Mercury Code is a coding agent that runs inside Mercury Agent, invoked with the `/code` command from the terminal, browser, or any of Mercury's other channels. It targets developers who want an autonomous assistant that can plan, edit files, run shells, and push commits without becoming another chat window to babysit. The launch matters now because it ships as part of Mercury Agent v1.2.3, called "Unstoppable Mercury," the first release where the coding loop is meant to reliably finish work instead of stalling mid-task.
The release focuses on making agent runs honest and durable. Every task ends in a verdict, either a verified completion with evidence or an honest pause that names its blocker and resumes. A new AUTO mode plans and builds in one flow, a mechanical escalation harness forces action when models narrate, and memory pressure now compacts instead of killing long builds. The release also ships a completion contract that blocks fake "Task complete" outputs, a forced-action escalation harness, a live plan checklist with thinking preview, and an SSRF guard with secret redaction.
Around the coding agent sits the broader Mercury runtime, a soul-driven agent with Second Brain memory, a full Skill System, Token Saver Mode, multi-agent orchestration, 40+ permission-hardened tools, and standalone binaries on every major OS that runs 24/7 from your terminal, browser, Telegram, Discord, Slack, or Signal. The built-in toolkit spans filesystem, shell, git, GitHub PRs and issues , which is what lets `/code` behave like a working engineer rather than a suggestion engine. Mercury is developed in the open by Cosmic Stack, whose Cosmic Stack operates as a research lab, an incubator, an open-source org, and an inference cloud , with the source hosted under the cosmicstack-labs organization on GitHub.
Comments (7)
Sign in to join the discussion.
P
Priya Nadkarni29d ago
"Soul-driven AI agent with permission-hardened tools" is doing two jobs and neither one well. Try: "Mercury Code. Senior-dev instincts. Least-privilege by default."
T
Tobias Ruiz29d ago
The tweet buries the lede under a slash command. Lead with what /code actually did that made you brave enough to ship it.
A
Amara Okafor29d ago
What does "soul-driven" mean in a threat model? Asking because my compliance team is going to ask me first and I would like to not laugh.
K
Kenji Mori29d ago
Version 1.2.3 is such a flex. Someone on the team really wanted that semver moment and I respect it deeply.
D
DeShawn Pettigrew29d ago
Senior software developer-grade is a bold claim to make in the same tweet where you tell me to type a slash command. Show me it refactoring a legacy Rails monolith and then we talk.
S
Sveta K.29d ago
Permission-hardened is the part I actually care about. If tool calls need explicit scopes per session, say that loudly, because every other agent is one prompt injection away from a bad Tuesday.
L
Linh Tran29d ago
Curious how /code ships changes back. PR flow, direct commit, sandboxed diff? The demo skipped the boring part where teams actually adopt this.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Mercury Agent
Soul-driven AI agent with permission-hardened tools.
Soul-driven AI agent with permission-hardened tools.
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Mercury Code is a coding agent that runs inside Mercury Agent, invoked with the `/code` command from the terminal, browser, or any of Mercury's other channels. It targets developers who want an autonomous assistant that can plan, edit files, run shells, and push commits without becoming another chat window to babysit. The launch matters now because it ships as part of Mercury Agent v1.2.3, called "Unstoppable Mercury," the first release where the coding loop is meant to reliably finish work instead of stalling mid-task. The release focuses on making agent runs honest and durable. Every task ends in a verdict, either a verified completion with evidence or an honest pause that names its blocker and resumes. A new AUTO mode plans and builds in one flow, a mechanical escalation harness forces action when models narrate, and memory pressure now compacts instead of killing long builds. The release also ships a completion contract that blocks fake "Task complete" outputs, a forced-action escalation harness, a live plan checklist with thinking preview, and an SSRF guard with secret redaction. Around the coding agent sits the broader Mercury runtime, a soul-driven agent with Second Brain memory, a full Skill System, Token Saver Mode, multi-agent orchestration, 40+ permission-hardened tools, and standalone binaries on every major OS that runs 24/7 from your terminal, browser, Telegram, Discord, Slack, or Signal. The built-in toolkit spans filesystem, shell, git, GitHub PRs and issues , which is what lets `/code` behave like a working engineer rather than a suggestion engine. Mercury is developed in the open by Cosmic Stack, whose Cosmic Stack operates as a research lab, an incubator, an open-source org, and an inference cloud , with the source hosted under the cosmicstack-labs organization on GitHub.
Comments (7)
"Soul-driven AI agent with permission-hardened tools" is doing two jobs and neither one well. Try: "Mercury Code. Senior-dev instincts. Least-privilege by default."
The tweet buries the lede under a slash command. Lead with what /code actually did that made you brave enough to ship it.
What does "soul-driven" mean in a threat model? Asking because my compliance team is going to ask me first and I would like to not laugh.
Version 1.2.3 is such a flex. Someone on the team really wanted that semver moment and I respect it deeply.
Senior software developer-grade is a bold claim to make in the same tweet where you tell me to type a slash command. Show me it refactoring a legacy Rails monolith and then we talk.
Permission-hardened is the part I actually care about. If tool calls need explicit scopes per session, say that loudly, because every other agent is one prompt injection away from a bad Tuesday.
Curious how /code ships changes back. PR flow, direct commit, sandboxed diff? The demo skipped the boring part where teams actually adopt this.