ClaudeR - The Modern Researcher's Toolkit

ClaudeR connects MCP-configured LLM agents, like Claude Code or Codex, to the R session you are already working in. The agent runs code where your cleaned data and fitted models actually are, sees the plots it draws, reads the objects in your environment, and can edit the file open in your editor. A companion Python package does the same for a live Python session, and both work in RStudio and Positron.

Several agents can share one session and divide the work between them, or you can run separate sessions side by side so each agent works on its own dataset without touching the others.

It covers much of what a paid research AI tool does, with the work happening in your own session and a record of every line an agent ran.

Quick Start

# Install
if (!require("devtools")) install.packages("devtools")
devtools::install_github("IMNMV/ClaudeR")

# Set up your AI tool
library(ClaudeR)
install_clauder()          # For Claude Desktop / Cursor
install_cli(tools = "claude")  # For Claude Code CLI

# Start the server in RStudio
claudeAddin()

AI agents: See llms-install.md for automated setup instructions.

Demo

Single agent via Claude Desktop App	Multi-agent: Codex + Claude Code via CLI	GPT 5.4 Codex: Data analysis + Quarto report

• Console logging from the addin fixed (R 0.14.2). Ticking "Also log my own console commands" in the addin armed nothing on 0.14.1, while calling start_console_logging() in the console worked. Cause: globalCallingHandlers() (used to observe warnings and messages) can only be called with an empty handler stack, and inside the Shiny observer that handles the checkbox it errors, which aborted the function before the console callback was registered. The callback is now registered first and unconditionally; the handlers are installed only when logging is started from the console, and otherwise warnings are read from last.warning after each command. Started from the addin, message() output is the one thing not captured, and the addin says so.

• Control what reaches the agent, and stop sending plots nobody saw (R 0.15.0 / clauder-mcp 0.15.0). From two user requests. A new AGENT CONTEXT panel in the addin does two things. "Send plots to the agent automatically" can be turned off: the agent is told a plot was drawn and asks for it with execute_r_with_plot, saving roughly 4,500 tokens per image when you mostly want text. Your Plots pane is unaffected either way. "Tools available to the agent" hides whole groups you are not using (editor, session, coordination, background jobs, manuscript audit); all 41 tools cost about 8,600 tokens of schema on every request, and a programming-only session needs about 1,200 of that. Both settings are per session and take effect without a restart, though some MCP clients only notice a changed tool list on reconnect. Separately, a bug fix: p <- ggplot(...) returned a full image to the agent even though assignment displays nothing, so the agent was sent plots that never appeared in your Plots pane. Capture now requires the plot to have actually been shown.

• Logging fixes (R 0.14.1). Two problems from a user report. Agent output was never written to the log, only the code, because the entry was written before the code ran; the log now records what each call printed. And the output capture added in 0.14.0 could stop console logging entirely: it opened its own output sink before registering the console callback, so if that sink could not be opened, nothing was logged at all. Opening it is now optional and failing back to logging commands and their values, the sink is restored if another execution unwinds it, and an agent run no longer pops a sink it did not open. Logged output lines are marked so that replay, history and notebook export skip them.

• Console logging now captures output, not just results (R 0.14.0). From follow-up on the logging request. The log recorded the value a command returned, so anything printed as a side effect was missing: cat(), progress output, and print() called inside a function. Running f() showed the call but not what f() printed, which is exactly the context an agent needs when you hand it an error you hit yourself. Standard output is now teed to the log and grouped under the command that produced it, alongside the warnings, messages, and errors already captured. Your console still shows everything as normal.

• Windows session-liveness fix (R 0.13.2). Checking whether a recorded R session was still alive used tools::pskill(pid, signal = 0). That is the standard idiom on Unix, but on Windows pskill always calls TerminateProcess regardless of signal, so the check killed the session it was asking about, then reported it as alive and left the stale discovery file in place. Starting a server could therefore terminate another RStudio session and leave agents routed to a dead port. Liveness is now probed without signalling. Separately, registering a session whose name is already held by a different live session now warns instead of silently overwriting its discovery file, which had left agents holding the wrong port and token.

• Unified console logging (R 0.13.0). The session log recorded what the agent ran, but not what you ran, so asking an agent to explain an error you hit in the console meant re-running the work through it. Tick "Also log my own console commands" under Logging and your console activity joins the same file, tagged by who ran what: the command, its printed result, and any warnings, messages, or errors. "Read the last 100 lines of the log" is now enough for an agent to see both sides of the session. Off by default; toggle it in the addin or call start_console_logging() / stop_console_logging().

• Editor tools fixed, plus approve-before-apply edits (R 0.12.5 / clauder-mcp 0.14.4). From two user reports. modify_code_section and insert_text now save to disk by default and report saved_to_disk, so an agent no longer believes it wrote a file when the change sat unsaved in the buffer. Bounded replacements may change the line count (the old equality constraint is gone). Both tools accept a path to target a specific file, open and focus it, and refuse to edit a different document instead of failing silently. get_active_document now reports the path, the document id, and whether the buffer differs from disk, so buffer state and file state stop being confused for each other. New suggest_edit tool: the agent proposes a change and waits for the user to approve it, using rstudioapi::showEditSuggestion() when the installed rstudioapi provides it, and otherwise staging the edit unsaved so the user accepts by saving or rejects with undo.

• Reviewer Zero now reasons, not just reconciles (R 0.12.4). Added a mandatory Pass 5 (content reasoning) to the base auditing protocol. The deterministic tools (reconcile_values, verify_references, check_cross_references, probe_scripts) find numeric, reference, cross-reference, and code defects, but they do not reason about meaning, and a manuscript can clear every one of them and still be wrong. Pass 5 is a gated, equal-weight pass with eight checks the tools cannot do: instrument and source attribution, whether each reported test is computable from the data that exists, whether each cited figure or table actually contains the claimed evidence, magnitude wording, convergence across studies, causal and generality framing, data existence for descriptive claims, and supplement and appendix integration (every supplement cited, and no body claim resting on an uncited one). In a controlled benchmark on a synthetic manuscript with a known defect set, this raised detection from below a native-tools baseline (16.8 of 24) to clearly above it (21.3 of 24), recovering exactly the reasoning defects the old tool-led protocol was missing, with no rise in false positives.

• Shared-connection identity, coordination visibility, and a stale-session guard (R 0.12.2 / clauder-mcp 0.14.2). Three fixes from a live three-persona field session. (1) Personas sharing one MCP connection were renaming each other, because set_agent_name changes the identity of the whole connection. New as_agent parameter on send_message, check_messages, and wait_for_message acts as a named persona for one call, with a separate read cursor per name. The bridge now enforces the pattern: a second set_agent_name with a different name is refused unless forced, every send confirmation echoes the name it was sent as, and the agent intro states where the current identity came from. (2) Coordination messages bypass R by design, so the console and the Agents panel showed nothing while agents talked. The addin now echoes each new coordination event to the console in full (no truncation), appends it to the session log, and shows a live coordination roster with last-seen ages. (3) A bridge still pointed at a dead R session used to report "success" while writing to a coordination log no live agent reads. Coordination calls now fail loudly when no live session exists, and announce it when the connection re-binds to a different live session. Also from pilot 2: check_cross_references understands S-prefixed supplement numbering, and the audit protocol documents unname() for htest fields and author-plus-year citation matching.

• Agent identity and cross-restart history (R 0.11.0 / clauder-mcp 0.13.0). Built from field reports of a multi-day, three-agent session. New set_agent_name tool: an agent sets its working name (for example "Claude-Stasis") once, and execution history, message attribution, presence, and its read cursor all carry that name. This fixes the case where several agents or personas share one MCP connection and collapse into a single random id. get_session_history gains include_past: it parses prior session log files on disk, so the audit of who ran what now survives R restarts. The coordination protocol now makes identity the first step of check-in.

• Researcher workflow release (R 0.10.0 / clauder-mcp 0.12.0). Three workflows that paid tools charge for, built on machinery ClaudeR already had. (1) Systematic review screening: two independent AI screeners from different model families judge every abstract against your criteria, and the new screening_report tool computes agreement, Cohen's kappa, PRISMA flow counts, and the conflict set, so the human reads only the disagreements. (2) Grant Panel Mode: grant_panel_prompt(rubric = "nih") convenes a mock study section, one reviewer per criterion, anchored weaknesses, and a ranked list of revisions that would move the score. (3) Response to Reviewers: reviewer_response_prompt() parses a decision letter into a point-by-point registry, reruns analyses so answers carry real computed numbers, gates until every point is answered, and exports the response letter with export_response_letter() plus Word comments in the manuscript.

• Coordination v2 + consensus gate (R 0.9.0 / clauder-mcp 0.11.0). The multi-agent board is now a typed, append-only event log on disk, designed from field reports of real multi-hour multi-agent sessions: typed signals instead of prose-grepping, per-agent read cursors (the shared-dataframe race is structurally impossible), reply threading, lease-based task claims, a fact store for shared state, auto-stamped presence, and a wait_for_message tool that blocks until a partner's event arrives, without touching the busy R session. New consensus gate: after propose_plan(), every execution response carries an agreement banner until all agents run confirm_agreement() with the required sentence verbatim. Only then is the plan marked approved. Plus referee_prompt(stance = "reviewer2"): a hostile-but-fair Reviewer 2 that opens with an unprimed three-sentence read (central claim, and would a strong venue accept it?) and ranks findings fatal / must-fix / minor, each tagged to the study it concerns.

• Referee Mode v2: configurable reviewers (R 0.8.0). referee_prompt() now takes lenses (run any subset), reviewers_per_lens (2 = adversarial prosecutor + verifier pairs, 3 adds a backwards reader), model (a tier for all subagents, e.g. "haiku" for quick passes or "opus" for submission-grade, or a named per-lens vector), and cross_vendor = TRUE to dispatch logic/methods reviewers to a different model vendor via codex/agy/qwen one-shots. Anti-collapse rules are now mandatory in the protocol: reviewer prompts must be written from scratch per lens and stance, the consistency lens reads back-to-front, and every finding carries a corroboration count across independent reviewers.

• Referee Mode (R 0.7.0 / clauder-mcp 0.10.0). The substantive manuscript review that paid services charge ~$50 a pass for, running free on the subscription you already have, and delivered where it belongs: as Word comments in your manuscript. reviewer_zero_prompt(referee = TRUE) (or standalone referee_prompt()) runs five content-only review lenses (argument logic, methods, internal consistency, evidence presentation, framing) as parallel subagents where the CLI supports them. Every finding must anchor to a verbatim quote and survive an independent verification pass before it lands in the document, severities are kept honest, and a clean report on a sound paper is a valid outcome. Alongside it, the new check_cross_references tool deterministically catches dangling references ("see Table 4" when Table 4 no longer exists) and tables or figures the text never mentions.

• Value-first auditing (R 0.6.0 / clauder-mcp 0.9.0). Built from field feedback after a full manuscript+supplement audit. read_file now transparently extracts .docx/.pdf (previously returned raw bytes), and the extractor preserves structure: headings marked, table cells emitted row-wise with separators (they were previously dropped entirely). New reconcile_values tool: enumerates every number in a manuscript and reconciles each against the corpus your code produced, respecting displayed precision (5038.5 matches 5038.46), commas, percents, scientific notation, and < .001 thresholds. The per-value values_registry makes numeric completeness a construction, not a diligence hope. Reviewer Zero now sets audit-clean print options (no more tibble 3-sig-fig false alarms), gates on the value sweep, and takes final verdicts from clean-room runs via probe_scripts(capture_output = TRUE). CrossRef lookups retry with backoff instead of silently truncating the reference check on 429s.