v0.9.9.5 release notes

v0.9.9.5 is a feature release. Modulatio can now consume MCP servers, the conversational Leader remembers a whole conversation, and the WebOS console gained a batch of terminal-parity polish. This page is the delta on top of v0.9.9.4. Every new surface was cleared by a two-lens (security + quality) adversarial review before shipping.

MCP servers — bring outside tools onto the team

Modulatio is now an MCP client. Plug in a Model Context Protocol server and its tools become first-class Modulatio tools:

pip install "modulatio[mcp]"

# a local (stdio) server Modulatio launches as a subprocess
modulatio mcp add-stdio files --command /path/to/files-server --arg /home/me/notes

# a remote (http) server with an auth token
modulatio mcp add-http hub --url https://mcp.example.com/api --auth bearer --token <TOKEN>

modulatio mcp test files      # connect + show the tools it offers

Each discovered tool registers as mcp__<server>__<tool>. The Leader can call any enabled server’s tools — each authorized once/session/always the first time, or mark a read-only server trusted for zero prompts. A producer gets a specific MCP tool only when a skill’s loadout names it (that grant is the authorization — jobs never block). MCP is a fourth capability provider on the same control plane as the built-ins and the SERVICES pool: the tool registry, the write-only key vault, the permission gate, and the spend comptroller. Secrets never enter agent context or results, and a server that won’t connect degrades to an honest “unavailable” instead of crashing a run.

The opt-in [mcp] extra keeps the base install lean — without it, a configured server is inert with an install hint. Full details in the MCP servers reference.

The conversational Leader rides the model’s full window

Talking to the Leader used to be capped at a role-sized context budget, so a long conversation would quietly compress away his memory of it. Now the chat lane rides the model’s own full context window by default — a 500K- or 1M-token model keeps the whole thread in view. The operator can cap it lower from the SETTINGS tab for cost control (every conversational turn re-sends the whole thread, so cost grows with depth — /new resets it). Every other lane — producers, QC, the Leader’s decompose and verify — keeps its role-bounded window; the exception is conversation-only.

The Leader can also read and change the team’s own configuration directly now (models, agents, services) — “what models and agents do we have” is answerable in the chat, where the permission gate previously refused the config directory. Credential files stay fenced behind the secret floor.

WebOS console — terminal-parity polish

Add model uses the live picker. Pick a provider, then pick from its real model list — fetched with your configured key server-side (the key never crosses the boundary), free models flagged, with a typed-id fallback. Parity with the terminal’s two-level picker.
Reload services (Agents tab) and Refresh (Services tab) — the terminal’s /reload, mirrored: apply model/roster/config changes without a server restart (refused while the Leader or a job is busy), and reconnect any MCP servers against the current config.
Copy works like the terminal. Drag-select in a TV — the view won’t yank to the bottom mid-selection while a run streams — then Ctrl+C / Ctrl+V; Ctrl+C with nothing selected copies the Leader’s last message whole, and the selection is now visible inside his speech block.
Static assets carry Cache-Control: no-cache so a browser can’t pin a stale interface across an upgrade.

modulatio auth login-xai signs in to xAI (Grok) with a SuperGrok / X Premium+ subscription through a PKCE OAuth flow Modulatio owns end to end. Tokens are stored write-only and auto-refreshed. If your subscription tier isn’t authorized for API access, the flow says so and points you at an API key.

`doctor` checks the model-call stack

modulatio doctor now probes the model-call import path and names any missing dependency with the remedy — a library version drift that would otherwise fail every agent call is surfaced as a plain diagnosis instead of a traceback.