An AI shell that learns your work and automates it — with your approval at every step.

Capture → compile → operate

Conch turns work you already do into governed automation. Capture your real work (opt-in), let Conch mine the recurring shapes, compile a pattern into a reviewed Architecture Card, then operate it as a supervised workflow with approvals and an audit trail — locally, across your machines and channels, or through Capitol. Capitol pairs each exact executable workflow version with a readable, searchable Procedure (the SOP); the Card authorizes the design, while the Procedure documents what that pinned workflow does.

# turn on capture, then let conch find work worth automating
/install capture
/compile suggestions            # recurring shapes from your recent work
/compile from-suggestion 1      # → a reviewed Architecture Card

Draft a card straight from a specific source:

/compile from-session                 # the conversation you just had
/compile from-mission <id>             # a durable mission's journal
/compile from-history 50 "nightly backup and prune"
/compile from-email                   # the capture mailbox
/compile from-scribe "Onboard a new vendor"   # a Scribe guide via MCP *
/compile from-procedure <workflow-id> --version 2
# exact existing Procedure + workflow version → a draft Card

Browser capture — a Chrome extension records semantic clicks, navigation, and form-submit field names (never values, never passwords, never clipboard contents) on origins you explicitly allowlist, delivered to your local kernel over native messaging — no ports, no cloud. The extension ships inside the package, and setup creates a stable copy plus a native-host launcher that survives upgrades:

# set up the extension + native host, load it in Chrome,
# allowlist an origin in its options page — then:
/install capture browser
/compile from-browser example.com "automate this web procedure"

* Capture is opt-in (/install capture); compiling/operating processes needs the optional works module (/install works); from-scribe reads a Scribe workspace over MCP. Procedure commands require Capitol's ontology API rollout; before that backend epic deploys, they fail closed instead of compiling on read. Browser capture is proven first on Chrome for macOS; Linux host paths ship but are experimental until equivalently proven.

Components & capabilities

Conch starts as a local-first AI shell. Ask for a command and approve it before it runs, or open a full chat session that can execute shell, SSH, and MCP tools under graded permissions, with searchable conversations, memory, and cost tracking. It works with Anthropic (Claude, including Opus 5.5), OpenAI (including GPT-6 Sol and Luna), OpenRouter, Bedrock, a local Ollama, or any OpenAI-compatible server (llama.cpp, vLLM).

Its agent control plane carries approved work beyond one terminal: durable missions survive reboots and keep working across your machines, channels, and Capitol, with every risky step routed back to you as an approval. The shell is the base; everything else is an optional install from inside it.

Capabilities

• Local-first — run entirely against llama.cpp. Cloud providers are optional.
• Fleet-aware models — a local llama-idx registry discovers your self-hosted backends; /llamaidx shows status, /model switches to any of them.
• Live token stats — tokens, cost, and true tok/s after every reply, server-reported by llama.cpp and Ollama (/tks toggles).
• Sandboxed execution — run approved commands in an ephemeral Docker container or an E2B cloud sandbox instead of locally, same approval gate (/sandbox docker | e2b | off).
• Safe code edits — automatic per-turn git checkpoints on conch-only refs (never your branch history, secrets excluded); /checkpoint list, /undo to roll back.
• Nearly stdlib-only — pygments is the lone hard dependency; installs are small and auditable.
• Event-sourced mission kernel — every decision journaled, budgeted, and replayable.
• Origin-bound approvals — remote requests can't approve themselves; risky steps come back to you.
• Fail-closed everything — graded permissions, hidden key input, 0600 secrets, no trackers.