The AI coding assistant that
doesn't lock you in.
Your providers. Your credentials. Your data. Your cost. Anvil runs in your terminal, drives from your browser, and remembers across sessions — with a mind map of everything it's thinking.
35 providers · encrypted vault · 138 commands · zero telemetry
test utils::tests::merges_sorted ... ok
test result: ok. 1 passed; 0 failed; 0 ignored
All green — the test covers interleaved values, empty inputs, and duplicates.
Other assistants come with a leash.
Anvil is the inverse.
Most AI coding tools route through one vendor, hold your keys, and phone home. Anvil hands you the reins — every layer is yours to control.
Any provider
Anthropic, OpenAI, Gemini, xAI, Ollama, and 30 more. Switch mid-session with /model.
Your keys
Bring your own API keys, held in a local encrypted vault. Anvil never sees a key it didn't get from you.
Your data
Sessions, memory, and your mind map live on your machine as plain files. They travel with you.
Zero telemetry
No phoning home. What happens in your terminal stays in your terminal.
One model's answer looks the same
whether it is right or wrong.
So stop asking one. Anvil puts a panel of models on the same question — each through its own client and its own key — and shows you where they disagree instead of averaging it away.
Ask a panel
/consult asks several models at once and reports agreement, divergence, and what the minority saw that the others missed. It never picks a winner for you.
Or a chain
/relay hands one problem down a line of models in order, each seeing the real transcript of the work before it — redacted at every vendor boundary.
It refuses to fake it
A consult of one is not a consult. A single model working alone is a normal turn, not a relay. Anvil says so rather than pretending.
See the money first
Every turn priced against a live registry. A budget cap refuses before the provider call — and an unpriceable turn reports unknown, never a comforting $0.
Use a secret you never see
Paste a credential and Anvil vaults it, leaving a handle it resolves only when a tool actually runs. The model reasons about a name; the transcript keeps a name.
Lend without giving away
Hand a colleague the label of a vault entry for a set number of minutes. They can use it, cannot read it, and cannot lend it onward. It expires on its own.
Somebody did that.
Anvil knows who.
Several people can work in one live session, each with their own identity, models and permissions — enforced where the work happens, not just hidden in the interface.
Attributed to a person
Actions carry a verified identity — subject, role, capabilities, workspace. A peer's message is labelled as theirs, and tool calls name who caused them on the durable record.
Nothing unrestricted by omission
An empty scope grants nothing. Whole-machine reach is a flag somebody set on purpose, never an accident of a blank field.
Refuse, then offer a door
A blocked request can ask for a one-time or permanent increase. It reaches AnvilHub, an admin decides, and the answer returns to the running session.
Roles you author
Define roles with their own capabilities and workspace scope rather than picking from a fixed menu. The built-in roles stay put and cannot be redefined under you.
Joining is not handing over
Bringing an Anvil into an organization grants other people access. Ownership stays with one person — by design, not by omission.
Run them all from one place
Claim a deployment from the web: name it, give it a persona, pick its primary agent, watch it live. Install or remove capabilities — applied live when it is online, queued when it is asleep.
Expertise that costs nothing
until you actually use it.
A real agent library is dozens of specialists. Telling the model about all of them costs thousands of tokens of roster on every single turn — spent before any work happens, and growing every time you add an agent. Anvil refuses that trade.
Twelve names, not a roster
The model sees twelve squad names and a one-line brief each — about 300 tokens, and that cost is fixed. No agent names, no counts. Add a hundred specialists and it does not grow by a line.
Ask for a capability
You ask for security testing, not for a name you had to memorise. The host resolves the concrete specialist from the full roster and loads it for that job — then lets it go.
Grow it without a release
Install a whole new squad from the marketplace and it becomes routable immediately. No Anvil update, no waiting on us to ship a version that knows your team.
An empty squad says so
Ask for a capability nobody on the team has and Anvil refuses — rather than quietly handing the job to a general-purpose agent and letting you believe a specialist did it.
Every machine you own
/fleet sends one task to every eligible Anvil on your account and collects the answers — routed by capability, so the work lands on a machine that can actually run it.
No silent half fan-out
It names what was sent, what was refused, and who was never eligible — the offline laptop, the host without that capability. A fan-out that reports only its successes looks identical to one that lost half your fleet.
It thinks in a map.
A living, force-directed graph of your work — wired into Anvil's memory so it reflects what it's actually learning. Drive it from the browser: drag anything, zoom to the cursor, click any node for its full notes.
Force-directed
Nodes repel so the graph stays readable at hundreds of nodes.
Memory-aware
Today's memories auto-populate; toggle to a color-coded Memory Graph.
With atmosphere
An ambient starfield drifts behind it — on the web and in the terminal.
Terminal and browser, one session.
Start in the terminal, hand off to a paired browser tab, and back again — it's the same live session the whole time, never a copy. Nothing is lost in the handoff.
- ✦ Pair a browser with one command; drive the exact session you're running.
- ✦ The web viewer renders like a product — message cards, glowing code panels, real markdown.
- ✦ Hand off without flashing or losing history; wake the terminal and it's all there.
It remembers what matters.
Anvil remembers across sessions — working, episodic, semantic, procedural, reflective, and a long-term knowledge archive. Click any layer to browse what it knows.
- ✦ Every memory is a card with a colored type badge and full text on click.
- ✦ Memories link to each other and to your mind-map thoughts.
- ✦ All local — it's your knowledge, and it stays yours.
Everything, on your terms.
300+ commits since v2.2.5 — SSO, the seven-layer memory, live remote control, the mind map, and now v2.2.33.
One command. Seven targets.
Native binaries for every platform your team runs.
Windows irm https://anvilhub.culpur.net/install.ps1 | iex
Need a FIPS-validated build for regulated environments? See Anvil FIPS.
