version, when Hive runs from a uv-made virtualenv on Windows. The python.exe uv put in a virtualenv (from uv's #17821 until #19199) set PYTHONHOME for the interpreter it started, and every process Hive started inherited it. A server that uvx runs on another Python version then loaded Hive's standard library and died at import (SRE module mismatch), which Hive reported as "Connection closed"; under a Windows service, every health check of comfy-mcp failed this way. hive now drops that PYTHONHOME when it starts. It recognises uv's by the marker uv sets beside it, so a PYTHONHOME you set yourself is kept.
• Retired Claude model. `clau

Hive v1.0.10: Her Own Model

Each queen can run on a model of her own, set from her profile in the dashboard, and hive serve stops within seconds.

New

• A queen's own model. Every queen used the main model (Settings → AI Model → LLM). Now a queen can have her own: open her profile (Org Chart → her card) and, under Model, pick a saved provider, edit the section as JSON, or go back to the main model. It is stored under queen_llm in configuration.json, by queen id, in the main slot's shape, and gets the same key check when saved. Her running chats switch on their next model call; her context window follows her model; changing the main model leaves her on hers; her colony workers use it too unless worker_llm is set. The key can be literal, api_key_env_var or in the credential store; the subscription logins (Claude Code, Codex, Kimi, Antigravity) work only for the main model. See configuration.md.

Fixed

• hive serve stops within seconds. Stopping it started a last memory reflection for every session (up to five model calls each) and waited a minute for open connections, which dashboard event streams never close, so a service manager often killed it first. The reflections now get 6 seconds and connections 3; if the process is still there 5 seconds after that, every thread's stack goes to stderr and it exits.

For code built on Hive

• GET /api/queen/{queen_id}/llm and PUT /api/queen/{queen_id}/llm ({"section": {...} | null}) read and set a queen's own model.
• LiteLLMProvider.reconfigure() takes an optional api_key_resolver, to replace the callable that re-resolves the key before each call.

aden-tools is released at the same version. It hasn't changed, but the two packages pin each other exactly.

If you run hive serve under NSSM with AppRotateOnline 1, NSSM itself can stay in STOP_PENDING for a minute after Hive has exited; AppRotateOnline 0 avoids it (files still rotate at each start).