DeepSeek Added in 0.9.17

The DeepSeek provider runs DeepSeek's models with your own DeepSeek API key. DeepSeek has no agent of its own, so OAR runs Claude Code on the relay's machine and points it at DeepSeek's Anthropic-compatible API. The agent works in the conversation's folder with Claude Code's tools; DeepSeek's model does the thinking. Turns are billed to the DeepSeek account the key belongs to.

Because the agent is Claude Code, a DeepSeek chat has most of what a Claude chat has: all four relay modes, question cards, the plan board, sub-agents, steering and background tasks. The table of all providers is on Providers at a glance.

The API key #

You need a DeepSeek API key and nothing else: there is no CLI to install, and no Claude account is involved. Open Settings → Providers → DeepSeek, the last sub-tab.

The line under the fields says where you are:

Then start a conversation as usual: in New Chat, pick DeepSeek. See Starting a conversation.

Where the key is kept, and who sees it #

• The key is stored in the relay's database on the relay's machine. The panel never shows it again, and it is never sent to a browser: the settings only say that a key is saved.
• It reaches only the workers of DeepSeek chats. Where sessions run under tmux, it goes to them through a file only the relay's user can read. Inside the worker it becomes Claude Code's token for DeepSeek's endpoint.
• It is sent to the base URL and nowhere else: to DeepSeek's API, or to the address under Advanced. When the relay's own requests to DeepSeek (the model list, the balance) fail, the error text has the key taken out.
• It cannot be set through the environment: a DEEPSEEK_API_KEY in the relay's own environment is not passed on, the workers get the panel's key.
• The agent can read it. Claude Code holds the key in its environment, so the commands the agent runs can see it, as with any key a CLI holds. The relay replaces the key with [redacted] in everything a session worker posts (replies, thoughts, tool activity, sub-agent text, questions, task output) before it stores, shows or logs it, and the DeepSeek worker does the same in its own log. Where a text was cut, the part of the key in it goes too, from 8 characters on. A key you changed or removed stays redacted until the relay restarts. Task output the relay reads from a task's file is cleaned by the shape of keys and long tokens, not by the saved key's own text. This keeps the key out of what OAR stores and shows. It is no barrier against the agent, which can change the text before printing it, call the relay's local API itself or read the relay's database.

Saving, changing and removing the key #

• The key must be 8 to 512 characters, without spaces. Otherwise saving fails with The DeepSeek API key has an invalid format.
• Saving a key while DeepSeek is on asks DeepSeek for its models (see Models). If that fails, the notice says why, and the list falls back as described there.
• Saving a different key clears the discovered models and stops the DeepSeek workers that are running, including one in the middle of a turn, so change the key between turns. The next message starts them with the new key.
• Removing the key is refused while any DeepSeek conversation has sent a message or has one queued: Cannot remove the DeepSeek API key while N DeepSeek conversation(s) still exist. Switch DeepSeek off for new conversations instead. When it goes through, conversations that have not sent a message yet move back to Copilot. Switching DeepSeek off moves nothing; it only takes DeepSeek out of New Chat.

Models #

Where the list comes from. The relay asks DeepSeek's GET /models with your key and keeps the IDs that start with deepseek-. It asks when you save a key, when you switch DeepSeek on again, when you change the base URL, and on Refresh in 🤗 Select Models. A request gives up after 10 seconds. Until a list has been read, the pickers offer deepseek-flash and deepseek-v4-pro. A failed read, or one that lists no DeepSeek model, keeps the list the relay had; after Refresh the notice reads DeepSeek model discovery failed. Cached DeepSeek models were kept: ….

Which ones show. 🤗 Select Models → DeepSeek chooses which models the composer offers; 💾 Save enabled models keeps the choice. With no choice made, all of them are offered. The Default model is always offered and cannot be unticked.

Per message. A conversation can change its model between turns. A switch restarts the conversation's Claude Code with the new model before the next message; the conversation goes on where it was. A model the relay does not offer is refused with DeepSeek model "…" is not available.

Images. DeepSeek V4 Pro takes no image input, and neither does a model OAR does not know. Such a model gets every attached image as a file path instead. Once a conversation holds images, a switch to such a model is refused, because the history it resumes would fail the request: DeepSeek V4 Pro takes no images, and this conversation already has some. Stay on DeepSeek Flash, or start a new DeepSeek conversation for DeepSeek V4 Pro.

Reasoning effort #

DeepSeek's models always think. The effort picker offers how much: low, high or max, the same three for every model. high is the default.

There is no none: Claude Code never asks DeepSeek to switch thinking off, so OAR does not offer it, and the 🧠 dialog has no Thinking controls for DeepSeek.

The effort is set when Claude Code starts. A different effort therefore restarts the conversation's Claude Code between turns, like a model switch, and the conversation goes on where it was.

An effort that comes from elsewhere, for example from an agent on another relay, is mapped to DeepSeek's levels: none and minimal run as low, medium and xhigh as high, ultra and ultracode as max. A message without an effort, or with a name DeepSeek has no level for, keeps the conversation's effort. A new conversation without an effort starts at high; one created with a name DeepSeek has no level for starts at low.

The folder and the session #

A DeepSeek chat runs the same worker as a Claude chat, built on the Claude Agent SDK, on the relay's machine: one per conversation, in the conversation's working folder. Its Claude Code is the build bundled with the Agent SDK, or the claude binary the relay uses for Claude chats when one is set (CLAUDE_CODE_EXECUTABLE, which a CLI install from the Claude sub-tab sets). Claude Code is pointed at <base URL>/anthropic, DeepSeek's Anthropic-compatible API. The first turn of a worker shows the activity line Starting Claude Code for DeepSeek (cold start)….

Its own configuration folder. Claude Code in a DeepSeek chat keeps its state, sessions and transcripts in deepseek-claude in the relay's data folder, readable by the relay's user only. It never uses ~/.claude (or CLAUDE_CONFIG_DIR): the machine's Claude login, your global settings, skills, plugins and MCP servers are not in reach. A repository's own CLAUDE.md and .claude/ folder still apply, except that they cannot reroute DeepSeek's traffic.

No Claude login. The worker hands Claude Code the DeepSeek key and nothing of the machine's Claude configuration. A Claude login token in the relay's environment does not reach it. OAR also switches off Claude Code's own traffic to Anthropic, such as update checks and telemetry.

What works as on Claude. These come from the same runner as Claude chats and behave as described on Claude:

• all four relay modes; plan ends with the Plan ready for review board;
• tools on the relay's machine, question cards with multi-select, the sub-agent viewer, background tasks in the task panel, Stop, and steering (the send button reads Steer during a turn);
• OAR's preview and relay_tool agent tools;
• the Session root button in the file browser, which opens the session's folder under deepseek-claude/projects;
• the background command limit from the Claude sub-tab;
• resuming after a restart, from the stored session ID. Deleting a DeepSeek conversation deletes its Claude Code session in deepseek-claude; the spend it added stays in the usage card's totals.

Sub-agents stay on the chat's model. Claude Code names its models by aliases. OAR pins them: a sub-agent the agent asks for as "opus", "sonnet" or "fable", or one without a model, runs on the conversation's model; one asked for as "haiku" runs on DeepSeek Flash. Claude Code's own small calls, such as titles and summaries, also go to DeepSeek Flash. A sub-agent's badge names the DeepSeek model. A model switch restarts Claude Code so the pins follow it.

Context and compaction. DeepSeek's models have a window of 1,048,576 tokens. The composer's Context window chip shows that one window, as 1.05M; there is nothing to choose. The 🧠 dialog shows the context as Claude Code reports it. A DeepSeek chat compacts at 400K tokens unless you set another window with the dialog's auto-compact slider: far below the window, so a turn after DeepSeek's cache has expired does not send most of a million tokens again at the cache-miss price.

Attachments. On a model that takes images, images up to 5 MB go inside the message; larger images and other files are passed as paths for the agent to read. See Models for the models that take no images.

Not on DeepSeek. Claude's plan meters and the usage-limit pause do not apply. DeepSeek replies have no usage line; the cost is on the usage card.

Folder settings cannot reroute DeepSeek #

A .claude/settings.json or .claude/settings.local.json in the folder can set environment variables and a key helper for Claude Code. In a DeepSeek chat the relay pins the base URL and the key over them, and with them the proxy, TLS trust, provider and credential settings: the folder's settings cannot point Claude Code at another address, swap the key for another one or route the requests through a proxy of their own. The key is not put on the command line for this.

Added in 0.9.18 No hooks. Claude Code in a DeepSeek chat starts with all hooks and the status line off: neither those of DeepSeek's own settings folder nor the chat folder's (.claude/settings.json, settings.local.json) nor a plugin's run. They would run as the relay's user with the key in their environment, and a folder's SessionStart hook even before the relay had pinned the endpoint and key. The setting is on Claude Code's command line, so a settings file cannot turn them back on. Hooks in managed settings, the machine administrator's, still run. In 0.9.17 those hooks ran; going back to 0.9.17 turns them on again.

If Claude Code does not take the pinned settings within a minute, the turn fails before anything is sent, with Claude Code did not take the DeepSeek connection settings before the first message, so nothing was sent. Send the message again; if it fails again, update the relay and look in its log for the reason.

What goes to DeepSeek #

Everything the agent reads and writes during a turn goes to DeepSeek: your messages, the files it opens, the output of its commands, and the context Claude Code adds to every session, such as the folder's path and the CLAUDE.md files. During a measured turn, DeepSeek's API was the only host Claude Code talked to.

Commit attribution #

There is no Vanilla: Claude Code's own lines would credit Claude for DeepSeek's work. A folder's own choice, set in the 🧠 dialog of a Claude conversation in that folder, applies to DeepSeek chats there too, with Vanilla read as OAR. A DeepSeek chat's own 🧠 dialog has no attribution select.

Another endpoint #

Advanced → Base URL takes another address, for a proxy for example. Leave it empty for https://api.deepseek.com. The address must use https, or http on this machine (localhost, 127.0.0.1 or ::1), and must not contain a user name, password, query or fragment. The key is sent there and nowhere else. Saving a new address clears the discovered models and reads them again from it. A DeepSeek worker that is already running keeps the address it started with.

WebSearch and WebFetch #

Claude Code has two web tools, and they run in different places:

The usage card counts the searches on a Search row of their own. The count comes from Claude Code's own totals, which count each search its WebSearch tool makes.

The usage card #

📊 Check Usage has a DeepSeek card while DeepSeek is switched on. DeepSeek is pay-as-you-go, so the card shows money: the account balance, read live with the saved key, and what this relay's DeepSeek turns have cost, estimated from their tokens at DeepSeek's prices (see How the estimate is made).

The spend is shown in the currency of the balance (the one with money on it), or in US dollars when there is no balance to go by. The card shows everything at once, figures first, and fits a phone's width. From top to bottom:

The bars change colour as they fill and turn red at 90 percent, as on every card; see Reading a card. The foot links to DeepSeek usage and Top up on DeepSeek's platform.

Once you have imported DeepSeek's usage export, the card shows DeepSeek's own figures on the days the export covers, and the labels read Spend, actual and estimated and Spend by model, actual and estimated. See Import usage export.

The charts. They are drawn without a chart library and work in both colour schemes. On Daily spend a model keeps its colour from one read to the next: DeepSeek Flash first, then DeepSeek V4 Pro, then other models by name; with more than four models, the fourth and later share Other models. Every column and bar has a text equivalent for screen readers and a tooltip on hover, focus or tap, and each chart has an As a table fold with the same figures. Daily spend is left out when there was no spend in the last 30 days, and its note says when the daily history on this relay starts, if that is later. The cache saving is the cache hits at the cache-miss price, less what they cost.

When the balance cannot be read. A read gives up after 5 seconds. An answer is kept for a minute, a failure for 15 seconds. When a read fails, the card is marked Partial, says The balance could not be read: …, and shows the last balance it read, as read with its time. When DeepSeek reports the balance as too low for API calls, the card says DeepSeek reports this balance as not enough for API calls: DeepSeek turns fail until it is topped up. with a Top up link.

An agent can read the same card with relay_tool's usage action.

Monthly budget #

Monthly budget (optional) on the DeepSeek sub-tab adds a budget bar to the card: this relay's estimated DeepSeek spend of the calendar month against the amount.

• Budget currency: USD ($) (preselected for a new budget), CNY (¥), or The balance's currency (may change). The bar is labelled with its currency, for example Monthly budget (CNY), and counts the spend in that currency, whatever the balance is in. A budget saved before the currency could be chosen follows the balance's currency.
• The bar is marked estimated. With days of an imported usage export in the month (and the budget in the card's currency), it is not, and its note says the spend is actual on those days and estimated on the others. It changes colour at 50 and 75 percent of the budget and turns red at 90.
• It counts by UTC day and starts over on the 1st of the month (UTC); the bar shows when it resets.
• Leave the amount empty, or enter 0, for no bar. Amounts keep two decimals.

Starting balance and top-ups #

DeepSeek tells OAR the balance, not what you paid in. The balance bar therefore needs something to measure against, and OAR keeps that per currency on the relay:

• Without a starting balance, the bar measures from the balance right after its last top-up. Until the relay has seen a top-up, it measures from the highest balance it has read, so the bar starts empty and fills with the spend. The line under the bar names the amount and the day it was read, for example Measured from ¥110.00, the highest balance seen (2026-10-20).
• A top-up (or a grant) is any rise of the balance between two reads. The relay reads the balance when the card is opened (or an agent reads the card), so it notices a top-up at the next read. From then on the bar measures from the balance after it.
• Starting balance (optional) sets the amount to measure from instead, for example your first top-up. It is in the balance's currency and applies to the bar of that currency. The line under the bar reads Measured from …, the starting balance set on the DeepSeek tab. Leave it empty to measure as without one.

What a top-up does to a starting balance. A top-up after you set it replaces it: the bar measures from the balance after the top-up, and the DeepSeek sub-tab says No longer used: the balance rose on <date> (a top-up), so the bar measures from the balance after it. Enter an amount again to measure from it. To use a starting balance again, type an amount into the field (the same amount works) and save. Saving the sub-tab's other fields does not bring the old amount back.

A top-up made before you set the starting balance does not replace it, even when the balance is now above the amount. Saving a starting balance reads the balance right then, and a rise replaces the amount only when the relay had read the balance after you set it. So you can top up, enter the new balance and save. If that read fails, the amount stays until a top-up after the relay's next successful read.

Import usage export #

The card's spend is an estimate. For DeepSeek's own figures, export your usage on DeepSeek's usage page (DeepSeek usage at the card's foot links there) and import it. The export is two CSV files for the period you choose: the amounts (tokens and requests per model and API key, with their prices) and the cost.

1. Export a period on DeepSeek's usage page.
2. In Settings → Providers → DeepSeek, under Import usage export, choose the Amount file and the Cost file, then press Import usage export.
3. The line below the button says what was imported (the rows, the days, the models) and whether the card uses it.

What the card shows then. On the days the export covers, the figures of the key saved on the tab replace the estimate: tokens, requests and cost per model, by the export's days. Every other day stays estimated, a gap between two imports included. The labels read Spend, actual and estimated and Spend by model, actual and estimated, a model's row adds how many requests the export counted, and a note names the end of the last export.

• The saved key only. The export lists every key of the account; the card counts the saved key's rows, from wherever the key was used, not only from this relay.
• One currency. The card uses only rows in its own currency, the balance's. When the export is in another one, the import says so, and the card stays estimated.
• Searches. On export days the export's cost includes the web searches, so the card adds no search estimate there.
• Days left out. A day with a period that has not ended yet (today, in an export taken during the day), and the export days up to and including the day this relay's daily history starts, because not all of their turns are in that history. A note says so.
• Time zones. The export's days run in the time zone it was taken in, the estimate's in UTC. At the edge between an export day and an estimated day, an hour can be counted twice or not at all.
• Monthly budget. With export days in the month, the budget bar is no longer marked estimated.

How the import checks the files. For every period and model, the amounts times their prices must add up to the cost file. Otherwise nothing is imported, and the line says The files do not match for …. Importing the same files again changes nothing. A later export replaces, for each key in it, the days it covers, also when it was taken in another time zone. The two files together may be up to 2 MB. The relay keeps the rows in a table of their own, with each key masked to its first and last characters; its own counts of your turns are not changed. Agents of a paired relay cannot import.

On the relay's machine, npm run deepseek:import -- amount.csv cost.csv <database> in OAR's folder does the same. Name the database, copilot.db in the data folder: without it, the command uses data/copilot.db beside the server.

How the estimate is made #

DeepSeek reports no cost per request, and Claude Code's own cost figure is at Claude's prices. OAR therefore counts the tokens each DeepSeek turn on this relay reports, per model, and prices them at DeepSeek's published prices (read 2026-10-19), in US dollars and in yuan. Per million tokens, off-peak:

• Peak and off-peak. Peak prices are exactly twice these. Peak is Monday to Friday 01:00–04:00 and 06:00–10:00 UTC (09:00–12:00 and 14:00–18:00 Beijing time); the card names those hours in your browser's time zone. Each turn is priced by its own time.
• Holidays. On Chinese public holidays DeepSeek charges off-peak prices all day. The estimate does not know those dates, so turns at peak hours on a holiday are counted at double the real price.
• Searches. Web searches are counted at $0.75 or ¥5 per 1,000 from 2026-10-31 16:00 UTC (2026-11-01 00:00 Beijing time), the same at any hour; earlier ones are counted at no cost. See WebSearch and WebFetch.
• Not billed. DeepSeek does not bill the count_tokens requests Claude Code sends at each session start, and the estimate does not count them.
• Unknown models. deepseek-v4-flash and deepseek-v4-flash-vision-exp are priced as Flash. A model without a known price gets no estimate: its row reads no price known, and the total says how many models it leaves out.
• This relay only. Turns on other relays, and other use of the key, are not in it. DeepSeek's usage page, linked from the card, has every request on the account.
• Actual figures. On the days an imported usage export covers, the card shows DeepSeek's own figures for the saved key instead of the estimate, from every use of the key.
• Restarts. A Claude Code process that starts over, or a report that arrives late or twice, is booked once. One case remains: a late report from an earlier process that arrives after a newer one started is booked in full, and so is the newer process's next report.
• Rows under a Claude name. Builds from before sub-agents were pinned to the chat's model booked some turns under a Claude model name. The relay folds those rows into the DeepSeek model that served them when it starts. A folded row is priced at one time, its last turn's, so a row that spans peak and off-peak turns is approximated: its token cost by up to a factor of two.

A turn time more than 5 minutes ahead of the relay's clock is taken as the relay's own time.

Limits #

• Thinking cannot be switched off, and none runs as low.
• DeepSeek V4 Pro takes no images; a switch to it is refused once a conversation holds images.
• The spend is an estimate, of this relay's turns only, and overcounts peak turns on Chinese public holidays. Only the days of an imported usage export show DeepSeek's own figures.
• A changed effort or model restarts the conversation's Claude Code between turns.
• Claude Code's server-side web fetch is not available on DeepSeek; WebFetch runs on the relay's machine.
• No usage-limit pause: a turn refused for the balance fails, and you send it again after topping up.

Error codes #

A failed DeepSeek reply ends with Error code: relay.<code>. and, for some codes, buttons that go straight to the fix:

DeepSeek settings opens Settings → Providers → DeepSeek. Other codes follow the same pattern, relay.deepseek- and Claude Code's reason, for example relay.deepseek-overloaded (HTTP 529) or relay.deepseek-model-not-found. None of them carries the Claude account advice.

Troubleshooting #