Heroku's MCP server ships inside the Heroku CLI, so Atlas connects with no separate service to install. Upgrade to Heroku CLI 10.8.1 or newer, run heroku login, then add the bundled server with atlas mcp add heroku -- heroku mcp:start. From there Atlas can read dyno logs, add config vars, scale processes, and write the Procfile as a reviewable diff.
What the Atlas and Heroku integration does
Heroku's MCP server ships inside the Heroku CLI, which means Atlas gains Heroku access in 2026 without a second daemon to manage. Atlas can read dyno logs, add config vars, scale processes, and write the Procfile as a reviewable diff, all from the terminal where your code already is.
Heroku is a hosting platform, and the classic debugging loop on it is split across two windows: the dyno logs in one, the repository in the other. Because Heroku bundles its MCP server into the Heroku CLI, Atlas can collapse that loop. The agent reads dyno logs, identifies the failure, and then edits the files responsible, because it is running in your working tree. Config vars and process scaling sit behind the same connection, so a fix that needs both a code change and an environment change stays inside one conversation instead of three tools.
How to add the Heroku MCP server to Atlas
Setup is three commands. Upgrade to Heroku CLI 10.8.1 or newer, run heroku login, then add the bundled server with atlas mcp add heroku -- heroku mcp:start. The version floor matters: mcp:start is only present in Heroku CLI 10.8.1 and later, so an older CLI will simply fail to launch it.
The bundled path is the one to prefer because heroku mcp:start reuses your existing Heroku CLI session. No API key has to sit in atlas.json, and no long-lived token needs rotating. If you cannot upgrade the CLI, or you want to run the server from npm, the alternative is atlas mcp add heroku --env HEROKU_API_KEY=$HEROKU_API_KEY -- npx -y @heroku/mcp-server. That route works, but it does put a HEROKU_API_KEY into the environment Atlas launches the server with, so treat the key like any other production credential.
Why heroku mcp:start beats storing an API key
heroku mcp:start reuses your existing Heroku CLI session, so no API key has to sit in atlas.json. That is the main reason to choose the bundled server over the npm package in 2026: the credential lifecycle stays inside the Heroku CLI, where you already log in and log out.
Credentials in configuration files are the quiet failure mode of agent tooling. They get committed, they get copied between machines, and they outlive the person who created them. The Heroku CLI already holds a session after heroku login, and mcp:start simply borrows it. The npm alternative, atlas mcp add heroku --env HEROKU_API_KEY=$HEROKU_API_KEY -- npx -y @heroku/mcp-server, exists for environments where the bundled CLI path is not available. Use it deliberately, and keep HEROKU_API_KEY in your shell environment rather than pasting the literal value into atlas.json.
Debugging a crashing dyno with Atlas
The most common Atlas and Heroku task in 2026 is a crashing dyno. Ask Atlas to read the crashing dyno's logs and locate the module that fails to load at boot. Because the Heroku MCP server is bundled in the Heroku CLI, Atlas fetches those logs directly instead of asking you to paste them.
Boot failures on Heroku usually announce themselves as a module that will not load: a missing dependency, a require path that only resolved on your laptop, or a config var the process expected and did not find. Atlas reads the dyno logs, finds the failing module, and then does something a log viewer cannot: it opens the file in your repository and checks whether the import actually resolves. If the problem is a missing config var, Atlas can add the config var and scale the dyno, approving each command at Atlas's permission prompt, so nothing changes on Heroku without you saying yes.
Letting Atlas write the Procfile as a diff
Atlas writes the Procfile and app.json changes as a diff you review before git push heroku main. In 2026 that review step is what separates an agent that edits your Heroku app config from an agent that proposes an edit and waits, which is the behavior most teams actually want.
The Procfile decides what Heroku runs, and app.json decides how the app is provisioned. Both are files in your repository, which puts them squarely in Atlas's normal editing model: it proposes, you read the diff, you accept or reject. Atlas can reason about them with real information, because through the Heroku MCP server it already knows which processes are running and what the config vars contain. When the diff is right, the deploy is the ordinary git push heroku main. Nothing about the deploy path changes just because an agent wrote the file.
Keeping Atlas's Heroku access under your approval
Atlas's permission prompt is the control surface for Heroku work in 2026. When Atlas adds a config var or scales a dyno, it asks first, and you approve each command. Combined with the diff review on the Procfile and app.json, every write to Heroku passes through an explicit human yes.
There are two classes of Heroku action Atlas can take, and both are gated. The first is a direct platform action: adding a config var, scaling a process. Atlas surfaces these at its permission prompt, and you approve each command individually. The second is a repository change, such as the Procfile or app.json, which arrives as a diff you review before git push heroku main. Because heroku mcp:start reuses your existing CLI session, revoking access is equally simple: log out of the Heroku CLI and the MCP server has nothing to reuse.
Setup
- 01Upgrade to Heroku CLI 10.8.1 or newer, since the bundled mcp:start command is only available from that version onward.
- 02Run heroku login so the Heroku CLI holds an active session for the MCP server to reuse.
- 03Add the bundled server with atlas mcp add heroku -- heroku mcp:start
- 04If you need the npm route instead, run atlas mcp add heroku --env HEROKU_API_KEY=$HEROKU_API_KEY -- npx -y @heroku/mcp-server
- 05Ask Atlas to read the crashing dyno's logs and locate the module that fails to load at boot.
- 06Let Atlas add the config var and scale the dyno, approving each command at Atlas's permission prompt.
- 07Have Atlas write the Procfile and app.json changes as a diff you review before git push heroku main.
Frequently asked questions
- how to connect atlas to heroku mcp server
- Upgrade to Heroku CLI 10.8.1 or newer, run heroku login, then run atlas mcp add heroku -- heroku mcp:start. The MCP server ships inside the Heroku CLI.
- do i need a heroku api key to use atlas
- Not with the bundled server. heroku mcp:start reuses your existing CLI session, so no API key has to sit in atlas.json. The npm route does need HEROKU_API_KEY.
- what heroku cli version is required for mcp
- Heroku CLI 10.8.1 or newer. The bundled MCP server is started with heroku mcp:start, which older CLI versions do not provide.
- can atlas read heroku dyno logs
- Yes. Ask Atlas to read the crashing dyno's logs and locate the module that fails to load at boot. Heroku's MCP server exposes dyno logs to Atlas directly.
- can atlas scale heroku dynos and add config vars
- Yes. Atlas can add config vars and scale processes, approving each command at Atlas's permission prompt before it runs.
- how do i install the heroku mcp server from npm
- Run atlas mcp add heroku --env HEROKU_API_KEY=$HEROKU_API_KEY -- npx -y @heroku/mcp-server. This is the alternative to the CLI-bundled server.
- will atlas deploy to heroku without asking me
- No. Atlas writes the Procfile and app.json changes as a diff you review before git push heroku main, and platform commands go through Atlas's permission prompt.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas with Gemma 3 27B (local via Ollama): Setup, Cost, and Tradeoffs in 2026
Run Atlas on Gemma 3 27B (local via Ollama) in 2026: a 131,072 token context, Free (self-hosted) pricing, single-GPU inference, and the honest tradeoffs.
Atlas with Grok 4.3: The Cheapest 1M Context Reasoning Model (2026)
Grok 4.3 gives Atlas a 1M token context at $1.25 / $2.50 per Mtok, the cheapest 1M reasoning model from a US lab. Output caps at 30,000 tokens, so plan around it.
Atlas for Bun: Terminal-Native AI Coding in 2026
Atlas is a terminal-native AI coding agent for Bun. Swap Node compatibility shims for Bun.serve and bun:sqlite, iterate with bun test, and format with bun fmt in 2026.
Atlas with Qwen Max: Setup, Cost, and the 32,768 Token Limit in 2026
Run Atlas on Qwen Max in 2026. Alibaba's original flagship costs $1.60 per Mtok input, $6.40 per Mtok output, with a 32K tokens (32,768) context window.
Atlas with DeepSeek R1 (0528): The Open Reasoning Trace, 2026
Run Atlas on DeepSeek R1 (0528) in 2026. DeepInfra hosts the MIT-licensed open reasoning model at $0.50 per Mtok in, $2.15 per Mtok out, 160K tokens context.
Atlas with Qwen2.5-Coder 1.5B (Ollama): the 986MB small_model slot in 2026
Qwen2.5-Coder 1.5B (Ollama) in Atlas: a 986MB Q4_K_M pull with 32K tokens (32,768) of context, Free (self-hosted), sized for titles and summaries, not refactors.
Atlas with Fireworks AI (gateway) in 2026: Buying Latency with Money
Fireworks AI (gateway) serves Atlas open models with fast-router tiers: DeepSeek V4 Flash at $0.14 / $0.28 per Mtok on a 1,000,000 token context.
Atlas with DeepSeek Coder V2 16B Lite (Ollama): The 160K Local Agent Model in 2026
DeepSeek Coder V2 16B Lite (Ollama) gives Atlas a 160K token context from an 8.9GB download, Free (self-hosted). Setup, MoE speed, and the KV cache catch, for 2026.