Toolbelt for AI agents: 1DevTool gives Claude, Codex, and friends HTTP and database tools

You're debugging a flaky staging endpoint. You ask Claude to check whether /api/orders/42 returns the right shape. Claude can't actually call the endpoint — it's an LLM, not an HTTP client — so it writes you a curl command. You paste it into your terminal. The response comes back. You paste that into the chat. Claude reads it, asks for another endpoint to compare against. Same loop. You're the messenger.

Now imagine the same conversation about a Postgres query. "What's the row count on orders for last week?" Claude can't query the database either, so it writes you a SELECT. You paste it into the SQL editor. You paste the result back. Loop. Loop. Loop.

v1.18.6 is the release that hands AI agents the tools they kept asking you to use for them. There's a new request MCP tool that lets Claude, Codex, Gemini, and OpenCode fire HTTP requests through your active project's environment, and a new 1devtool-database MCP server that lets them query your configured database connections — without seeing the credentials, and without bypassing safe mode.

What changed

This release adds two MCP servers and a system tab that finally answers "is the right CLI even on my PATH?" Plus a sub-agent history viewer so you can see what nested agents are actually doing.

How it works in practice

HTTP requests, run by the agent through your environment

Every supported AI client — Claude, Codex, Gemini, OpenCode — now gets a request MCP tool. It takes a method, URL, optional headers and body, and runs the call through your active project's HTTP client setup: the bearer tokens, basic auth, base URL, and any project-level headers you already configured. It returns status, headers, body, and latency in one structured response.

So instead of being the messenger, you ask: "Hit /api/orders/42 on staging and tell me whether total matches sum(items[].subtotal)." The agent does it. It uses the staging environment, the staging bearer, the staging base URL — because that's what your active project is set to — and reports back.

There's also a run_saved_request tool that lets agents fire any saved request from your HTTP collection by name. "Re-run the staging health check" picks the right URL, environment, and auth without you having to spell out the URL. And list_environments and list_saved_requests let agents browse your collection so you can ask follow-ups like "which environment is this request in?" without pasting URLs every time.

Database queries, agent-issued but credential-safe

The other new MCP server is 1devtool-database. It exposes every database engine 1DevTool supports — Postgres, MySQL, MongoDB, Redis, Elasticsearch, and the rest — to AI agents through four tools: list_connections, describe_schema, preview_table, and query.

The important thing is what doesn't move: credentials never leave 1DevTool. Agents see connection names via list_connections, never connection strings. When they call query or preview_table, the call goes through 1DevTool's existing connection pool. Each connection's safeMode flag and the same read-only / write detection that the manual SQL editor uses still gate every agent-issued query. A read-only connection stays read-only no matter what the agent asks.

When an agent runs a query, the result streams into a new "AI query" tab next to your own queries in the Database panel, and the SQL is captured into history. You can audit exactly what the agent asked and what it saw.

Detected CLIs — finally, an answer to "is this on my PATH?"

The new Detected CLIs section under Settings → General → System scans every supported AI agent (Claude, Codex, Gemini, OpenCode, Amp, Qwen, Aider) plus runtimes, package managers, dev tools, and database CLIs in parallel. Each row reports a state: detected, override, ambiguous (multiple paths found — pick one), not-found, unverified, timeout, or error. You see the actual paths, the detected version, and how long the probe took.

A Cancel button appears after four seconds, so a flaky NFS share or a broken symlink in your PATH can't lock the dialog. Per-row Retry handles individual timeouts. Slow paths get quarantined automatically with an "Include anyway" escape hatch. And if you have a CLI 1DevTool doesn't know about, you can add it to the list with a name and binary — it shows up alongside the built-ins everywhere 1DevTool reaches for a CLI.

This is the answer to a question every multi-tool setup eventually faces: "Why doesn't this work on this machine?" You stop guessing.

Sub-agent badge and history viewer

Modern AI CLIs spawn sub-agents constantly — Claude fires off bash sub-agents and tool-call agents, OpenCode runs summarizer passes, Aider chains edits. Until now, you had no way to see what those nested agents were actually doing without opening their log files by hand.

A small sub-agent badge now appears on any running terminal whose CLI has spawned a nested agent. The badge shows which CLI is running underneath and how long it's been alive. Click the badge and a separate window live-tails the sub-agent's session log, color-coded per CLI: Claude amber, Codex green, Gemini violet, Amp pink, OpenCode blue, Qwen indigo, Aider red. The viewer parses JSONL, JSON, and markdown session formats the same way — assistant text, thinking, tool calls, tool results, system messages, and turn durations all show up with the labels each CLI actually uses.

Smaller things, all worth a line

OpenCode joins the MCP settings tab as a first-class client — one-click install / uninstall of channels, database, design, and http MCP servers into ~/.config/opencode/opencode.json, matching the existing flow for Claude, Codex, and Gemini.

Drag-and-drop from Finder / Explorer can now target any sub-folder at any depth in the file tree — it used to always land at the project root. Hovering a folder highlights it as the drop target, and dropping into a collapsed sub-folder auto-expands it so you immediately see what landed.

Opening a file from the file tree now auto-focuses the editor panel in whichever slot it lives — no more hidden-behind-the-Browser-tab moments.

The MCP settings tab splits Tools from External clients into two sub-tabs. The Tools sub-tab lists every MCP tool registered with 1DevTool — name, profile, description, output kind, mutates / long-running flags, enable state — so you can see exactly what's exposed to AI agents at a glance. The bridge health badge now reports how many tools are registered and the registry's revision, so a stale process or a bridge restart is immediately obvious.

The browser viewport in the Send-to-Terminal dialog is now tracked live, not snapshotted — it updates whenever you resize the panel, toggle the debug console, or switch device-emulation modes.

And terminal session recording (the MP4 export and share-menu flow) has been retired. Transcript copy-paste, the new Sub-Agent History viewer, and channel history cover the same ground with less footprint — recording will return in a different form if there's demand for it.

Before vs after

The old workflow	v1.18.6
AI agent writes a curl for you. You run it. You paste the response back.	Agent calls request directly. Response lands in the chat in one round-trip.
AI agent writes a SELECT. You paste it into the SQL editor. You paste results back.	Agent calls query against the right connection. Results stream into an audit tab.
Paste connection strings into chat so the agent can "help".	Agent sees connection names via list_connections — credentials never leave 1DevTool.
Read three blog posts to guess whether Claude is on this machine's PATH.	Detected CLIs scans every supported CLI with versions, paths, and per-row diagnostics.
Tail Claude's session log in a separate terminal to see what its sub-agents are doing.	Click the sub-agent badge → color-coded live viewer in a separate window.
Manually paste files into the project root, then move them with the file tree.	Drag-and-drop straight onto any sub-folder at any depth.

Who benefits most

Anyone who has ever been the messenger between an AI agent and a tool. The request and 1devtool-database MCP tools turn that loop into a single sentence to the agent.

Teams that share databases. Safe mode plus credentials-never-leave-the-app means the agent can investigate without you handing it the keys.

Multi-machine developers. The Detected CLIs section is the fastest way to answer "what's actually installed on this box?" without leaving the app.

Try it

Update to v1.18.6 and open Settings → MCP. Install the http and database MCP servers into whichever AI client you use. The next time you'd reach for the messenger pattern — "I'll paste it", "let me run that for you" — try asking the agent to do it directly. The clipboard hop disappears.