Claude Code & Cowork

Claude already knows Chutes.

Claude gets the toolkit's deepest integration: nine skills, OS-keychain credential storage, and a stdio MCP server. It can call Qwen, Google, Z.ai, DeepSeek, Moonshot AI, MiniMax AI, and Unsloth through Chutes' OpenAI-compatible API while model facts stay tied to the live catalog.
Claude install sourceCredential storeMCP serverClaude recipesConnect other agents

Quick config

The minimum to point this client at Chutes. Every value below is rendered from the live catalog or is a stable endpoint fact — copy it and go.

Install in Claude
text
/plugin marketplace add Veightor/chutes-agent-toolkit
/plugin install chutes-ai@chutes-agent-toolkit
Plugin version
chutes-ai v1.3.0
Toolkit commit
b2a58f7c459d
Skills
Nine skills across four product lanes
TEE catalog
13 listed models report confidential_compute=true, verified Jun 25, 2026.

60-second setup

01

Install the plugin

Claude command palette
text
/plugin marketplace add Veightor/chutes-agent-toolkit
/plugin install chutes-ai@chutes-agent-toolkit
02

Ask for setup

Ask: Set me up with a Chutes account and API key. Claude uses the hub skill and stores the cpk_ key outside the transcript.

03

Check credentials

No secrets printed
bash
python plugins/chutes-ai/skills/chutes-ai/scripts/manage_credentials.py check

Why Claude users should care

Four product lanes

Use Chutes, Build on Chutes, Operate on Chutes, and Run agents with Chutes are split into focused skills Claude can choose from plain-English requests.

Keychain storage

The credential manager checks macOS Keychain, Linux Secret Service, or encrypted-file fallback, while CHUTES_API_KEY stays available for CI/CD overrides.

MCP tools

The same repo ships a stdio MCP server. Read tools are verified; write and deploy tools keep their [BETA] labels.

TEE evidence

Claude can fetch and parse evidence for a confidential-compute chute. Shape-valid evidence is not the same as a completed DCAP cryptographic verdict.

The nine skills

Skill names and status labels mirror the toolkit. Click any card to open that skill's pinned source file.

Demo prompts for Claude

Prompt
"Set me up with a Chutes account and API key"
Claude uses
chutes-ai registers, creates the key, and stores it in the OS keychain.
Prompt
"Which Chutes model for tool-calling under $1/M?"
Claude uses
Live /v1/models plus the model picker filters.
Prompt
"Build me a failover pool of the 3 cheapest vision models"
Claude uses
chutes-routing emits an inline routing string or saves an alias.
Prompt
"Prove my model runs in a TEE"
Claude uses
chutes-tee fetches and parses TDX/GPU evidence; shape-valid is not cryptographic attestation.
Prompt
"How much did I spend this week and am I near my quota?"
Claude uses
chutes-usage-and-billing reads spend, balance, and quota endpoints.
Prompt
"Add Sign in with Chutes to my Next.js app"
Claude uses
chutes-sign-in [BETA].
Prompt
"Make Chutes available in Cursor via MCP"
Claude uses
chutes-mcp-portability; write tools remain [BETA].

Claude Desktop gets the same toolkit as MCP tools

Prefer tools over skills? Install the stdio MCP server from the toolkit. Read tools for models, quota, usage, aliases, chutes, discounts, and API keys are verified; mutating tools keep [BETA] labels.

Install MCP server
bash
uv tool install chutes-mcp-server \
  --from plugins/chutes-ai/skills/chutes-mcp-portability/mcp-server

Manual skill copy

If the marketplace flow is unavailable, copy the skill tree into the active Claude skills directory. The source remains the same plugin tree.

Direct install
bash
cp -r plugins/chutes-ai/skills/* ~/.claude/skills/

Privacy is on by default

The live model object exposes confidential_compute. Ask Claude to prove a model runs in a TEE and the toolkit can fetch TDX/GPU evidence for inspection. Keep the honesty boundary: a shape-valid parse is useful evidence, but it is not a cryptographic attestation verdict unless the optional DCAP validation path actually ran and passed.

Private inference guide

Troubleshooting

Symptom
200 from /v1/models, completions fail
Likely cause
/v1/models is public and did not validate the key.
Fix
Validate against an authenticated endpoint or ask Claude to check the keychain profile.
Symptom
Requests hit anonymous rate limits (429)
Likely cause
The client sent X-API-Key, which is silently ignored on inference.
Fix
Use Authorization: Bearer with the cpk_ key everywhere.
Symptom
default:latency or default:throughput rejected
Likely cause
Saved routing aliases need a pool configured once in the dashboard.
Fix
Configure the pool at chutes.ai/app Model Routing, or use concrete model IDs from /v1/models.
Symptom
Snippet model ID stopped working
Likely cause
Model IDs can churn with the catalog.
Fix
Refresh from GET https://llm.chutes.ai/v1/models and choose a current ID.
Symptom
chutes-mcp-server not found
Likely cause
The MCP server is not installed in PATH.
Fix
Run the uv tool install command from the toolkit repo.