---
title: "DeepSeek integration with Bolta"
description: "Can you use Bolta with DeepSeek? Not in the chat app, but yes through a host agent like Claude Code running a DeepSeek model. See what is confirmed, the steps, and the limits."
canonical_url: "https://bolta.ai/integrations/deepseek"
markdown_url: "https://bolta.ai/integrations/deepseek.md"
last_updated: "2026-11-21"
content_type: "integration"
publisher: "Bolta"
---

# DeepSeek integration with Bolta

Source: https://bolta.ai/integrations/deepseek
Last updated: 2026-11-21

Can you use Bolta with DeepSeek? Not in the chat app, but yes through a host agent like Claude Code running a DeepSeek model. See what is confirmed, the steps, and the limits.

## Summary

Use a DeepSeek model as the drafting brain, with Bolta holding the approval gate

## Who this is for

- Developers
- Automation teams
- AI-assisted marketing teams

## Core capabilities

- Run a weekly batch: ask for five LinkedIn drafts and review them in Bolta on Friday
- Draft in bulk with a DeepSeek model, then edit and approve only the ones you like
- Check your approval inbox from the terminal without opening the dashboard
- Ask for a read of recent account analytics before planning next week's topics
- Turn release notes or a changelog pasted into Claude Code into a Bluesky and X draft pair
- Set up a recurring routine from a plan you read and confirm, instead of clicking through settings

## Limitations and boundaries

- Confirmed from DeepSeek's own docs: its API has an Anthropic-compatible endpoint at https://api.deepseek.com/anthropic, and a guide for pointing Claude Code at it with your DeepSeek API key.
- Confirmed from DeepSeek's own docs: that endpoint ignores mcp\_servers, so DeepSeek itself never holds the Bolta connection. The host agent does.
- Not found in DeepSeek's docs: any connector or MCP feature in the DeepSeek chat app, or a remote MCP setting in DeepSeek Harness.
- Not tested by us: how well Bolta's tools behave when the host agent runs a DeepSeek model. Draft quality and tool-call reliability can differ from other models, so review every draft.
- Integration availability depends on current APIs, authentication, scopes, and account permissions.
- Generated content can be incorrect and should be reviewed before publication.
- Availability depends on the current Bolta plan, connected network, account permissions, and integration coverage.
- Current prices and plan limits must be verified on the Bolta pricing page.

## Pick a host agent that supports MCP

Claude Code is the best documented option for DeepSeek. DeepSeek also publishes guides for Codex, OpenCode, OpenClaw and Hermes. This page uses Claude Code.

## Get a DeepSeek API key

Create the key on the DeepSeek platform. This key pays for the model only. It is not a Bolta key and Bolta never sees it.

## Point Claude Code at DeepSeek

As DeepSeek's Claude Code guide describes, set ANTHROPIC\_BASE\_URL to https://api.deepseek.com/anthropic and ANTHROPIC\_AUTH\_TOKEN to your DeepSeek API key. The guide also sets model-name variables such as ANTHROPIC\_MODEL, so copy the full list from the guide, because model names may change. Then start claude.

## Add Bolta to Claude Code

Run: claude mcp add --transport http bolta https://mcp.bolta.ai/mcp. If you would rather use the plugin, the Claude Code guide on the Bolta MCP docs page has the exact commands. Do not do both, or you will have two Bolta entries.

## Sign in to Bolta

In Claude Code run /mcp, pick bolta and finish the sign-in in your browser. Tick the brands this connection may reach. One connection is one organization, and a brand you add later needs a reconnect and a new tick.

## Optional: add the Bolta skill

Run npx skills add boltaai/bolta-skills --skill bolta in a terminal so drafts follow your voice and the approval rules.

## Test with a read

Ask: List my Bolta brands and connected accounts. It should name the brands you ticked and change nothing. If it lists the wrong brands, reconnect and tick the right ones.

## Ask for a first draft

Ask: Draft a LinkedIn post about our latest update and save it for my review. The draft appears in Bolta. Nothing goes out until you approve the exact text there.

## No browser on that machine? Use an API key

Prefer sign-in. If your host agent cannot open a sign-in window, you can create a Bolta API key (it starts with bolta\_sk\_) and send it as a Bearer header instead. We have not run that route end to end for this page. A pasted API key also skips the hold that puts sign-in posts on your review list as AI drafts, so keep the key out of shared files and still ask for drafts only.

## Why the DeepSeek API cannot hold the Bolta connection

People often search for deepseek mcp social media setups, so here is what DeepSeek's own docs say. DeepSeek's API takes tools that your own client defines. Its Anthropic-compatible endpoint, https://api.deepseek.com/anthropic, ignores the mcp\_servers field and does not support mcp\_tool\_use blocks. A remote MCP server such as Bolta's cannot be attached at the model. That is why the connection lives one layer up. The host agent keeps the MCP connection and calls Bolta's tools, and DeepSeek is only the model reading and writing text. You sign in to Bolta in the agent, not in DeepSeek.

## Why try a DeepSeek model, and what stays the same

One reason people try a DeepSeek model for drafting and weekly batches is model cost. We have not benchmarked it, so we make no claim about price or quality. Whichever model writes the text, Bolta's rule does not change. Drafts save freely. Approving, scheduling and publishing need your yes, and Bolta refuses or sends the post to review until you give it. A draft you do not like costs you an edit, not a bad post. Review every draft more closely than you would with a model you have already tested. Keep tasks small at first, such as one draft or one read. Whatever the agent reads from Bolta, such as brand notes and drafts, is sent to DeepSeek's API as part of the model request. Check DeepSeek's terms if that matters for client work. If tool calls fail often, switch the host agent back to its default model and compare.

## DeepSeek Harness: what we have and have not confirmed

DeepSeek Harness is an MIT-licensed project from DeepSeek, started with npx @deepseek-ai/dsh web and built from plugins. Confirmed: the project and its docs exist, and it uses a plugin system. Not found: a remote MCP setting or an official MCP client plugin in the docs we read. Not tested: third-party plugins that claim MCP support, with https://mcp.bolta.ai/mcp. Until someone tests it, we do not promise a Harness setup. Use a host agent instead.

## Which host agent to run with your DeepSeek key

Host agent: Claude Code; DeepSeek guide: Yes, the best documented; Bolta page: /integrations/claude-code | Host agent: Codex; DeepSeek guide: Yes; Bolta page: /integrations/codex | Host agent: OpenClaw; DeepSeek guide: Yes; Bolta page: /integrations/openclaw | Host agent: Hermes Agent; DeepSeek guide: Yes; Bolta page: /integrations/hermes-agent | Host agent: OpenCode; DeepSeek guide: Yes; Bolta page: Use the MCP guide for other apps

## Chat app, API or Harness: what each can do for Bolta

DeepSeek is several things under one name, and only one of them gives Bolta somewhere to connect. Route: DeepSeek chat app; Can hold the Bolta connection: Not found; Why: No connector or MCP setting found in the docs we read | Route: DeepSeek API; Can hold the Bolta connection: No; Why: The Anthropic-compatible endpoint ignores mcp\_servers | Route: DeepSeek Harness; Can hold the Bolta connection: Not confirmed; Why: Plugin system documented, no remote MCP setting found | Route: Host agent on a DeepSeek key; Can hold the Bolta connection: Yes, on the agent; Why: The agent keeps the MCP link and DeepSeek is only the model

## Model names change, so copy them from DeepSeek

DeepSeek's Claude Code guide maps its own models onto the model-name variables Claude Code expects. We list ANTHROPIC\_BASE\_URL and ANTHROPIC\_AUTH\_TOKEN on this page because they are stable, and leave the model-name values to the guide, since those may be renamed. If a Bolta call starts failing after a DeepSeek model update, check the guide first. A changed name is a cheaper fix than a changed setup.

## What DeepSeek sees when the agent works with Bolta

The model only sees what the agent puts into the request. If you ask for a draft, that can include your voice notes, brand facts, earlier drafts and the text of the request. Your social network logins stay with Bolta, and your DeepSeek key is never given to Bolta. We do not know DeepSeek's retention terms, so for client work read them before you route a brand's notes through the model.

## Can I connect Bolta to the DeepSeek chat app?

We found no connector or MCP feature in the DeepSeek chat app, so we cannot show a way to do it. The route that works on paper is a host agent such as Claude Code running a DeepSeek model, with Bolta connected to that agent.

## Does the DeepSeek API support MCP servers?

Not directly. DeepSeek's docs say its Anthropic-compatible API ignores mcp\_servers. The MCP connection has to be held by the agent that calls the API.

## Can I use DeepSeek models with Bolta through Claude Code?

DeepSeek documents how to point Claude Code at its API, and Claude Code can connect to Bolta. We have not tested how well Bolta's tools behave with a DeepSeek model, so start with small tasks and review the results.

## What is DeepSeek Harness and does it work with Bolta?

It is a plugin-based developer preview from DeepSeek. Its docs that we read do not mention remote MCP, so we do not confirm a Bolta setup there. Use a host agent for now.

## Will DeepSeek post without my approval?

It should not. The model only writes text and asks the agent for tools. Bolta is the one that waits for your yes before anything is approved, scheduled or published, and it refuses actions your role does not allow. Tool-call behavior on a DeepSeek model is untested here, so read each draft before you say yes.

## Do I need a DeepSeek API key or a Bolta API key?

For this route you need the DeepSeek key to run the model. For Bolta, sign in through your host agent. Only use a Bolta API key, which starts with bolta\_sk\_, if the agent cannot open a browser. We have not run that route for DeepSeek, and a pasted key bypasses the hold that sign-in drafts get, so ask for drafts only.

## Which brands can DeepSeek reach through Bolta?

Only the brands you tick at sign-in, and only with the permissions of your Bolta role. One connection is one organization. A brand you add later needs a reconnect and a new tick.

## Which social networks can I publish to?

Through Bolta you can publish to LinkedIn, X, Instagram, Facebook, Threads and Bluesky, once the account is connected in Bolta. DeepSeek does not post anywhere itself.

## Is the DeepSeek route free?

We do not set DeepSeek's prices, so check the DeepSeek platform. Bolta's own plans are on the pricing page. Using your key means DeepSeek bills you for model use separately.

## Does this work for client work with several brands?

Yes in principle, since you tick every brand the connection may reach and tell the agent which brand to use. Because tool-call quality on DeepSeek is untested here, name the brand in each request and check the draft shows the right one.

## Will drafts be as good as with another model?

We have not compared them, so we make no claim either way. Bolta gives the model your voice context, but the writing is still the model's. Review the first few drafts closely and keep what you would have written yourself.

## Can DeepSeek set up a posting routine for me?

It can ask Bolta to plan one, and Bolta shows you the plan first. The routine is saved switched off and only turns on after a second yes from you. We have not tested that flow on a DeepSeek model, so read the plan carefully.

## Claude Code still answers like its default model

The environment variables were probably set in a different terminal. Set the variables from DeepSeek's Claude Code guide, including ANTHROPIC\_BASE\_URL and ANTHROPIC\_AUTH\_TOKEN, in the same shell you start claude from, then start it again.

## I get an authentication error from the model

ANTHROPIC\_AUTH\_TOKEN must be your DeepSeek API key, not a Bolta key or an Anthropic key. Check it on the DeepSeek platform and that the key has not been deleted.

## Bolta does not appear in /mcp

Run the claude mcp add command again from the same project or user scope you are using, then restart Claude Code. If you also installed the plugin, you may have two entries, so keep one.

## The agent names the wrong brands

Brands are chosen at sign-in. Reconnect through /mcp and tick the brands you want, then ask again and name the brand in the request.

## Bolta tool calls fail or come back malformed

This is the part we have not tested. Retry once with a simple read, then try the same request on the agent's default model. If only the DeepSeek run fails, use it for drafting text and keep Bolta steps on the other model.

## DeepSeek is not Claude Code, even when they run together

In this setup Claude Code is the app and DeepSeek is the model behind it. The Bolta connection, sign-in and tools belong to Claude Code. If you simply want the app, read the Claude Code page.

## Relevant links

- [Related page: /integrations](https://bolta.ai/integrations)
- [Related page: /integrations/claude-code](https://bolta.ai/integrations/claude-code)
- [Related page: /integrations/codex](https://bolta.ai/integrations/codex)
- [Related page: /integrations/openclaw](https://bolta.ai/integrations/openclaw)
- [Related page: /integrations/hermes-agent](https://bolta.ai/integrations/hermes-agent)
- [Related page: /integrations/mcp](https://bolta.ai/integrations/mcp)
- [Bolta pricing](https://bolta.ai/pricing)

## Supporting sources

- [DeepSeek official website](https://www.deepseek.com/)
