CanXP AI
Login
One binary. For every serious research project.

Researchbeyond the answers.

CanXP Research is a terminal-native agent that reads and writes files, runs commands, searches the web, connects to your data, and loads expertise for your field. Over 230+ domain skills.

Windows, macOS, Linux No runtime required TUI, CLI, HTTP, ACP
canxp-research
canxp-research run "tell me about frogs"
Starting session
Field: general · Agent: code
Frogs are tailless amphibians in the order Anura, with more than 7,000
known species. They are ecological indicators because their permeable skin and
dual aquatic–terrestrial life cycle make them sensitive to environmental change.
Complete · session ses_7fa2…
230
domain skills
13
field options
5
primary agents
6
ways to run

Research, not just chat

One agent across the whole evidence trail.

The same tool can audit a manuscript, run an analysis, search your notes, query a live system, and update the code that makes it reproducible. Built for research across medicine, genomics, chemistry, data science, academia, engineering, software, finance, law, security, and cybersecurity.

01

Search every source together

Fuse results from workspace files, a local vector index, Obsidian, and approved MCP tools into one ranked answer with provenance on every hit.

research_search → files + index + vault + MCP
02

Connect live data through MCP

Reach an EHR, warehouse, internal API, or vault without placing credentials in the model context. Sources remain explicit and permission-aware.

canxp-research mcp add
03

Load expertise on demand

Skills add domain-specific methods only when relevant. Your selected field carries the right expertise without spending context on all 230 skills.

/field · /skills
04

Run anywhere

Use the full-screen TUI, a non-interactive command, a long-lived HTTP server, browser interface, daemon, or an ACP-capable editor.

run · serve · web · acp · daemon

Field-aware reasoning

It reasons like your field demands.

A field changes the agent’s specialist persona and the skills available in every session. Carry the expertise you use—not all 230 skill descriptions.

/field
Or set once in config: { "field": "medicine" }

medicine

Evidence-based clinical reasoning

23 skills
Context ~2,800 tokens

genomics

Reference-aware genomic analysis

35 skills
Context ~3,500 tokens

data-science

Estimands, assumptions, and leakage checks

28 skills
Context ~2,700 tokens

phd

Study design and scholarly writing

38 skills
Context ~4,600 tokens

engineering

First principles, standards, and tolerances

17 skills
Context ~1,400 tokens

software

Correct, tested, project-native code

10 skills
Context ~1,200 tokens

finance

Bias-aware quantitative research

67 skills
Context ~2,400 tokens

chemistry

Structure-precise chemical reasoning

12 skills
Context ~960 tokens

legal

Jurisdiction-aware IRAC reasoning

0 skills
Context 0 tokens

security

Threats, controls, and risk weighting

0 skills
Context 0 tokens

cybersecurity

CWE, CVE, ATT&CK, and defense

0 skills
Context 0 tokens

general

The base research assistant

0 skills
Context 0 tokens

all

Every skill, without a field persona

230 skills
Context ~19,700 tokens

Agents with boundaries

Choose a mode. Get enforceable behavior.

Each mode has its own system prompt and permission set. Use Plan mode to plan without disrupting your work.

Default

code

Read and write files, run commands, search, and delegate. The full-capability research and coding agent.

Read-only answers

ask

Explain, inspect, and search without changing your project. For answers rather than diffs.

Edits denied

plan

Explore and design an approach while edit tools are denied at the permission boundary.

Diagnostic

debug

Reproduce, isolate, form hypotheses, and test them instead of guessing at a fix.

Parallel work

orchestrator

Break large research programs into independent units, delegate them, and synthesize the result.

Need your own agent?

Agents are portable Markdown files with YAML frontmatter in .canxp/agent/.

canxp-research agent create

Your data, with provenance

Search everything. Know what answered.

Research Search asks configured sources in parallel, merges their rankings, and labels every result with its origin. Empty and failed are never treated as the same thing.

indexSemantic search over the local vector index
filesLiteral workspace search with ripgrep
vaultFull-text search across an Obsidian vault
mcpAn approved EHR, warehouse, or internal API tool

Where do we handle patient consent?

3 documents, ranked across 2 sources.

1. src/consent.ts [files#1, index#2]

14–22: export function recordConsent(…)

2. docs/consent-policy.md [index#1]

2–6: Consent must be re-captured…

Sources:

index: ok (4) · files: ok (11)

vault: empty · mcp: excluded

Command reference

One binary. Six interfaces.

Work interactively, automate a single run, expose an API, open the web client, speak ACP, or keep a daemon warm in the background.

/field/agents/models/skills/mcps/sessions/new/status/editor/warp
$canxp-research

Open the terminal interface

$canxp-research run "…"

Run once without the TUI

$canxp-research run -c "…"

Continue the latest session

$canxp-research run -s <id> "…"

Continue a specific session

$canxp-research run --agent plan "…"

Plan with edits denied

$canxp-research run -m provider/model "…"

Choose a model for one run

$canxp-research serve --port 4399

Start the HTTP API

$canxp-research acp --cwd <path>

Start an ACP agent backend

$canxp-research mcp add

Connect your own data

$canxp-research config check

Validate effective configuration

Install CanXP Research

A native binary for your OS.

Install, open a new terminal, and type canxp-research. Sign in with your CanXP account when prompted.

Signed installers

macOS package is signed and notarized; Windows uses a signed installer.

Predictable paths

Binary, config, sessions, and cache live in standard user directories.

Updates in place

Run canxp-research upgrade, or let automatic updates use the release channel.

Designed for automation

Your research specialist can work for other agents.

Claude Code, Codex, Cursor, scripts, and CI can invoke the binary and consume newline-delimited JSON. Your calling tool never needs your MCP credentials.

CLIcanxp-research run --format json
HTTPcanxp-research serve
ACPcanxp-research acp
ask-canxp.sh
#!/usr/bin/env bash
set -euo pipefail

QUESTION="${1:?Pass a research question}"

canxp-research run --format json --agent ask "$QUESTION" \
  | python3 -c '
import sys, json
out = []
for line in sys.stdin:
    try: event = json.loads(line)
    except ValueError: continue
    if event.get("type") == "text":
        out.append(event["part"]["text"])
print("".join(out).strip())
'

Questions

Everything important before your first run.

Installation, fields, permissions, automation, and where your local state lives.

Start with one command.

Install CanXP Research, open your terminal, and give your work an agent that understands the discipline behind it.

canxp-research run "tell me about frogs"