Blog

Latest news, tutorials, and updates from Octomind

RSS feed
The Octomind Tap registry — agent domains radiating from one command, from developer to family to finance

The Tap Registry Grew Beyond Code: 108 Agents, 27 Domains, and Workflows That Verify Their Own Work

This week the default Octomind registry expanded from a developer toolbox into a full ecosystem: new life-domain agents from parenting to tax, four new self-verifying workflows, a project knowledge system, and container-proof dependency installs — all one command away, no installation.

A terminal walkthrough — curl install, device login, and an agent fixing a bug through the Octomind Hub

Your First Five Minutes with Octomind: One Install, One Login, Zero API Keys

A transcript-backed walkthrough of getting started with Octomind and the Hub: one curl install, one browser login, and an agent reviewing and fixing real code — every command run today in a clean Ubuntu container, costs included.

The Octomind panel showing cloud machines and the model hub — one key, one wallet, transparent prices

Octomind Cloud Is Live: Real Machines for Your Agent, Every Model Through One Key

Your coding agent dies with your laptop lid, your API keys live in five dashboards, and your AI bill is a surprise. We built Octomind Cloud to fix all three: persistent machines from $0.05/hr — with real Docker inside — a hub of 21 models behind one key at published per-token prices, and one wallet with caps you can see.

An AI agent constrained by layered controls — scoped tool permissions, a filesystem sandbox, and guardrail rules

Give Your AI Agent Less Power: Roles, Sandbox, and Guardrails That Hold

Prompts are not a security boundary. How to constrain an AI coding agent in Octomind with scoped tool permissions, filesystem sandboxing, and project guardrails that block dangerous commands before they run.

Octomind 0.38.0 — a device-authorization login flow signs the CLI in through the browser and stores a per-machine hub key plus a refreshing panel session, a new /usage command draws spend and quota bars across the CLI, WebSocket and ACP lanes, a user-scope .env lets one credential serve every project, and two speculative supervisor features (stale tool result masking and mid-turn plan recitation) are removed after an A/B showed they raised cost per solved instance

Octomind 0.38.0: One Command to Sign In, and Two Features We Deleted

Octomind 0.38.0 adds `octomind login` — an RFC 8628 device flow that signs your terminal in to an Octomind account without ever typing a password there — plus a `/usage` command that shows account spend and quotas in the CLI, WebSocket, and ACP lanes. It also deletes two supervisor features we shipped three weeks ago, because the benchmark said they cost more than they saved: 70% solved at $3.28/solved, both metrics better after the deletion.

An orchestrating AI agent delegating subtasks to specialized context, review, and test-writing sub-agents

Stop Asking One AI Agent to Do Everything. A Practical Guide to Sub-Agents

One agent struggles with big tasks — it loses context and produces half-finished work. How to split work across specialized sub-agents in Octomind: static agents, tap roles, dynamic agents, and parallel async delegation.

Octomind 0.37.1 — the Supervisor verify-gate gains ground-truth verification: it checks a completion claim against the runtime git diff of changed files and the last command's recorded output (which the agent cannot edit), replaces name-based mutation tracking with observational working-tree fingerprinting so a verification only counts when it ran on an unchanged tree, extracts and recites the request's explicit prohibitions, adds a free deterministic file:line reference check and a plan-completion pre-gate that refuses done while the live checklist has open items, breaks out the compression model's own spend in /info, and freezes a new 15-instance k=4 SWE-bench-Live baseline (70% solved, $3.39/solved)

Octomind 0.37.1: Ground Truth Over the Agent's Story

Octomind 0.37.1 hardens the Supervisor's verify-gate: it now judges a "done" claim against the actual working-tree diff and the actual command output — not the agent's narrative — and the check-after-mutation gate stops classifying tools by name: it fingerprints the working tree itself, so a change made through any tool only clears after a real check on an unchanged tree. Plus two free deterministic pre-gates (plan-complete, file:line references), recited prohibitions, compression spend in /info, and a new 15-instance benchmark baseline.

Octomind 0.36.0 — project instructions move to the AGENTS.md open standard (auto-loaded, zero config, custom instructions/constraints files removed), the WebSocket API gains request_id correlation with ack frames and input hardening, empty provider completions now fail loudly instead of stranding the user

Octomind 0.36.0: Your Project Speaks AGENTS.md Now

Octomind 0.36.0 adopts the AGENTS.md standard for project instructions and hardens the WebSocket API with request correlation, acknowledgments, and a fix for the stuck-on-working hang.

A developer navigating an unfamiliar codebase as a knowledge graph of imports, calls, and dependencies

Dropped Into a 200K-Line Codebase? Get Oriented in an Afternoon with Octocode

How to use the Octocode MCP server — semantic search, signatures, and a GraphRAG knowledge graph — to understand an unfamiliar codebase fast. A practical onboarding workflow for the project nobody has time to explain.

Octomind 0.35.0 — the Supervisor gains a Condenser: task-aware, research-grounded narrowing of oversized tool outputs by verbatim line-range selection (never summarization), conditioned on a cached distilled agent profile; benchmarked on SWE-bench-Live: success 87%→93%, cost per solved task down 34%

Octomind 0.35.0: Your Agent Reads Too Much

Octomind 0.35.0 ships the Condenser — a task-aware filter that decides what the agent actually needs to see from every oversized tool output. Same model, success up 87%→93%, cost per solved task down 34%.

A cost dashboard showing AI token spending dropping after tuning Octomind compression and capability settings

Your AI Coding Bill Is Mostly Waste. The Config That Cuts It.

A practical guide to controlling AI agent costs in Octomind — automatic capability loading, cache-aware compression, a cheap decision model, and spending limits. These settings quietly cut a token bill without changing your model.

A terminal connecting the Octocode MCP server to Claude Code, Cursor, and Windsurf for semantic code search

Add Semantic Code Search to Claude Code, Cursor, and Windsurf in 5 Minutes

A step-by-step guide to wiring Octocode into your AI assistant as an MCP server — install, index, and connect Claude Code, Cursor, or Windsurf so your agent finds code by meaning instead of grepping blindly.