imperal-mcp

imperal-mcp - The Exoskeleton for Your AI Coding Agent 🐝

Stop babysitting your AI. Put it in an exoskeleton: it remembers everything, sees your whole codebase, can undo any change and can't fake a result - or hallucinate one past you. Fewer wasted tokens, fewer broken builds. Works with Claude Code, Cursor, OpenAI Codex, Gemini CLI, Windsurf, Claude Desktop, Cline and Goose.

pip install "imperal-mcp[code]" && imperal-mcp login && imperal-mcp setup

The code tools work offline, no account needed. Memory across sessions, the kernel's frame and the hallucination check run on an Imperal Cloud account (imperal-mcp login) - sign-up is free, plans and credits: imperal.io/pricing.

---

The problem

AI coding agents are brilliant - and work blind:

• They forget. Every new session, every switch from Cursor to Claude Code, starts from zero. You explain your architecture again. And again.
• They burn tokens. To look at one function they read the whole 2,000-line file, then grep the repo and read three more files.
• They hallucinate. A function that doesn't exist. A file it never opened. "I updated the caller" - which one? "All tests pass" - nobody ran them. Said with total confidence, and you only find out in production.
• They break things. They edit a shared helper without knowing who calls it, and a failed experiment leaves your working tree dirty.

The fix: put your agent in an exoskeleton

imperal-mcp wraps the agent you already use. Same model, new abilities:

Your agent today	In the exoskeleton
Forgets everything between sessions and tools	Total memory - every step, finding and decision of every session, across terminals, IDEs, the web panel and Telegram
Reads whole files to find one function	X-ray vision - a live map of your code: it reads the 28 lines it needs, not 2,000
Edits blind	Knows the blast radius - every caller and file a change reaches, before the edit
Leaves broken experiments behind	A safety net - invisible checkpoints; any change undone in one call, your .git untouched
Invents functions, files and results	Hallucination check - every name, file and result it states is matched against what it really read and ran
Says "done" when it isn't	A referee - a task closes only with proof the kernel saw: a diff, a check that passed, a finding with sources
Works alone	A team - sees your other agents working right now and talks to them

An illustration of one task - your own numbers: token_savings() measures them in every session.

WITHOUT the exoskeleton
  1. cat service.py (1,800 lines)            -> 72,000 chars dumped into context
  2. grep -rn "process_order" .              -> 140 noisy matches
  3. read 3 more files to trace callers      -> +95,000 chars burned
  4. edit a shared function blind            -> 2 callers broken, a syntax error left behind
  5. "I also updated retry_backoff()"        -> that function doesn't exist
  6. "Done, tests pass!"                     -> it never ran them
  7. open Cursor tomorrow                    -> total amnesia

WITH the exoskeleton
  1. frame                                   -> the task, what is proven, the code around your focus, the world
  2. read_symbol("service.py", "Order.run")  -> only the 28 lines that matter
  3. impact_of_change(["Order.run"])         -> every caller, before you touch a line
  4. checkpoint("before-refactor")           -> undo in one call if it goes wrong
  5. task(goal, items)                       -> an item closes only when the kernel saw the proof
  6. "I also updated retry_backoff()"        -> held back: no file it read holds that name
  7. tomorrow, any tool                      -> recall: the whole history is there

---

Hallucinations: caught before they reach you

You can't trust a sentence just because it sounds sure. So the exoskeleton doesn't trust the model's words - it checks them against the record of what actually happened in this task.

On every step, the names the agent's thinking leans on - functions, files, sizes - are matched against what the task really read, what its commands really returned and what your code map really holds. A name found nowhere comes back in the next frame as NAMED BUT NOT FOUND, before the agent builds on it.

On the final answer, the same check runs once more. Name a symbol nobody read, a file nobody opened, a size nobody measured - and the answer doesn't reach you. It goes back to the agent:

KERNEL FACTS: your answer states what the kernel did not see happen:
- `retry_backoff` -- no file this task read and no output it got contains this name
Correct the answer to the facts, or do the missing work first.

The agent then reads the code it skipped, or corrects the claim. Either way, what you get is backed by something real.

On "done", words count for nothing: a plan item closes only on a proof the kernel measured - a diff, a check that exited 0, a finding with its sources. "Tests pass" without a test run is not a pass.

What it can't do - and we'd rather tell you: it doesn't make the model smarter. An agent can still draw a wrong conclusion from true facts. What it can no longer do is pass off what it never saw.

---

Powered by ICNLI - the new HTTP for AI

HTTP gave every computer one way to talk to every other. ICNLI gives people and AI agents - and agents among themselves - one way to work together that can be proven: whatever the model, whatever the tool.

ICNLI (Infrastructure Contextual Natural Language Interface) is the open protocol at the heart of the exoskeleton - and the reason it can promise honesty, not just speed.

Under ICNLI, an agent's work is a state the kernel can prove, not a story the model tells:

• Every step is measured. What a call read, changed or returned is recorded by the kernel - never taken from the model's words.
• Proven vs. claimed. The step frame shows what is proven and what is only claimed; a plan item is done only when its proof (a diff, a check that exited 0, a finding with sources) was seen.
• Decisions from facts. Progress, loops and stalls are computed from what the steps returned, and stated to the agent in plain words.
• One state, every surface. The same session state serves the terminal, the IDE, the web panel and Telegram.

MCP is how your agent talks to the exoskeleton. ICNLI is how the exoskeleton thinks. The specification is open (CC BY-SA 4.0) at icnli.org.

---

Quick start

# 1. Install (Python 3.7 - 3.14; the code map included)
pip install "imperal-mcp[code]"

# 2. Sign in to Imperal Cloud (memory, the kernel's frame, apps)
imperal-mcp login

# 3. Wire every shell found on this machine: server + hooks + protocol
imperal-mcp setup            # Claude Code, Codex, Antigravity CLI, Gemini CLI

That's it. From the next prompt on, your agent starts each step from the kernel's frame, remembers across sessions and gets its final answer checked.

Claude Code only, in this project:

imperal-mcp claude-setup --with-hooks

Adds the imperal server to .mcp.json (keeps your entries, refuses to overwrite malformed JSON), appends the exoskeleton protocol to CLAUDE.md, and installs a PostToolUse hook that syntax-checks every Edit/Write.

Claude Code CLI:

claude mcp add imperal -- imperal-mcp

Cursor / Claude Desktop / Windsurf / Cline (.mcp.json or claude_desktop_config.json):

{
  "mcpServers": {
    "imperal": {
      "command": "imperal-mcp",
      "args": [],
      "env": { "IMPERAL_WORKSPACE": "/absolute/path/to/your/repo" }
    }
  }
}

Any other MCP stdio client: run imperal-mcp with IMPERAL_WORKSPACE set to your repository root.

Pricing: sign-up is free; plans and credits are on imperal.io/pricing. In imperal-mcp only frame spends credits; every other tool is free, and the local code tools work offline.

---

For the geeks: how it works

The loop

1. frame - at the start and after each change. The kernel's step frame: the task and plan with what is proven, the code around the focus from the exact graph, what the change reaches, the machines and repos, the path so far, and a DECISION section to act on first. With imperal-mcp setup, the shell's prompt hook puts it in context automatically.
2. task(goal, items) - declare the request and the plan; an item closes only by a proof the kernel saw. An item that cannot be finished is marked blocked with the reason.
3. Code - orient -> outline -> read_symbol; search_code / grep_code for exact text; impact_of_change before editing a shared symbol. The same tools reach a server's repo with target="host:<alias>" or target="ssh:<connection_id>".
4. Safety - checkpoint before risky edits, checkpoint_diff, checkpoint_rollback; check_syntax after.
5. Memory - every call, thought and answer is written to the vault by the shell's hooks; look back with recall, search_history, state_lens, vault_log_read.
6. The final check - the stop hook sends the final answer to the kernel: an answer that names a file, symbol or size the task never read is handed back to the agent, and open plan items are listed to you.

Hooks (installed by imperal-mcp setup)

Event	What happens
session-start	The session is bound to its kernel state; a missing install is reported
prompt	Your request is recorded; the kernel's frame for this step is put in context
post-tool	Every call and its result go to the vault; edits are syntax-checked
stop	The final answer is checked against the facts; open plan items are named

Tool catalog (34 tools, 4 resources, 1 prompt)

Kernel & task (Imperal Cloud)

Tool	What it does
frame(root?, target?)	The kernel's step frame - task, proofs, code around the focus, reach of the change, world, path
task(goal, items?)	Declare the request and the plan; items close only by a measured proof
state_lens(section?, focus?, conversation?)	What the kernel holds, by section (law, plan, reality, memory, conversation, path) - yours or another live session's
vault_log_read(ref, offset?, limit?, grep?)	An output the kernel kept in the vault (vault://logs/...), paged or grepped

Memory & other sessions (Imperal Cloud)

Tool	What it does
recall(ref, offset?, n?, conversation?)	A past request or answer whole - of this session or another of yours
search_history(query, limit?)	Your past conversations that hold a text - Webbee, the panel, Telegram, other shells
live_sessions(minutes?, limit?)	Your other sessions working now, with their newest steps
message_session(conversation, text)	Write to another live session as an agent; its answer comes back in your frame
recall_context(query, limit?)	Standing rules, preferences and repo notes, with PII scrubbed
remember_fact(fact, category?, tags?)	Keep a durable fact or rule across every surface
save_repo_note(note, repo?, citations?)	Keep a citation-backed note about a repository

Code map (imperal-mcp[code], local or on a server via target)

Tool	What it does
orient(query, k?)	One-call orientation: repo profile + top-ranked symbols and matches for the task
repo_profile()	Languages, file counts, entry points, test hints, git ref
search_code(query, k?, path_glob?, kind?)	Hybrid symbol + semantic + lexical search
grep_code(pattern, path_glob?, regex?)	An exact text or regex inside the code
outline(path)	A file's structure with exact line ranges, without its body
read_symbol(path, name)	Only one function, class, method or line range (L120-180)
graph_slice(symbols, depth?)	Definitions, signatures, callers and callees
contract_slice(symbols)	Signatures, returns, raises and real call-site argument shapes
impact_of_change(symbols, full?)	Every file and call site that depends on what you plan to change
check_syntax(path)	Tree-sitter (and Python ast / JSON) syntax check after an edit
coding_context()	Project tree and git status snapshot
token_savings()	Measured characters and lines returned vs. the real files on disk

Safety net (local)

Tool	What it does
checkpoint(label?)	An invisible ShadowGit snapshot - your .git history untouched
checkpoint_diff(target?, paths?)	Diff of the workspace against a checkpoint
checkpoint_rollback(target?)	Restore a checkpoint (a safety snapshot is taken first; marked destructive, so your client asks)

Apps on Imperal Cloud (ICNLI)

Tool	What it does
validate_ir(app_ir)	Validate an app.ir.json locally - the envelope and every declarative step
smoke_ir(app_ir, function, args?)	Dry-run a handler in an isolated sandbox: {ok, result, trace}
deploy_ir(app_ir, app_id)	Deploy the app to your Imperal Cloud account
list_apps() / get_app(app_id)	Your apps, their manifests, tool schemas and action types
run_read_tool(app_id, function, args?)	Run a read tool; PII scrubbed from the answer
run_write_tool(app_id, function, args?)	Run a write tool; destructive ones ask you first, money ones need panel approval
upload_file(path, name?)	Upload a file to your cloud storage, get a storage:// URI

Resources & prompt: imperal://exoskeleton-protocol (the operating discipline), imperal://ir-spec, imperal://ui-catalog, imperal://examples, and the build_imperal_app prompt for designing, validating, smoke-testing and deploying an app.

CLI

imperal-mcp                  # Start the MCP server over stdio
imperal-mcp login            # Sign in via browser (device code + PKCE); --no-browser prints the URL
imperal-mcp logout           # Revoke the refresh token and delete local credentials
imperal-mcp setup [shells]   # Wire claude | codex | antigravity | gemini (default: the ones found); --remove undoes it
imperal-mcp claude-setup     # Configure .mcp.json and CLAUDE.md for one project (--workspace, --with-hooks)
imperal-mcp install-hooks    # Install the Claude Code PostToolUse syntax-check hook
imperal-mcp hook <event>     # Run a hook event: session-start | prompt | post-tool | stop (JSON on stdin)
imperal-mcp status           # Workspace readiness and index profile
imperal-mcp watch            # Messages from your other agents, one line each, as they arrive (for a shell monitor)

Privacy

See PRIVACY.md for every data flow in detail.

Python 3.7 - 3.14

On Python 3.10+ imperal-mcp runs natively. On older system Pythons (3.7 - 3.9) a zero-dependency launcher provisions an isolated Python 3.12 runtime via uv on first launch - the full code map works even on legacy systems.

---

Built like a vault

Your agent's memory is your work - your code, your decisions, your mistakes. So it's guarded like one.

What leaves your machine

Stays on your machine	Goes to your private space in Imperal Cloud
Your code map, symbol reads, syntax checks and checkpoints - all local tools work offline	The steps of a session the shell's hooks record: requests, calls and their results, thoughts and answers - that's your agent's memory
Your credentials (~/.imperal/credentials.json, 0600)	Facts and repo notes you choose to keep
Secrets and personal data - removed before anything is sent	Apps you deploy

Scrubbed on your machine, before sending. Private keys, passwords, tokens, JWTs, API keys, authorization headers, credentials in URLs, emails, phone numbers, payment cards and SSNs are masked locally by imperal-mcp itself - before a single byte leaves your machine.

How it's locked down

• Your own sealed space. Every user's memory lives in its own isolated space, in a private storage network that is never exposed to the internet. Each internal service holds its own narrow key and reaches only what it needs.
• Secrets sealed with keys that can't be taken out. Your connected accounts' tokens and credentials are encrypted in Vault with non-exportable keys - a full copy of the database reveals nothing.
• Zero keys on disk. No platform secret is ever written to a server's disk: every service gets its keys from Vault at start, bound to that one machine.
• Distributed infrastructure. Isolated environments across independent data centers and providers - no single provider holds the whole platform.
• Durable by design. Every agent turn runs as a durable workflow: a restart or an outage costs time, never your work.
• Encrypted in transit. Everything between your machine and the platform goes over HTTPS; you sign in with OAuth 2.0 device code + PKCE, and no password ever passes through imperal-mcp.

And on your machine

• Workspace confinement. Local tools resolve every path and refuse anything outside your workspace root.
• Your config files are safe. setup, claude-setup and install-hooks refuse to overwrite a JSON file they cannot parse; setup --remove takes away exactly what it added.
• You stay in charge. Destructive actions ask you first; money moves only with your approval in the panel. A message from another agent informs - it approves nothing.
• Your memory, your call. See everything your agents remember and delete any of it - or all of it - from the Imperal panel. Delete your account and your space goes with it.

---

FAQ

Does it replace Claude Code / Cursor / Codex? No. It's an exoskeleton, not a new agent: you keep your tool and your model - the exoskeleton makes them remember, see, check and prove.

Which model does it use? Yours. The kernel does the bookkeeping - measuring, remembering, checking - and your agent's model does the thinking.

What does it cost? Sign-up is free. Plans, included credits and top-ups are on imperal.io/pricing. In imperal-mcp only frame (the kernel's step frame) spends credits; every other tool - the code map, checkpoints, memory, history, messaging - is free.

Does it work without the cloud? The code map, symbol reads, blast radius, syntax checks and checkpoints run fully offline. Memory across sessions, the kernel's frame and the hallucination check need your Imperal Cloud account.

How is this different from an MCP server that searches code? Search tells your agent where things are. The exoskeleton also remembers what it did, knows what a change will break, holds it to its plan and checks every claim against what really happened - under one protocol, ICNLI.

Can I remove it? imperal-mcp setup --remove takes away exactly what setup added. imperal-mcp logout revokes the sign-in and deletes the local credentials.

---

Author

Created by Valentin Scerbacov (valentin@imperal.io) and built at Imperal.

---

License, protocol & trademarks

• Code: GNU General Public License v3.0 (GPL-3.0-only).
• Protocol: ICNLI, an open specification under CC BY-SA 4.0.
• Privacy: see PRIVACY.md - local execution, credential storage and PII redaction.
• Trademarks: Imperal, Imperal Cloud, Webbee and ICNLI are trademarks of Imperal, Inc. Claude, Claude Code, Claude Desktop and Anthropic are trademarks of Anthropic, PBC. Cursor is a trademark of Anysphere, Inc. OpenAI and Codex are trademarks of OpenAI, Inc. Windsurf is a trademark of Exafunction, Inc. (Codeium). Gemini is a trademark of Google LLC. All third-party names indicate MCP interoperability only; no endorsement or affiliation is implied.