> ## Documentation Index
> Fetch the complete documentation index at: https://evolink.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.
# MCP authorization, approval, and troubleshooting
> Cost boundaries, account access, retention, and step-by-step troubleshooting
## MCP costs and approval
This page describes released MCP 1.6.1. Call `estimate_cost`, then show the model, final inputs, output count, and estimated cost. Call the generation tool only after explicit approval of this task.
See [models, costs, and quotas](/docs/en/cli-mcp/models-and-billing#pricing) for estimate states, public-pricing limitations, approval, and budgets. Released max\_cost\_usd only compares pre-submission estimates; it does not cap final charges.
estimate\_cost does not accept max\_cost\_usd; generation tools accept it.
Native MCP has neither a CLI local `quote_id` and `--confirm` command nor a session-level server approval button. The agent must retain the model, inputs, estimate, and approval for this task; changed inputs or prices require renewed approval. Host tool permission, account authorization, and purchase approval do not replace one another.
Model, balance, and task queries do not submit generation. Generation is billed; file-service limits apply separately. Failure, timeout, or closing chat does not establish that no charge occurred or that a refund completed. Usage queries are not complete settlement receipts.
See [shared access rules](/docs/en/cli-mcp/models-and-billing#authorization); this page retains sign-in and sign-out actions for this route.
## Authorization, permissions, and quotas
Remote MCP uses Passport browser OAuth. The host stores connection credentials, and the service uses the existing OAuth-linked Key. Remote MCP and CLI can share account access, Key quotas, and permissions. Pausing a shared Key affects all its consumers; host session sign-out is a separate operation.
Local stdio uses the personal API Key supplied to its MCP process and follows that Key's permissions and quotas. Do not paste Keys into chat or use CLI sign-in as proof of native MCP authorization.
Disconnect or sign out through the host's MCP or connector settings. Revoke account authorization through account authorization management when needed. Closing a window does not necessarily revoke access. Resolve insufficient balance, daily/total quotas, or model permissions through account settings.
## Storage and expiry
Default quote, reference, upload-authorization, and original URL lifetimes are maintained in [tasks, references, and results](/docs/en/cli-mcp/tasks-and-files#retention). Actual returned values apply; receipt support depends on file-service deployment.
## MCP troubleshooting: identify your connection
Remote MCP uses `https://mcp.evolink.ai/mcp`, Streamable HTTP, and host OAuth sign-in. Local stdio uses the host-configured `@evolinkai/mcp` process and a platform Key. Check credentials and tool discovery for the actual connection.
| Symptom | Troubleshooting section |
| - | - |
| Browser authorization fails or repeats | [Sign-in](#login) |
| Skill present but actual tools missing | [Skills and permissions](#skills-discovery) |
| MCP missing, text-only replies, or denied tools | [Connection and tool discovery](#mcp-connection) |
| Model missing, invalid inputs, or incomplete pricing | [Models and pricing](#models-and-pricing) |
| Generation timed out or outcome unknown | [Tasks and recovery](#tasks-and-recovery) |
| Attachment unreadable or upload failed | [Uploads](#uploads) |
| Original file 403, gray preview, or link-only result | [Downloads](#download) |
| Problem persists | [Support information](#support) |
Check that this host connection is enabled and authorized and that tools are visible, then actually call `check_balance`. Remote MCP exposes 15 tools; local stdio exposes 13. CLI success does not prove native MCP access, and MCP troubleshooting does not require CLI installation.
### Browser sign-in and authorization
Use sign-in or authorization in this MCP or connector entry. Find the host command or menu in its dedicated guide. Signing into a separate CLI does not authorize this connection.
Sign into the correct EvoLink account, check scopes, and approve. Keep the host authorization flow waiting. Do not share full authorization URLs or callback parameters. Restart a failed flow from the original host.
Refresh the connection's tools, check for `check_balance` and `search_models`, and actually query balance. Reauthorize after 401. For 403, also check account permissions, Key status, and host tool permissions.
Local stdio does not use this browser flow. Check the platform Key received by the process, startup environment, and stderr. Do not mix remote OAuth with personal Keys. See [other clients](/docs/en/mcp/other-clients).
### Skill installed but tools unavailable
1. Read the `evolink-mcp` tool instructions. Native MCP cannot execute terminal commands from the CLI skill.
2. Check actual server configuration and tools in the host. Reading a skill does not register a server, and plugin installation does not automatically complete OAuth.
3. Reload the session or plugin and check tool permissions. Instructions without a recorded tool call do not prove access.
4. If tools are listed but calls fail, inspect the tool error and original connection. Do not use paid generation as a connection test.
### MCP connection and access: missing tools, connection failure, 401/403
1. Check the connection type. Remote uses `https://mcp.evolink.ai/mcp`, Streamable HTTP, and OAuth; local stdio uses a launch command and personal Key. A documentation page, `llms.txt`, or REST endpoint is not an MCP server.
2. Enable EvoLink, complete this host’s browser authorization, then refresh or open a new session. Back up existing configuration and merge only EvoLink. Check JSON quotes/commas, YAML spaces, and configuration scope without overwriting other services.
3. Call `check_balance` and `search_models`. If tools list but calls fail, inspect authorization. If balance works but generation tools are absent, check tool toggles and workspace policies.
| Symptom | Action |
| - | - |
| Duplicate server name | Inspect the existing entry and choose one source; avoid manual/plugin duplicates |
| 401 / expired authorization | Reauthorize through the affected host, preserving task IDs |
| 403 / denied tool | Distinguish host policy, account/Key limits, and network rejection; record the failed stage rather than assuming a UA change fixes all 403s |
| Local process cannot start | Check Node 18+, npx path, host environment, and launch errors; reload and verify |
| Invalid/unauthorized local Key | Inspect its status and scope in the dashboard and update host secret settings, not chat |
| Configuration change has no effect | Refresh the process that owns the connection; a remote Gateway does not refresh with your laptop terminal |
Setup pages: [ChatGPT](/docs/en/mcp/chatgpt), [Claude](/docs/en/mcp/claude), [Claude Code](/docs/en/mcp/claude-code), [Codex](/docs/en/mcp/codex), [Cursor](/docs/en/mcp/cursor), [OpenClaw](/docs/en/mcp/openclaw), [Hermes](/docs/en/mcp/hermes-agent), [other clients](/docs/en/mcp/other-clients).
### Models and pricing: empty search, invalid parameters, incomplete estimate
1. Search live models and use the returned canonical ID. A family, display name, or alias may not be a valid submission ID. Broaden search wording without discarding required reference capabilities.
2. Inspect MCP `get_model` for this model. Check required fields, spelling, enums, numeric limits, and mutually exclusive references. Do not copy another model’s `duration`, `size`, or `quality` fields.
3. Re-estimate and inspect `input_valid`, `problems`, `estimate.status`, `basis`, and `warnings`. Input, price, or identity changes require a new estimate and approval.
Supply missing input for `needs_input`; `no_price` is unusable and `partial` is not a total. Unknown duration or reference multipliers cannot be filled with a starting price. Preserve the budget; do not silently remove caps, switch Keys, or reduce specifications to bypass a rejection.
Published public-default estimates are not account-specific authoritative Quotes, and `max_cost_usd` is not a final settlement cap. See [costs](/docs/en/mcp/billing) and [command/tool reference](/docs/en/mcp/tools).
### Timeout, running tasks, or unknown outcomes
Call `get_task` for an existing `task_id`; set `wait_seconds` to 0–45 for a bounded wait. Use `list_tasks` for multiple IDs. Timeout ends that wait only, not generation.
Preserve `client_request_id`, inputs, identity, and error when no task ID returns. Recover through the same tool and identifier; do not change accounts or identifiers and generate again. `list_tasks` accepts only its documented parameters, not invented filters.
Record actual errors and times, and query the original task and account usage. Failure does not prove no charge or a completed refund; unknown is not failed. There is no usable public queue-cancellation tool. Disconnecting MCP or closing a page does not cancel the task.
Changed inputs or prices require new approval for unsubmitted tasks. Recover submitted or unknown requests first. See [task tools](/docs/en/mcp/tools#task-progress-and-results) for parameters and examples.
### Unreadable attachments, failed uploads, or unknown outcomes
1. Confirm that the host can read the supplied file. Remote MCP cannot directly read a desktop path, and chat attachments do not automatically become model URLs.
2. Use `upload_file.file_url` for accessible URLs. Remote base64 is limited to 1 MiB. For larger files, call `prepare_upload`, upload from an environment with file access, then call `get_upload`. Single-use uploads allow up to 95 MiB; authorization lasts about 15 minutes and status about one hour.
3. Local stdio can use `upload_file.file_path` on the MCP process machine. Check path, permissions, format, file-service limits, and model limits; a desktop path is not automatically readable on a cloud host.
4. Preserve and query the original `upload_id`. Unknown does not mean nothing uploaded, and receipt compatibility depends on file-service deployment. Request a replacement after confirmed failure or expiry. References usually last 72 hours; re-upload expired inputs.
See [upload tools](/docs/en/mcp/tools) for parameters and the [reference workflow](/docs/en/mcp/overview#files) for full steps. Use an accessible URL when the host lacks file capabilities. Do not troubleshoot upload failures by regenerating.
### Downloads and previews
Query the original task, confirm completion, and use its original URL. Open originals in a browser or save from a file-capable environment; see [MCP delivery](/docs/en/mcp/overview#delivery).
HTTP/format checks, the Python product User-Agent example, expired links, and gray previews are maintained in [download troubleshooting](/docs/en/cli-mcp/tasks-and-files#download). Recover unknown tasks instead of regenerating or editing previews.
### Information for support
Follow the [support checklist](/docs/en/cli-mcp/tasks-and-files#support) with versions, failed stage, IDs, and redacted errors. Preserve diagnostics from this route; do not share Keys, tokens, private URLs, or environment dumps.
### Looking for CLI installation troubleshooting
CLI commands, secure storage, SSH callbacks, and skill discovery now live in [CLI troubleshooting](/docs/en/cli/billing#cli-install) and the [CLI setup guide](/docs/en/cli/setup). This entry preserves old-link compatibility. Use this page and your client guide for native MCP setup.
This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.