Skip to content

Open-source agent harness

You run an agent
inside Elpis.
It becomes Elpis.

The open-source home for coding agents—context, continuity, tools, evidence, and deterministic work in one place.

Context ledgerLean continuityProvider choiceDeterministic work graphs

One terminal turn

See the harness
at work.

Scroll to play. Reverse midway to rewind; after the final frame, scrolling up closes the demo.

masih@atlas ~/Desktop/p/Elpis

Elpis is starting · preparing session

ELPIS~/Desktop/p/Elpisgpt-5.6· local
Reading product truth before touching code.
read docs/GUIDE.md docs/context.md docs/sessions.md
inspect documented Ace pruning example231 lines
◆ Ace · pre-admission
readme.md:117 Completed tool evidence before admission: 231 lines, 18,930 characters.
docs/context.md:14 RTK filtering → deterministic cap → optional Ace pruning.
docs/context.md:26 Native compaction remains an independent process.
docs/sessions.md:6 GOAL.md, ES.md, and applicable rules form lean continuity.
[Ace · same result · 248 characters retained]
Relevant conclusion retained; exact tool-call envelope and immutable audit remain available.
Documented example: 18,930 → 248 characters · readme.md:117–142
I’m oriented. Elpis admits explicit continuity sources, keeps provider history authoritative for exact resume, and records pruning evidence without changing already-sent history.
elpising…0s · esc to interrupt
Enter send · Shift+Enter newline
Tab focus · Alt+C focus/hide · Ctrl+click open file
Esc interrupt · Tab focus Ledger · /context full viewcontext not measured

/context

See what filled
the window.

A full, rendered history report—not JSON and not a source-toggle panel. This illustrated request is exactly 50% full; the headline is measured while category attribution is estimated.

/context illustrated measured snapshot

Context Usage

gpt-5.6-terra · 129.2k / 258.4k tokens (50% used)

Token usage by category

User messages: 7.8k tokens (3.0%)

Agent responses: 45.6k tokens (17.6%)

Tool calls: 54.3k tokens (21.0%)

System prompt: 3.1k tokens (1.2%)

Development rules: 5.4k tokens (2.1%)

Other (overhead): 13.0k tokens (5.0%)

Free space: 129.2k (50% left)

Headline is measured; category sizes are attribution estimates.

Category Breakdown · retained ■ reclaimed ✦

User messages7.8kAgent responses45.6kTool calls54.3kOther categories21.5k

Ace Pruning Audit & Low-level Breakdown

documented example 231 lines · 18,930 characters reviewed

retained 248-character replacement · immutable before/after artifacts

prune_report.md · ace.json · manifest.json

Checkpoints · Esc Esc to backtrackAvailable after a user message; selecting one forks from that prior point.

Dashboard

The whole run.
One glance.

Timing, context, tokens, and backend cost from one recorded successful run.

Elpis · recorded run

Completed · revision 12

successful
Latest outcome
Completed
Elapsed · 930 ms
Backend cost
$1.25
reported total
First token
240 ms
3 sampling requests
Requests
3
1 retry · pruning off
Where 930 ms wenttool blocking dominated
Sampling 200Tools 500Other 230 ms
Recent activity
OutcomeTotal tokensTTFTCost
Completed26.4k240 ms$1.25

Deterministic work graphs

Bounded work.
Visible evidence.

Play a task from assignment to review. Inspect what each agent receives, changes, and hands back.

Interactive example · work graphs are opt-in

Elpis deterministic work graph

Coordinator-owned tasks · dependency-ready dispatch · evidence checks

Open source

Install Elpis.
Run it locally.

Linux x86_64 preview. Inspect the source, then run the installer.

Installation and source ↗
curl -fsSL https://raw.githubusercontent.com/MasihMoafi/Elpis/main/scripts/install-elpis.sh | bash && ~/.local/bin/elpis