Inflara
Create account

Inflara setup & operating guide

From a clean machine to a supervised agent team.

This is the single, complete guide for installing Inflara Runner, keeping it online in the background, and orchestrating Codex, Claude Code, and Antigravity from AI Agent Center.

15–25 minutes macOS, Linux, Windows Codex, Claude, Antigravity

Inflara account and project

Create an account, create the project you want agents to work on, and add at least one concrete task.

At least one agent CLI

Install Codex, Claude Code, or Antigravity separately and complete that tool’s own sign-in before Runner launches it.

A machine appropriate for the work

A personal computer is fine for private work. Use a VM, cloud workstation, or dedicated sandbox for shared agents or full autonomy.

01

Create the Inflara side of the connection

Runner pairing and MCP access are separate credentials. Set up both once, for different jobs.

  • Create an Inflara account or sign in, then create the project you want the agents to work on.
  • Add at least one clear task with the expected outcome, relevant constraints, and acceptance criteria. The task becomes selectable in Agent Center.
  • Open Settings → Agentics → Remote access token for MCP/API. Choose Supervisor (recommended), select all projects or named projects, create the token, and copy the ifl_... value immediately.

Know which credential you are using

inflara login pairs Runner and stores a local runner credential. The ifl_ MCP token lets your agent apps read and update Inflara through MCP. Never use one in place of the other.
02

Prepare the coding agents and install Runner

Runner launches the real local agent CLIs. Those tools must already be installed and signed in on this computer.

Install any agents you plan to assign—Codex CLI, Claude Code, or Antigravity CLI—using their vendor setup. Open each one once, complete sign-in and workspace trust, then verify the installed commands you use:

codex --version
claude --version
agy --version

You do not need all three. Missing commands are fine when you do not plan to use that agent type.

# macOS / Linux
curl -fsSL https://download.inflara.io/runner/install.sh | sh

# Windows PowerShell
irm https://download.inflara.io/runner/install.ps1 | iex

Open a new terminal if inflara is not found, then run inflara help to confirm the self-contained CLI is on your path. You do not need Node.js, pnpm, or an Inflara source checkout.

Shared machines need deliberate isolation

Provision the VM or sandbox and its OS user before pairing Runner. Map only approved folders, use least-privilege tokens, and sign agent CLIs into accounts appropriate for the team’s work.
03

Pair Runner, install the agent integration, and map a project

These commands connect the machine, install the global Inflara Supervisor skill and MCP entries, and tell Runner where the project lives locally.

inflara login
inflara setup-agents --mcp-token "ifl_..."
inflara projects list
inflara projects map "<project name or id>" /absolute/path/to/project
inflara doctor
inflara smoke
  • inflara login opens a browser approval page. Approve this machine with the Inflara account that owns or can run agents in the project.
  • inflara setup-agents installs the global supervisor skill and MCP configuration for supported local agents. The alias inflara runner setup-agents does the same thing.
  • Map each Inflara project to one absolute local folder. Git repositories use isolated worktrees; ordinary folders run directly in place and serialize edit jobs.
  • Restart every open Codex, Claude Code, or Antigravity session after setup so it loads the new skill and MCP configuration.
Avoid putting the MCP token in shell historySupply it through a temporary environment variable instead of a command argument.Optional
# macOS / Linux
export INFLARA_MCP_TOKEN="ifl_..."
inflara setup-agents
unset INFLARA_MCP_TOKEN
# Windows PowerShell
$env:INFLARA_MCP_TOKEN="ifl_..."
inflara setup-agents
Remove-Item Env:INFLARA_MCP_TOKEN
Create deterministic repository contextRunner creates a repo map automatically for jobs; prebuild it and scaffold shared knowledge when you want to inspect or customize the inputs.Optional
# Optional: generate the context map used by agents
inflara repo map --project <project-id>

# Optional: add shared knowledge files without overwriting edited files
inflara repo init-knowledge --project <project-id>

init-knowledge does not overwrite knowledge files you have edited.

04

Keep Runner online

The background service is the recommended everyday mode. You do not need an open terminal once the service is installed.

# Recommended: start automatically and update only while idle
inflara runner install-service --auto-update

# Optional: save parallel worker capacity in the service
inflara runner install-service --auto-update \
  --workers codex=2,claude_code=1,antigravity=1

Background service

Recommended. Starts after you sign in to the computer. On macOS and Linux, installation starts it immediately.

Foreground

Best for debugging. It stops when you press Ctrl+C or close the terminal.

Run once

Claims and processes one available job cycle. Useful for controlled tests.

# Runs only while this terminal stays open
inflara runner start

# Optional: parallel workers in one runner process
inflara runner start --workers codex=2,claude_code=1,antigravity=1

# One polling cycle for debugging
inflara runner run-once

Do not start duplicate runners

One runner process can host multiple workers. Use --workers instead of opening multiple same-token terminals. On macOS, foreground start is blocked while the managed background service is active unless you deliberately use --force.

On Windows, install-service creates a logon task. Sign out and back in, or start it immediately with:

schtasks /Run /TN "Inflara Runner"

Confirm the connection with inflara doctor, then open Agent Center → Machines and Agent Center → Agents. The runner and every detected CLI you intend to use should show as active.

Runner settings are managed from the web

After pairing, you configure the runner with toggles instead of config files. In Agent Center → Machines, the Assistant column has two switches per runner you own: Assistantenables private Assistant conversations that use this machine's normal local agent profile, and Push & deploy additionally lets the Assistant perform explicitly requested git pushes, pull requests, and deploys with your local credentials. Changes reach the runner over its own connection and apply within seconds — no config-file edits and no restart.
Set Assistant options from the CLI insteadThe web toggles are the recommended path; these flags persist the same settings locally.Optional
# Enable the Assistant on this runner
inflara runner start --enable-assistant

# Also allow explicitly requested git push / PR / deploy actions
inflara runner start --enable-assistant --allow-assistant-external-actions

# Turn external actions back off
inflara runner start --no-assistant-external-actions

Both settings are off by default. Push & deploy is a runner-owner opt-in that shares your local credentials with explicitly requested actions — leave it off on shared machines.

05

Configure your supervisor in AI Agent Center

The supervisor is the project manager for agent work. It scopes the request, drafts or selects a workflow, assigns roles, and monitors execution.

  1. 1

    Open Agent Center → Supervisor

    If the header says Setup needed, select Setup to open Supervisor settings.

  2. 2

    Choose the supervisor

    Select one active agent, then optionally choose its model and effort level. The supervisor should be reliable at planning, tool use, and coordination.

  3. 3

    Choose role assignment

    Auto is easiest for a new team: Balanced combines reliability and consumption, Highest reliability prioritizes success, and Lowest consumption favors efficiency. Manual lets you drag named agents into planning, coding, review, test/fix, recovery, and final-QA pools.

  4. 4

    Review the assignment preview and save

    Check for offline agents, missing project mappings, or unavailable models. Select Save setup when the preview matches the team you intend to run.

Inflara Agent Center supervisor setup with automatic role assignments
Supervisor settings show the selected supervisor, assignment policy, role coverage, and the reason behind each automatic assignment.
06

Queue and supervise your first orchestrated run

Start with one well-scoped task and Robust mode. You can move to faster or larger teams after the basic path is healthy.

  1. 1

    Choose the workflow strategy

    Use Flexible when the supervisor should choose or compose the workflow. Choose Existing workflow when you want it to start from a saved graph.

  2. 2

    Choose Quick, Robust, or Full team

    Quick uses the smallest viable workflow. Robust adds planning, implementation, review, and test/fix loops and is the best default. Full team adds plan competition, specialists, recovery, and final QA for high-risk work.

  3. 3

    Set the scope

    Choose the project context and select one or more tasks. You may leave tasks empty when the supervisor should only draft or revise a workflow.

  4. 4

    Write the supervisor request

    State the outcome, constraints, acceptance checks, files or systems in scope, and anything that requires your approval. Then select Queue supervisor.

  5. 5

    Open the session

    Follow the transcript for status and questions. Use the Workflow tab to inspect the graph and each step. Approve or deny requests, send corrections, and cancel when the run should stop.

  6. 6

    Review the evidence

    A healthy run has structured handoffs, changed-file and artifact evidence, visible budget usage, and a passing verification gate. Failed runs should preserve the final blocker or fix plan instead of claiming success.

Quick

Low-risk, narrow, reversible work.

Robust

Meaningful coding work with regression risk.

Full team

Production, security, data, or multi-specialist work.

Build and run a workflow manuallyUse Agent Center → Workflows when you want direct control over the saved graph instead of supervisor-led composition.Optional
  • Create a blank workflow or start from Single Worker, Robust Workflow, Full Team With Tests, or Plan Competition.
  • Select steps and configure prompts, agents, models, execution modes, branches, conditions, loops, joins, recovery, and budgets. Graph edits save through the validated workflow API.
  • In the run panel, choose tasks and inputs, review the selected run count, and select Run with inputs.
  • Watch the run console and observability panels; answer interactions and inspect each node’s artifacts and status before accepting the result.
Inflara Agent Center workflow graph with planning, implementation, and verification steps
07

Operate Runner day to day

Most days, the background service needs no attention. These commands cover health, logs, artifacts, capacity, and safe cleanup.

inflara runner status
inflara doctor
inflara runner logs
inflara runner logs --follow
inflara runner artifacts status --json

Health and capacity

Agent Center → Machines shows runner versions and heartbeats. Agents shows detected CLIs, models, effort options, availability, and worker capacity.

Sessions and evidence

Sessions holds recent agent runs. Workflow artifacts stay bounded on the local computer; use status and prune commands to apply your retention policy.

# Preview artifact cleanup, then confirm it
inflara runner artifacts prune --dry-run --older-than 14d
inflara runner artifacts prune --yes --older-than 14d

# Remove a local mapping or the automatic startup service
inflara projects remove <project-id>
inflara runner uninstall-service
Use multiple workers or full autonomyIncrease concurrency only when the machine and mapped repositories can support it. Full autonomy belongs on dedicated sandboxes.Optional
# Parallel workers in one runner process
inflara runner start --workers codex=2,claude_code=1,antigravity=1

# Full autonomy: dedicated VM/sandbox only
inflara runner start --full-autonomy-agents codex,claude_code,antigravity

# Return every agent to its default permission mode
inflara runner start --default-autonomy-agents all
Run local orchestration evalsReplay saved evaluation specs in scratch worktrees and track whether agent behavior improves or regresses.Optional
inflara eval list
inflara eval run --id <eval-id>
inflara eval report
08

Update Runner without leaving the old service running

A manual update replaces the binary but cannot replace code already loaded by a running process. Wait for active work to finish, update, then restart the foreground process or background service.

inflara update
inflara doctor
PlatformRestart after a manual update
ForegroundPress Ctrl+C in the runner terminal, then run inflara runner start again.
macOS serviceRun inflara runner install-service again. It reloads the LaunchAgent with the updated binary and preserves saved settings.
Linux serviceRun systemctl --user restart inflara.service.
Windows taskRun schtasks /End /TN "Inflara Runner", then schtasks /Run /TN "Inflara Runner"; signing out and back in also starts the logon task.

Auto-update waits for Runner to be idle

With install-service --auto-update, Runner does not replace itself while it has active jobs. Managed macOS and Linux services restart it with the new binary after the idle update. On Windows, start the scheduled task again if it exits after updating.

Runner update versus agent-session refresh

A Runner binary update needs a runner-process restart. If the release also changes the MCP configuration or Inflara Supervisor skill, run inflara setup-agents again and start new Codex, Claude Code, or Antigravity sessions so those apps reload the files.
09

Troubleshoot in the shortest path

Start with doctor, then narrow the problem to the agent CLI, MCP setup, project mapping, or background service.

inflara doctor
inflara smoke
inflara runner status
inflara runner logs --follow
inflara command not foundOpen a new terminal so the PATH change loads. If needed, rerun the public installer.
Agent is missingInstall and sign in to that agent CLI, verify its version command, then rerun doctor and restart Runner.
MCP or skill is missingCreate a valid MCP token, rerun inflara setup-agents, and start a new agent session. Use --agents, --skills-only, or --mcp-only only for targeted repair.
Project cannot runConfirm your project role includes agent execution, map the project on this runner, and make sure the selected supervisor’s machine owns that mapping.
Runner is offlineInspect logs and service state. Reinstall the service to restore saved startup configuration; do not launch a duplicate foreground process beside it.
Run is waitingOpen the Agent Center session. Answer its pending question or approval, then send a reply or resume the session when the UI offers it.

Still blocked? Keep the output from inflara doctor --json and the relevant runner log lines. Do not include tokens or other secrets when sharing diagnostics.

Ready to orchestrate

Keep Runner online. Let Agent Center coordinate the work.

Start with one project, one agent, and one Robust run. Add concurrency and larger workflows only after the evidence path is healthy.

Open InflaraOrchestration white paper