Baseten Skills

Agent DX bundle — baseten skill tuned for Baseten backend MCP, Docs MCP and CLI.

The MCP makes token usage and wall time more efficient - our evals (below) show that, while agents can still achieve goals with raw REST API usage with similar pass rate. Additionally, the MCP tool annotations allow agent harnesses formal gating of destructive operations, providing additional safeguards.

What you can do without leaving the chat:

• Debug live: "Why do I see this log line" "Fix my deploy" → agent pulls logs, finds stack trace, proposes fix.
• Operate: Promote dev → prod, bump autoscaling for traffic spike, run a test predict.
• Keep the overview: "What's deployed, healthy, cold?" One-shot status across your account, easy cleanups.
• Skip the doc dive: Agent gets pointers to Baseten docs, blogposts and more in context.
• Wire up automations: Plug it into your own agents or internal tools for reactive ops without glue code.
• Install once, works everywhere: npx add-mcp, your API key, done. Uniform setup across 14+ coding agents.
• Read-only by default, mutations gated via harness policy check.

Set Up

Requirements:

• For interacting with your Baseten workspace, provide an API key with management permissions (you can get it from the webapp). We recommend using a purpose-dedicated key, so it can be independently revoked without impacting other workstreams.
• Node >= 18 (for the install tools)

Installation

Agent-driven (recommended)

Set your key, then paste this into your agent:

Linux/Mac:

export BASETEN_MCP_KEY=...

Windows (PowerShell or cmd.exe):

$env:BASETEN_MCP_KEY = "..."   # PowerShell
set BASETEN_MCP_KEY=...         # cmd.exe

Then paste this into the agent of your choice:

Install the Baseten agent toolkit following instructions from `github.com/basetenlabs/baseten-skills`.

- use `npx skills add` and `npx add-mcp`
- all global and for all agents (`-g -y`)
- The baseten backend server needs auth header `Authorization: Bearer $BASETEN_MCP_KEY` (or `Authorization: Bearer $env:BASETEN_MCP_KEY` in PowerShell). Run the install commands in a shell that has `BASETEN_MCP_KEY` set so the env var expands; don't read or inline the key value.

Manual

Linux/Mac:

export BASETEN_MCP_KEY=...
npx skills add basetenlabs/baseten-skills -g -y
npx add-mcp https://api.baseten.co/mcp -g -y --header "Authorization: Bearer ${BASETEN_MCP_KEY}"
npx add-mcp https://docs.baseten.co/mcp -n "baseten_docs" -g -y

Windows:

$env:BASETEN_MCP_KEY = "..."
npx skills add basetenlabs/baseten-skills -g -y
npx add-mcp https://api.baseten.co/mcp -g -y --header "Authorization: Bearer $env:BASETEN_MCP_KEY"
npx add-mcp https://docs.baseten.co/mcp -n "baseten_docs" -g -y

• -g global, -y auto-confirms all detected harnesses.
• Harnesses that support env-var interpolation: point the MCP config at $BASETEN_MCP_KEY instead of baking the key in.
• Some agents prompt before reading skill reference files (they live outside your workspace). In Claude Code, pre-approve via settings.json: "permissions": { "allow": ["Read(~/.claude/skills/**)"] }.

The truss CLI is separate, needed for deployment authoring; see CLI docs:

uv tool install truss

Components (skill, 2x MCPs, CLI) can be installed selectively, but work best in combination.

Getting started & Usage

After installation, most agents require a restart.

Check if the MCP servers connect with /mcp or /mcps (if not connected, verify the BASETEN_MCP_KEY in the harness config file).

You can start asking any questions or tasks related to Baseten, from chatting about the docs, to brainstorming solution approaches, deploying and iterating on models or managing your workspace. Most agents trigger the skill as needed automatically; alternatively you can invoke it with /baseten.

Evaluation results

We measured the baseten skill against the bare Claude Opus 4.7 baseline across 16 tasks spanning model authoring, integration, operate, debug, and tune workflows. Five configurations × 4 runs × 16 evals = 320 runs.

Configuration	Pass rate	Wall (s)	Cost ($)
Naked model (no skill, no MCP, no docs)	0.89	107	0.56
+ docs MCP	0.85	110	0.66
+ docs MCP + skill	0.87	136	0.73
+ docs MCP + baseten MCP	0.91	99	0.54
+ docs MCP + baseten MCP + skill (full kit)	0.97	99	0.55

Highlights (95% CIs from cluster bootstrap over evals):

• Full kit lifts pass rate from 0.89 to 0.97 vs. naked Opus 4.7 (Δ +0.08, CI excludes 0). Quality gains compound when skill and MCP are paired: adding either on top of the other is significant on its own.
• The baseten MCP cuts wall and cost roughly in half on backend-heavy tasks with no quality cost. On operate tasks (promote, autoscale, status), wall drops from 124s → 53s and cost from $0.82 → $0.35 when MCP is added to a skill-loaded agent. Similar magnitudes on debug and tune.
• Opus has strong baseline Baseten knowledge — most authoring tasks pass without the toolkit. The toolkit's measurable value concentrates on tasks that need live workspace state (operate, debug, tune).

Full methodology, marginal effects across all metrics, per-eval breakdowns, and per-group analysis: Full eval report.