Tools and integrations of the Nous agent

Shell, files, web search, email and Telegram. A REST API and an MCP server. Schedules, background subtasks and plans that can wait for your approval.

Get Nous on GitHub Read the quickstart

Plan · dag_create · release notes5 nodes

1. Collect merged pull requestsbash · gh pr listdone

2. Summarise each changeBackground subtaskdone

3. Draft the noteswrite_file · notes/1.2.0.mddone

4. Publish the releaseExternal action: waits for the ownerneeds approval

5. Post the summary to TelegramRuns after step 4queued

Node timeoutA reaper stops a node that runs too long

DeliveryResults reach the conversation at least once

WaitingStep 4 of 5

Publish the release: approve or decline

Nous tools and interfaces

Nous calls its own tools in a direct Anthropic API loop, and you reach it from the outside in four ways.

Work bash, read_file, write_file, run_python, send_file
Web web_search across Tavily, Exa and Brave, web_fetch; browser_task is opt-in
Email send_email, only to an allowlist
Interfaces REST with streaming chat, an MCP server, a Telegram bot, a dashboard
MCP client tools of external MCP servers, off by default

Works withreal integrations

Telegram bot

MCP server

Pull requests

Claude Code jobs

run_python

Memory store

Containers

Model API

MCP server toolsnous_chat · nous_recall · nous_status · nous_teach · nous_decide

RESTPOST /chat · POST /chat/stream · GET /decisions:8000

DashboardMemory, the graph, decisions, the ledger, DAGs/dashboard

TelegramStreaming replies, attachments in and files outoptional

Companion appWeb and Android beta: approvals and live cards

Memory toolsrecord_decision · recall_deep · learn_fact · learn_skill

Nous works without being asked. Four ways it starts work.

The agent runs heartbeat checks

Health, Google Drive, email (opt-in) and checks you describe in plain words. Findings are deduplicated, escalated and summed up in a daily digest.

Heartbeatfindings

Disk on the build runner above 90%Health check · first seen todayescalated

Same alert from last weekMatched a known false positiveclosed

"Check the docs site for dead links"A check described in plain wordsno findings

Daily digestThree findings, one escalateddigest

DeduplicatedOne fingerprint per finding

The agent runs scheduled subtasks

Cron and one-shot schedules feed a background worker pool, so long work runs while the chat stays free.

schedule_task · spawn_taskworker pool

Weekly dependency reportCron · Mondays 09:00scheduled

Remind me to rotate the tokenOne-shot · Fridayscheduled

Summarise the build logsBackground subtaskrunning

Check broken linksBackground subtaskdone

ScoredBy the separate scorer job

Agent plans wait for approval

Multi-node plans with node timeouts and token accounting. A step can wait for your answer; that wait is opt-in.

dag_createparked

Step 4: publish the releaseWaits for the owner's answer

ParkedA durable plan; nothing runs while it waits

ResumesWhen the answer arrives

Token accountingPer node, per plan

Wall-clock reaperStops a plan that runs too long

The agent delegates to Claude Code

An opt-in shell held to an allowlist lets Nous launch Claude Code jobs in their own worktrees and follow their pull requests.

Coding jobopt-in

Pin Node 22 in CIOwn worktree, own branchPR open

ShellHeld to an allowlist

FollowedChecks and review on the pull request

RecordedIn the execution ledger

ReportedBack to the conversation

Nous checks every claim

After each turn, what Nous says it did is checked against the ledger of tool calls that really ran.

The Nous execution ledger

A durable record of every side-effecting tool call. The claim check after each turn reads it.

A background send refuses to run without its record. Sends are keyed, so a retry never sends a second copy. An undo journal keeps the before-state of the files, schedules and checks Nous changes.

Execution ledger · turn 14side effect

bash · npm testexit 0read-only

write_file · .nvmrcbefore-state keptlocal write

bash · git push origin pin-node-22exit 0external

bash · gh pr createexit 0external

Claim check

"I opened the pull request"in the ledger

"I emailed the team"No send_email call ran; the reply is correctedno record

Nous action tiers and a separate scorer

Tools are classed by side effect, and riskier tiers get stricter checks. Finished background subtasks are scored by a separate, deterministic job, built to run under its own database role.

Tier 1: read-only passes

Reading files, searching memory or the web: no gate.

read_filepass

recall_deeppass

web_fetchpass

bash · git statuspass

Tier 2: local writes

Files, facts and schedules get a consistency check first.

write_filechecked

learn_factchecked

schedule_taskchecked

Undo journalBefore-state kept

Tier 3: external or irreversible

Sends, pushes and deletes pass the full gate: a critic and a confidence check.

send_emailAllowlist onlyfull gate

git pushfull gate

Delete a bucketA censor asks firstblocked

ThenRecorded in the ledger