Intelligence is the engine. Crowe Logic is the instrument.
A coding agent for your repositories, under an execution gate you control.
Crowe Logic reads, edits, tests and ships code in your repos, from the terminal or the desktop app. You set the operating mode, Plan, Read, Edit or Execute, before the agent acts, review every diff before it lands, and see the engine behind every model tier by name.
The free tier needs no card and no keys: CroweLM Flash, twenty turns a day, the full tool loop. Personal, Pro and Max open the whole table.
Why
Any engine. One instrument. Human authority.
Increasingly capable engines are being placed in the hands of individuals, teams, businesses and institutions at extraordinary speed. The model alone is no longer the whole system.
Crowe Logic is the instrument around the intelligence: a proprietary operating environment that gives one or many engines a controlled place to reason, collaborate and act. Models change. Engines get replaced. Several can operate together. The instrument stays responsible for context, routing, authority, evidence, verification, execution boundaries and the human's final say.
The human is the operator. The model is the engine. Crowe Logic is the harness that decides how the engine connects to data, tools and software, what it is permitted to do, how its actions are observed, and where a human must authorize.
A model can be replaced. The instrument remains.
Michael Crowe Crowe Logic, Inc. 2026.09.19
The Gate
You decide what the agent is allowed to do.
Plan, Read, Edit and Execute are the operating modes. Each transition is controlled.
The harness checks the mode on every tool call before anything runs, whatever the engine asked for, so the gate and the tool list can never disagree. Below Execute there is no shell; in Edit, every change is a diff you approve or reject before it touches the tree. A risk table and a read-only verifier check actions on the way through. The CLI names its levels read only, edit, execute and full.
The Engines
One interface, multiple intelligences.
Five tiers. Every engine named.
One name per job, each at normal, high or max effort, plus Auto, which picks for you. The name stays put when the engine behind it moves, so your pins and scripts do not have to.
Crowe Logic does not train these engines. Each tier runs on a named frontier model, hosted on Microsoft Azure or reached through OpenRouter, with our grounding, knowledge base, tools and orchestration on top. Ask any tier what it runs on and it will tell you.
Where a turn starts. Your knowledge base and repository context are mounted into the session, and it reads a screenshot as readily as a paragraph.
For the question the default answered vaguely. Sharp, structured, willing to commit. The other half of the dual-mode pair.
Multi-step work over very large contexts, with twice the thinking room at every rung.
Follows the patterns already in your project instead of importing its own.
Brief and action oriented, for tight iteration. Lowest credit cost per turn, and the engine of the free tier.
Routes per turn
CroweLM Auto
Reads the task and picks the tier, so a bare crowe already lands on the right engine. Vision, Grower and Kernel personas sit beside it, each on Claude Fable 5.1 by Anthropic.
On every plan, including free
Your own keys, by name
GPT-5.5, GPT-5.4, DeepSeek V4 Pro, Kimi K2.7 Code, Kimi K3, GLM 5.3 Flash, Claude Opus 4.8 and Claude Sonnet 5 sit in the same /model table under their own names, on keys you hold. No CroweLM identity, no credit meter on those turns.
Multi-engine operation
Dual mode and synthesis
Run two CroweLM tiers on one prompt and watch both stream. A third turn merges the strongest claims, judges a winner and says why, or diffs the two so the gaps are on the page.
Comparison is a switch
/replay and /fork
Re-run any past turn on a different engine. Fork a session to try another approach without losing the original. Comparing models is a switch, not a rewrite.
The Work
Read. Edit. Execute. Verify.
Built for the repository, not the chat box. Tools run on your machine, so the trust decision is made before the call, not after.
Read
Your GitHub repositories in the sidebar, a file tree, and your knowledge base and repository context mounted into the session.
A Read session cannot call a writing tool, because the tool is not in the list it was given.
Edit
Every edit arrives as a diff you approve, reject or amend before it touches the working tree. You see what changes in your repository, not a description of the fix.
Execute
A real shell, git and an in-app browser, only once the mode allows it. Worktrees let several agents work the same repository in parallel without stepping on each other.
Verify CLI 0.11.0 rc
Every tool call is written to an evidence ledger: the gate's decision, the risk, the resources touched with their revision before and after, and how the call ended.
crowe verify runs a policy over that record and returns a judgment. When the ledger cannot prove an obligation was met, the verdict is unknown, not a pass.
In the 0.11.0 release candidate on PyPI: pip install --pre crowe-logic.
Live cost per turn
Tokens, cache hits and credits on every turn, in the HUD as you work. Dual mode shows both sides. You always know what a session is burning before the invoice does.
Repos, shell, git, worktrees
Your GitHub repositories in the sidebar, a real shell, a file tree, git and an in-app browser. Worktrees let several agents work the same repository in parallel without stepping on each other.
Reviewable diffs
Every edit arrives as a diff you approve, reject or amend before it touches the working tree.
Fig. 01. The layered model from the Crowe Logic system architecture. Above the rule is what runs today: the operator's authority, the control plane with its four operating modes, named engines, routing and verify, and an evidence ledger written in the Mycelium EI observation contract. Below the rule is direction: the Mycelium EI relational network across sensor, farm, research and world data is not shipped. Architecture, not a feature list.
The Product
Desktop, CLI, web and MCP, on one Crowe ID.
The same agent against the same gateway, whichever surface you open.
Crowe Logic desktop runs the same agent against the same gateway, and adds the things a terminal cannot draw: reviewable diffs you approve or reject, your GitHub repositories in the sidebar, a real shell, a file tree, git, worktrees and an in-app browser. It updates itself.
crowe, from pip install crowe-logic. Python 3.10 or newer, on macOS, Linux or Windows. The full tool loop and every operating mode on the free tier; crowe login signs in with Crowe ID and opens the rest of the table.
The same console in a browser at crowelogic.com/app, behind Crowe ID. Desktop, web and CLI share one Crowe ID and one plan.
A remote Model Context Protocol server at api.crowelogic.com/mcp. Any MCP client connects with a Crowe Logic personal access token and gets the same control plane the console, the desktop app and the CLI use: crowelm_ask, crowelm_models and crowe_account. Every tier names the engine it runs on.
Screenshots from Crowe Logic desktop 0.24.8.
Start
Two commands and you're running.
Install, type crowe, and you are on the free tier in a repository you know. Review the first diff, run the tests, and decide from inside the console.
$ pip install crowe-logic
$ crowe
Python 3.10 or newer, on macOS, Linux or Windows. First launch runs on the free
tier with no card and no keys: CroweLM Flash, twenty turns a day.
crowe login signs in with Crowe ID, with Microsoft
or Google, and opens the rest of the table. Monthly billing, cancel any time, your history stays on your machine.
Or the app
The same agent, with a window.
Desktop 0.24.15. macOS disk images for Apple Silicon and Intel, signed and notarized. Windows installer signed with Azure Trusted Signing; SmartScreen names the publisher as Michael Crowe until an organisation validation is added. Linux AppImage and .deb. Every release ships with SHA256SUMS, and a workflow checks the update feed daily.
Pricing
Start free. Pay for the table.
Every paid plan opens every CroweLM tier, every effort rung, dual mode, and the desktop and web apps on the same Crowe ID. Credits are the meter, so a vendor rate change never reaches your invoice: a flagship turn costs 5, balanced 2, fast 1. Dual mode sums both sides and synthesis adds 5.
No card, no keys
- The full tool loop and every autonomy level
- Reviewable diffs, shell, git, worktrees
- Your own keys for the frontier picks
- Desktop, web and CLI on one Crowe ID
One developer
- Every CroweLM tier and effort rung
- Dual mode and synthesis
- Everything in Free
Daily driver
- Every CroweLM tier and effort rung
- Dual mode and synthesis
- Priority queue
- Everything in Free
Dual mode all day
- Room for 400 dual-mode turns with synthesis
- Everything in Pro
Your own keys
Already hold inference keys?
The frontier picks run on your own keys inside the same console, on every plan
including free. Point crowe init --node at your
endpoints and they appear in /model under their
own names. No separate tier, no credit meter on those turns.
By agreement
Teams and companies
Pooled credits, shared workspaces, admin cost reports, single sign-on, audit logs, data retention controls and a self-hosted node option, by agreement.
Talk to Crowe LogicCheckout runs through Crowe ID and Stripe and provisions immediately.
FAQ
Honest answers to the obvious questions.
What does a paid plan buy that a $20 chat subscription does not?
A repository workflow, not a chat window. Gated tool execution, diffs you review before they land, a real shell, git and worktrees, your repositories in the sidebar, replay and fork across engines, and dual mode with synthesis. Underneath, five named frontier engines instead of one, with a tool registry that reaches web search, browser automation, MCP servers and public records. If chat covers your work, keep the chat plan. If you ship code, start on the free tier in a repository you know and decide from inside the console.
What are credits and how far do they go?
Credits decouple what you pay us from what we pay upstream providers. A fast turn costs 1 credit, balanced 2, flagship 5. Dual mode sums both sides and synthesis adds 5. Personal's 750 credits are about 150 flagship turns a month, Pro's 3,000 about 600, Max's 6,000 about 1,200, or 400 dual-mode turns with synthesis. The HUD shows credits and cache hits live on every turn.
What actually runs underneath CroweLM?
Named frontier models. CroweLM runs on GPT-5.6 Terra and Zenith on GPT-6 Astra (both OpenAI, on Microsoft Azure), Depth on Grok 4.6 (xAI), Coder on GPT-5.6 Sol (OpenAI), Flash on GLM 5.3 Flash (Z.ai, on Cloudflare), and the Vision, Grower and Kernel personas on Claude Fable 5.1 (Anthropic, through OpenRouter or Azure). Crowe Logic does not train these engines; what we build is the grounding in your knowledge base and repository, the tool registry, the execution gate, the dual-mode synthesis and the workflow around them. The tier names stay fixed so your pins survive when we move a tier to a stronger engine, and any tier will tell you what it runs on if you ask.
Can I use my own API keys?
Yes, on every plan. The frontier picks (GPT-5.5, GPT-5.4, DeepSeek V4
Pro, Kimi K2.7 Code, Kimi K3, GLM 5.3 Flash, Claude Opus 4.8, Claude Sonnet 5) run on
keys you hold, and crowe init --node points
the console at endpoints you host. Those turns are billed by your provider,
not metered by us. The CroweLM tiers run through your Crowe ID.
What stops the agent running something it shouldn't?
Tools run on your machine, so the trust decision is made before the call, not after. The autonomy level, Plan, Read, Edit or Execute on the desktop (read only, edit, execute or full in the CLI), filters which tools exist for that turn. A Read session cannot call a writing tool, because the tool is not in the list it was given, and below Execute there is no shell. The harness checks the level on every tool call, so the execution gate and the tool list can never drift apart. A risk table and a read-only verifier check actions on the way through, and every edit is a diff you approve first. These controls limit what is available; they do not make every permitted command safe, so pick the lowest level the task needs.
Is my code trained on or sold?
No. Crowe Logic does not train models and does not send your sessions to any training pipeline. A turn goes from your machine to the edge that serves the tier you picked and back. That edge has its own data policy, which you can read on its site. Retention terms are in our Privacy page, and data retention controls and self-hosted options are available by agreement.
Can I cancel anytime?
Yes. Plans are billed monthly and cancel at the end of your billing period. No lock-in, no cancellation friction. Your session history remains in your local install regardless.
Who builds this?
Crowe Logic, Inc., founded by Michael Crowe in Phoenix, Arizona. The CLI and the desktop app are the products, and the same framework runs our other agents in adjacent markets. We are bootstrapped, not venture-backed, and the product exists because we needed it for our own work.