Observe
The Cully terminal and agent hooks turn tool calls into small events: an edit, a check, a search, a memory call.
Session intelligence · Harness improvement · Private memory
The harness runtime
for AI software work.
Your agent writes the code. Cully watches for repeated failures and unchecked edits, helps improve the harness around it, and brings useful context to the next session. Run Claude Code, Codex or Cursor with a second pair of eyes.
Get started with SolomacOS and Linux
curl -fsSL https://cully.net/install.sh | shInstall Cully, then run cully setup --all and cully run claude, cully run codex or cully run cursor. Setup once for your user, then work across projects—no project .cully folder needed. Open a new terminal before running cully. Installation options ↗
Open source · Works across agents. Learns across sessions. · Recall powered by Mem0 ↗
> fix the resource validation test
● ran go test ./auth/...
✕ FAIL TestValidateResource
● edited auth/resource.go
● ran go test ./auth/...
✕ FAIL TestValidateResource
● ran go test ./auth/...
✕ FAIL TestValidateResource
● edited auth/resource.go
● ran go test ./auth/...
✕ FAIL TestValidateResource
── Codexoauth:main27mContext 63%⚠ Loop 3x● 2 unchecked── Ctrl+] advisor ──
HarnessSetup 5/7 readyGaps Skills, Plugins
SessionContext 63% usedMemory used 1×MCP healthy
Example session · values are illustrative
CAUTION Cully noticed something: the same command failed 3 times after edits. Inspect the first failure before editing again.
$ cully harness
Your agent might be stuck.
One terminal for every agent
✳ Claude Code
◈ Codex
▰ Cursor CLI
＋ Any agent you add
01 / THE LOOP
Observe. Understand.
Remember. Advise. Improve.
Coding agents are powerful, but each one mostly sees the task in front of it. Cully keeps track of the broader development process around them.
Understand
A private session journal shows what happened: the same test failing three times, edits nobody checked.
Remember
Decisions, attempts, results and next steps are saved through MCP and found again by meaning.
Advise
A second pair of eyes for your coding agent. It warns about loops, context pressure and missing verification.
Improve
Cully evaluates its setup and guidance every five hours, proposes fixes for failing checks, and writes a change only after you approve it.
Across agents
Start in Claude Code, continue in Codex, finish in Cursor. Your saved context stays with you across agents.
02 / KEEP WORK MOVING
Catch the loop.
Find the next move.
The same test fails again. Edits pile up without a check. Cully brings those signals into view, helps you inspect what happened, and prepares a handoff when you need a fresh start.
01 ↗
ADVISORLOOP
Same commandfailed 3 times
A second pair of eyes.
The advisor watches the session while the agent works on the task. Click it in the terminal to preview a suggestion before anything changes.
Explore the advisor02 ↗
▶ REPLAYRead resource.go → Edited resource.go → go test ✕ → go test ✓8 files touched · 2 checks failed · 1 loop
See what the agent did, and how.
Replay a session step by step: the files it read, edited, created and deleted, and the commands it ran.
Explore session intelligence03 ↗
HANDOFF
✓ 34 tests passed
! Databricks integration unverified
Hand it to the next agent.
A structured handoff, or cully rescue when a session is stuck, so a different agent continues without rebuilding the context.
03 / THE IMPROVEMENT LOOP
A better harness.
Built from evidence.
Cully tests its own setup and guidance, turns failing checks into proposals, and lets you review the change before applying it. The next evaluation shows whether those checks improved.
01 ↗
01 / EVALUATESetup · harness · advisor
Check against known failuresIllustrative workflow · not a benchmark result
Check against known failuresIllustrative workflow · not a benchmark result
Test the harness, too.
cully eval exercises real Cully code against planted failures. Find where setup detection, warnings or the improvement loop miss their expected behavior.
02 ↗
02 / PROPOSEREVIEW
cully improveRead the failing checksPreview the plan
Turn failures into a plan.
Open a proposal, inspect the checks behind it, and ask your coding agent to plan a change to Cully’s guidance. Read the exact plan before approving a write.
Explore the improvement loop03 ↗
03 / CHECK AGAIN
↻ Run the next evaluation
Compare failing checks over time
See whether the fix holds.
Follow evaluation history and before-and-after task measurements. Fewer failing checks are useful evidence; differences between tasks still need your judgment.
Read what the scores meanPRIVATE MEMORY / ACROSS AGENTS
New session.
Less starting over.
Keep the decision, the failed attempt and the next step. Give your next agent useful context without making it reconstruct the whole conversation.
Save what mattered
Your agent saves a concise note: what it tried, what worked, what blocked it and what comes next. Session references connect related notes.
Find it by meaning
Ask about the work in your own words. Semantic recall through Mem0 finds relevant saved notes; PostgreSQL keeps the source records.
Keep your own context
Use the same private memory from Claude Code, Codex or Cursor. On an OAuth deployment, each signed-in contributor’s records stay scoped to their identity.
A shared deployment does not share your notes. The session journal stays local; durable memory contains the summaries your agent saves.
See how private memory worksYOUR AGENT SETUP
Know what’s configured.
Measure what changes.
Instructions, permissions, memory, skills and hooks shape how your agent works. Start with a read-only audit, then compare recent output before changing your setup.
01
HARNESSREAD-ONLY
cully harnessFind setup gapsKeep your settings
Check the setup around your agent.
See which of seven components Cully finds, with one suggestion per gap. The terminal keeps readiness, pending checks and active loops in view. Optional plugins stay optional.
Read the harness guide02
OUTPUT BASELINEOversized results
Estimated tokens / sessionLocal Claude Code or Codex history · estimates
Estimated tokens / sessionLocal Claude Code or Codex history · estimates
Compare the output you measured.
Find oversized tool results in supported Claude Code or Codex history, with extra file-read and Bash detail for Claude. Change one component and compare similar work to investigate its effect.
Understand the measurements03
YOUR WORKFLOW
cully workflowBuild · test · review · release
Give the advisor your working rules.
Create an optional private workflow profile with your build, test and release habits. Review it before saving. Your configured advisor model gets bounded, redacted context; the profile grants no access.
Set your workflow contextSetup detection checks file names and configuration keys across supported agents; it does not verify a live connection. Historical output estimates follow the selected agent’s supported local history: Claude Code or Codex. Cursor history measurements are unavailable. See what stays on your machine →
04 / THE TERMINAL
One terminal.
Every agent.
Cully is not another terminal emulator to learn. It is the observation surface that lets it understand what any agent is doing, using the agent exactly as before.
Same wrapper everywhere
cully run claude, codex, cursor or any executable on your PATH gets the same health bar and advisor panel.
Harness and session together
The Harness column shows readiness and gaps. Session keeps context, tokens, quota and Git changes beside observed memory use, analyses avoided and estimated net tokens.
An advisor you can click
Click the panel or press Ctrl+] to open every suggestion. Press Enter to put a setup request in your prompt; nothing runs until you send it. Once an hour it can also research current agent techniques from public sources, without your private task details.
Supported on macOS and Linux. Cursor support wraps the cursor-agent CLI. Cully's tool events come from each agent's hooks, so an agent without hooks gets the health bar and Git state but not loop detection or replay.
MEASURE THE OVERHEAD
Useful work.
Visible costs.
Memory and advice have a cost too. Inspect the events Cully observed, the context it injected and how its advice performs on your own history.
cully impact
Memory hits, empty lookups, analyses avoided, loops caught and oversized output flagged. Use --explain to inspect how the net estimate is built and see Cully’s measured injected cost alongside it.
cully savings
Estimate Cully results, arguments and schema tokens in supported Claude Code or Codex history. Compare early exploration with and without a memory hit, with a confidence label.
cully eval
Score setup, harness, advisor and memory against planted ground truth and published metrics. An optional live pass checks the advisor cites no figure it was not given.
These are different samples and estimates, not billing totals or causal proof of savings. Historical transcript reports support Claude Code and Codex; unavailable measurements are labeled. Live evaluation consumes model tokens and can use web research.
Read the measurement guideSELF / DEVELOPER PREVIEW
Return to your work.
Bring the context.
Cully Self brings dated personal work notes, local tasks and read-only questions to Android and iOS. Available as source builds while device verification continues.
A small brief, with its evidence.
See a sampled last 24 hours or seven days, recall a saved approach, and choose a next step. Dates and coverage stay visible. A last-planned handoff is a suggestion to review, not a live project status.
Ask in your own words.
Type or explicitly start the microphone. Downloaded models run on-device and use bounded read-only recall over your signed-in Cully memory. Personal task changes stay local; no Team actions or remote coding agents are exposed.
Model setup downloads roughly 469–499 MiB after confirmation. History retrieval still needs network access. Store distribution is not configured, and physical iPhone sign-in, voice and chat remain unverified.
05 / GET GOING
Three commands to
a smarter session.
Install Cully, set up its local services, and run your agent through it. Already installed? Run cully update --check to check for a release, then cully update to install a verified CLI update. Services use your deployment tooling.
01
Install Cully
Add Cully and its local advisor with one command.
02
Set it up
Run cully setup --all to start the memory services and connect your agents.
03
Run your agent through Cully
Run cully run claude, codex or cursor. Then try cully status, cully replay and cully handoff, and run cully improve to review what Cully suggests for itself.
06 / UNDER THE HOOD
Built around the agent,
not in its way.
Your agent talks to its model and tools directly. Cully draws around it, records small events and keeps durable notes in your own services.
Your coding agentsClaude Code · Codex · Cursor CLI · others
Keep working in the tools you already use.
Cully at workTerminal · hooks · journal · advisor · MCP
Health bar, loop detection, replay, handoff and rescue, plus tools to save and recall context.
Your recordsPostgreSQL · self-hosted Mem0
Private source notes, session links and semantic recall across your coding agents.
Run the memory stack locally or connect to an OAuth deployment. Verified sign-in keeps each contributor’s private records separate. Self-hosting guide →
PRIVACY
Your conversations aren't Cully's database.
The local session journal keeps the time, the kind of tool, whether it passed, project-relative file paths and the program of a command. It never keeps prompts, file contents, command arguments or tool output, and you can turn off path recording. Private memory holds only the short notes your agent chooses to save. Harness and cost reports read supported local Claude Code or Codex history for aggregate counts and estimates. Replay evaluation can send signal summaries to your configured advisor when you explicitly request a live pass; these measurements never save or upload raw transcript text.
PRIVATE MEMORY
Semantic recall, powered by Mem0.
Cully uses self-hosted Mem0 to index saved decisions, lessons and task context by meaning. PostgreSQL keeps the source records; Mem0 powers semantic search across them.
07 / COMMUNITY
From real sessions.
In their words.
How does Cully fit into your workflow? We'd love to hear.